Scope: This assessment examines whether China’s 2025–2026 international artificial-intelligence initiatives constitute a strategic attempt to build an alternative global AI ecosystem around capacity-building, standards, applications and institutional relationships, while comparing the documented responses of the United States and Europe over a five-year horizon through 2031.

Executive Summary / BLUF

The verified record supports the central proposition that Beijing is deliberately widening the definition of international AI competition beyond frontier-model performance and semiconductor capacity toward institutional architecture, human-capital formation, application deployment, standards coordination and long-term technology partnerships, although the stronger proposition that China has already established an international AI system capable of displacing the United States or Europe is not supported by currently available official evidence. President Xi Jinping’s 17 July 2026 WAIC address formally connected AI governance with capacity-building for the Global South, announced 5,000 AI training and seminar opportunities over five years, called for international AI application cooperation centres involving ASEAN, the Arab League, African Union, CELAC, Shanghai Cooperation Organization and BRICS, and announced planned deployment of the MAZU AI meteorological warning system in 30 countries. [Joining Hands to Build a Just and Equitable System For Global AI Governance — Ministry of Foreign Affairs of the People’s Republic of China — Jul 2026] Official source

The institutional dimension is equally material, because representatives of 29 countries signed the agreement establishing the World Artificial Intelligence Cooperation Organization, or WAICO, on 16 July 2026, with Shanghai designated as headquarters and the organization described in the agreement as an independent intergovernmental international organization intended to promote AI cooperation and governance. [Signing Ceremony of the Agreement on the Establishment of the World Artificial Intelligence Cooperation Organization Held in Shanghai — Ministry of Foreign Affairs of the People’s Republic of China — Jul 2026] Official source

This is not, however, evidence that Washington has remained focused narrowly on GPUs or benchmark leadership, because the United States’ July 2025 AI Action Plan explicitly states that American strategy should promote international adoption of the full US AI technology stack—hardware, models, software, applications and standards—and warns that failure to satisfy foreign demand would encourage partner countries to adopt rival technologies. [America’s AI Action Plan — The White House — Jul 2025] Official source

The European Union has likewise moved beyond regulation alone, because its International Digital Strategy combines international digital partnerships, global governance and standard-setting with a planned EU Tech Business Offer, secure connectivity, digital public infrastructure, cybersecurity and external deployment of AI-related capabilities, while Global Gateway provides an investment architecture whose stated ceiling is up to €300 billion across digital, energy, transport, health, education and research. [The International Digital Strategy for the European Union — European Commission — Jun 2025] Official source [Global Gateway — European Commission — current framework] Official source

The strategic contest therefore concerns neither models alone nor training programs alone, but the ability to connect compute, models, infrastructure, local applications, skills, finance, standards and institutional relationships into systems that partner countries continue using after the original intervention has ended, and the balance of evidence indicates that China is deliberately attempting to create such a system while the United States and European Union are pursuing competing ecosystem strategies of their own.

Saudi mBridge Withdrawal: SAMA Exits China-Led CBDC Platform Amid Geopolitical Scrutiny

China’s AI Challenge Is Moving From Models to Markets

The next phase of the artificial-intelligence contest will not be decided by which laboratory leads a benchmark for six months, but by which economic system persuades governments, engineers and companies to build on its technology for a decade. China has begun converting that proposition into policy: in July 2026, 29 countries signed the agreement establishing the World Artificial Intelligence Cooperation Organization, headquartered in Shanghai, while President Xi Jinping announced 5,000 AI training and seminar opportunities over five years, cooperation centres linked to six major regional groupings and deployment of the MAZU meteorological system across 30 countries. The United States and Europe understand the same contest, but their response is structurally different: Washington is mobilising an exceptionally powerful private technology stack, while Europe is still trying to connect regulation, compute, capital, energy and development finance into a deployable external offer.

Beijing is competing for the users who will choose the next stack

China’s proposition to developing economies is no longer limited to exporting telecommunications equipment or selling individual digital products, because the AI Capacity-Building Action Plan for Good and for All, the Global AI Governance Action Plan of July 2025, the AI Plus International Cooperation Initiative of September 2025 and WAICO now connect infrastructure, models, applications, open-source ecosystems, data, multilingual corpora, testing, certification and training. That architecture matters because technology markets are rarely won exclusively by the technically superior product: they are won when users, universities, ministries and developers accumulate enough knowledge and complementary investment around one environment that moving elsewhere becomes expensive.

The training component illustrates the mechanism more clearly than the headline rhetoric. China’s five-year commitment implies 5,000 opportunities, while a September 2026 programme at Peking University brought senior officials and core AI policymakers from 37 Global South countries into a six-day curriculum containing four thematic areas, nine lectures, six field visits and one roundtable. The strategic value is not the seminar itself but the institutional familiarity it creates among officials who later influence public procurement, national AI strategies, data regulation and infrastructure choices.

The number that matters is not model size but switching cost

The AI economy rests on a stack extending from accelerators and data centres through cloud orchestration, models, middleware, databases, applications, cybersecurity, standards and skilled labour, which means apparent freedom at the model layer can coexist with deep dependence elsewhere. A ministry may replace one language model in months while remaining tied for years to the same cloud environment, accelerator fleet, identity system, databases, security architecture and trained workforce; once those investments accumulate, technical compatibility becomes an economic fact.

China’s standardisation programme gives this strategy a longer horizon. Beijing set a target of more than 50 new national and industry AI standards by 2026 and participation in more than 20 international AI standards, while specific standards now cover generative-AI security, synthetic-content labelling and large-model testing. These instruments do not automatically become global rules, but standards matter because they determine how systems are evaluated, certified and procured after individual models have already changed generation.

The same logic explains why China’s cooperation on local datasets and multilingual corpora deserves more attention than another model release. A country whose public services operate in underrepresented languages requires annotated corpora, evaluation benchmarks, data pipelines and local engineers before sophisticated models can be deployed reliably; whoever helps build those assets enters the national AI architecture at a deeper level than a company merely selling temporary API access.

Washington has recognised that technological leadership is not enough

The American response now accepts the same competitive logic. Executive Order 14320 of July 2025, promoting the export of the American AI technology stack, explicitly includes AI-optimised hardware, data-centre infrastructure, cloud services, networking, data pipelines, models, applications and cybersecurity, demonstrating that Washington no longer assumes frontier-model leadership will automatically produce international market dominance.

The United States enters this contest with an advantage China and Europe cannot easily reproduce: American companies already operate at exceptional scale across semiconductors, hyperscale cloud, foundation models, enterprise software, cybersecurity and developer tooling. The weakness is the same concentration that creates the advantage, because governments seeking technological sovereignty can regard reliance on a small number of US providers as an exposure to foreign jurisdiction, export restrictions or future pricing power.

This is why American semiconductor policy contains a structural tension. Restricting access to advanced hardware can slow competitors’ technical progress, but every country prevented from building on US technology acquires a stronger incentive to invest in alternative accelerators, software and cloud infrastructure; export control therefore protects one layer of American advantage while potentially encouraging substitution across another.

Europe has capital and rules, but not yet one external machine

Europe’s response has changed materially since the debate was dominated by the AI Act. By April 2026, the European Commission reported 19 AI Factories deployed and 13 regional AI Factory antennas, while approximately €10 billion was associated with AI Factory infrastructure for 2021–2027 and a separate €20 billion investment facility was intended to support up to five AI Gigafactories. The Commission’s International Digital Strategy and proposed EU Tech Business Offer add an external dimension covering AI, secure infrastructure, digital public infrastructure, cybersecurity, market studies, pilot projects and financing.

The problem is not lack of instruments but institutional conversion. France, Germany, Italy, the European Commission, EuroHPC, Global Gateway, national development institutions and private companies control different parts of the offer, while the United States can frequently deliver technology through firms already operating globally and China can connect diplomatic and industrial policy through a more centralised state architecture.

The financing base is nonetheless substantial. Global Gateway seeks to mobilise up to €300 billion, while the wider G7 Partnership for Global Infrastructure and Investment targets up to US$600 billion by 2027; neither amount is an AI budget, but both can finance electricity, connectivity, data centres and digital infrastructure without which an international AI strategy remains rhetorical.

France, Germany, Britain and Italy are building different pieces of the European answer

France is pursuing the clearest compute-and-energy strategy. At the February 2025 AI Action Summit, President Emmanuel Macron announced more than €109 billion in prospective French and foreign AI investment commitments, while also stating that France wanted to increase annual AI-related training from approximately 40,000 people to 100,000. The calculation is straightforward: relatively abundant low-carbon electricity attracts data centres, data centres attract compute, compute attracts models and applications, and talent determines whether the resulting infrastructure produces domestic value rather than merely hosting foreign workloads.

Germany is building from a different asset base. Its September 2026 data-centre strategy seeks to at least double national data-centre capacity by 2030, while government policy also targets at least a fourfold increase in connection capacity dedicated to high-performance computing and AI. Germany’s comparative advantage is industrial demand in manufacturing, mobility, chemicals and engineering, but its constraint is equally physical: AI infrastructure requires affordable electricity, grid access and rapid permitting at a time when the German energy system is already carrying a demanding transition.

The United Kingdom has adopted the most aggressive investment-attraction model outside the United States. By January 2026, the government reported 38 of 50 AI Opportunities Action Plan commitments completed, £68 billion in investment pledged, five AI Growth Zones, approximately 200,000 students in AI-related higher-education programmes and an increase in government-referenced AI compute capacity from 2 to 21 ExaFLOPs between 2024 and 2025. Britain is therefore choosing integration with frontier US companies while attempting to preserve domestic leverage through sovereign compute, its AI Security Institute and the June 2026 AI Hardware Plan.

Italy operates at a smaller capital scale but has developed one of Europe’s more interesting external instruments. The IT4LIA AI Factory in Bologna carries a 50 Exaflop/s target, while the G7-backed AI Hub for Sustainable Development, developed with UNDP and connected to the Mattei Plan, reported by March 2026 130 African innovators engaged, more than 150 Italian organisations involved, more than 30 partnerships, 1.5 million GPU-hours mobilised and US$3 million in cloud credits allocated. Seven additional partnerships were signed in Bologna in June 2026 across agriculture, health, linguistic AI, education and infrastructure, giving Italy a potentially valuable role as a bridge between European compute and African application ecosystems.

The contest is becoming one of default systems rather than national champions

The emerging competition is therefore not simply OpenAI against DeepSeek, NVIDIA against Chinese accelerators or Silicon Valley against Shenzhen, because countries can combine American hardware, Chinese open models, European regulation, domestic data centres and locally developed applications within the same system. The strategically decisive question is which layer becomes the bottleneck: if five models can run only economically on one accelerator architecture, compute remains the dependency; if hardware is diversified but all government workflows sit inside one cloud environment, cloud and data architecture become decisive.

This is why the concept of “digital sovereignty” should be treated as an engineering and economic condition rather than a slogan. Sovereignty depends on whether workloads can move, whether data can be exported, whether models can be substituted, whether local engineers can maintain systems independently, whether cybersecurity controls survive migration and whether standards permit alternative suppliers. A country can own its data centre while remaining dependent on foreign accelerators, proprietary software and external updates; conversely, it can use foreign technology while preserving substantial autonomy through diversified suppliers and enforceable portability.

China’s greatest potential advantage is therefore not necessarily that it will produce the strongest model, but that it is systematically offering partner states several of the complementary goods required to make AI usable: training, public applications, infrastructure cooperation, local-language resources, standards dialogue and institutional relationships. Washington’s answer is a commercially powerful full-stack export strategy; Europe’s answer is becoming a federation of compute, regulation, national industrial strengths and development finance.

The next 24 months will decide whether Europe converts assets into market share

Between late 2026 and 2028, the critical European test will be whether the 19 AI Factories, planned Gigafactories, national data-centre programmes and EU Tech Business Offer become an integrated deployment system before competing ecosystems become embedded in third-country procurement and technical training. France must convert a material share of the €109 billion announced investment into commissioned infrastructure; Germany must turn its 2030 doubling and quadrupling targets into power-connected capacity; Britain must convert £68 billion of pledged investment into installed compute rather than dependency alone; and Italy must connect the AI Hub’s 1.5 million GPU-hours and African partnerships to EuroHPC and larger European financing.

The cost of inaction will fall first on European technology firms that lose developer and procurement relationships before their infrastructure reaches scale, then on governments forced to purchase critical AI capabilities from external platforms, and finally on industrial sectors whose data, software and workforce become progressively harder to migrate. China does not need to displace every American or European technology company to alter the market; it needs enough governments, engineers and institutions to regard its ecosystem as one of the default places to build. Once that happens, recovering market share becomes an infrastructure problem rather than a marketing problem.

China’s Global Technology Supply Chains: State-Backed Resource, Manufacturing and Logistics Networks Behind Strategic Industrial Expansion

---

Navigational Index

Institutional architecture and the Global South — How WAICO, capacity-building programs, regional AI cooperation centres and application deployment convert diplomatic relationships into recurring technological cooperation.

Technology stacks, standards and dependency — Why the strategically relevant unit of competition is increasingly the integrated ecosystem surrounding AI rather than the isolated frontier model, and why dependency can arise at the level of cloud, compute, standards, skills, software and data architecture.

Western response and European exposure — How the United States, European Union, United Kingdom, France, Germany and Italy are constructing alternative combinations of infrastructure, investment, regulation, standards and international partnerships, and where implementation differences remain visible.

---

Master Abstract

Beijing is converting AI policy into institutional infrastructure

China’s current international AI strategy should be understood as the accumulation of several initiatives rather than as a single announcement at WAIC 2026, because Beijing’s official record shows a sequence beginning with the Global AI Governance Initiative, continuing through the UN capacity-building agenda and the Global AI Governance Action Plan issued in July 2025, followed by the “AI Plus” international-cooperation initiative and culminating, so far, in the establishment of WAICO and the capacity-building package announced in Shanghai in July 2026. The Chinese Foreign Ministry itself presents these elements as a cumulative governance architecture, while Xi’s WAIC address explicitly frames AI as an international public good and connects international cooperation with governance rules, technical standards, development strategy and Global South capacity-building. [Global AI Governance Action Plan — Ministry of Foreign Affairs of the People’s Republic of China — Jul 2025] Official record [Joining Hands to Build a Just and Equitable System For Global AI Governance — Ministry of Foreign Affairs of the People’s Republic of China — Jul 2026] Official speech

The institutional significance of WAICO should nevertheless be described carefully, because the verified fact is that representatives of 29 states signed the establishment agreement and became founding members, whereas the long-term breadth, authority, financing mechanisms, voting rules and operational effectiveness of the organization cannot yet be inferred from the signing ceremony alone. The official agreement summary states that WAICO is intended to operate as an independent intergovernmental international organization headquartered in Shanghai, while the WAIC chair’s statement describes it as a mechanism intended to strengthen innovation, application and governance capacity, particularly among Global South countries, and to remain open to all countries. [Chair’s Statement of the 2026 World Artificial Intelligence Conference & High-Level Meeting on Global AI Governance — Ministry of Foreign Affairs of the People’s Republic of China — Jul 2026] Official source

Human capital is an instrument of ecosystem formation, but the scale should not be overstated

The announced 5,000 training and seminar opportunities for developing countries during the five years beginning in 2026 are strategically relevant because training engineers, regulators and public-sector officials can create durable familiarity with particular technological architectures and governance concepts, yet the number itself does not justify claims that China has already built a mass international AI-training system numbering in the millions. The first visible implementation evidence appeared in September 2026, when an AI capacity-building seminar opened at Peking University with senior government officials and AI policymakers from 37 Global South countries, demonstrating that the July commitment had begun to move from announcement into implementation, although one seminar does not establish the eventual geographic distribution, completion rate or downstream technology adoption associated with the full five-year program. [37国代表来了！5000个AI培训名额加速兑现——人工智能能力建设研讨班在京开班 — Beijing Municipal Science and Technology Commission — Sep 2026] Official source

The training program also sits inside a broader multilateral lineage rather than existing solely as bilateral Chinese outreach, because the United Nations General Assembly adopted Resolution A/RES/78/311, “Enhancing international cooperation on capacity-building of artificial intelligence,” without a vote on 1 July 2024, following a draft sponsored by China and a group of developing states, which means Beijing has succeeded in moving at least part of its preferred capacity-building agenda through an established UN framework rather than relying exclusively on China-created institutions. [Enhancing international cooperation on capacity-building of artificial intelligence — United Nations General Assembly — Jul 2024] Official UN record

The more consequential architecture joins skills to deployed applications

The announced cooperation centres involving ASEAN, the League of Arab States, African Union, CELAC, Shanghai Cooperation Organization and BRICS potentially matter more strategically than the absolute number of training places because regional institutional structures can create channels through which training, local model adaptation, cloud services, public-sector AI applications, standards discussions and financing can subsequently be combined, although the public record reviewed here does not yet establish the organizational form, host locations, budgets or technology vendors associated with every proposed centre. Xi explicitly announced these six institutional tracks in July 2026, meaning that their existence as policy commitments is established, while their effectiveness remains an empirical question for subsequent monitoring. [China to provide 5,000 AI training opportunities for developing countries: Xi — State Council of the People’s Republic of China — Jul 2026] Official source

MAZU illustrates the application layer of this strategy because China has stated that the AI-enabled meteorological early-warning system will be used in 30 countries, thereby presenting AI not as an abstract frontier-model technology but as infrastructure attached to climate and disaster-risk management, while later Chinese official reporting described the system as having entered application in those countries. This is strategically important because practical public-sector applications can generate demand for data integration, local technical training, maintenance and institutional interoperability, although the present official evidence is insufficient to establish whether those deployments require continuing dependence on Chinese cloud services, proprietary infrastructure or Chinese model providers. [中国经济面面观｜中国方案，如何让AI行稳致远 — Cyberspace Administration of China — Aug 2026] Official source

Pinglu Canal: China’s New Southwest–ASEAN Trade Artery and the Reordering of Regional Economic Geography

The “Digital Belt and Road for AI” analogy is analytically useful only with qualifications

The comparison with the Belt and Road Initiative becomes analytically meaningful when it refers to the construction of durable cross-border networks that combine infrastructure, financing, skills, standards and institutional relationships, but it becomes misleading if it suggests that China already possesses a centrally integrated international AI architecture equivalent in physical scale or contractual depth to transport, energy and telecommunications projects associated with earlier BRI phases. What is visible in 2026 is instead an emerging architecture in which China is attempting to connect governance institutions, training, open technology ecosystems, regional cooperation centres and application-level public goods, while the financial scale, vendor composition and contractual persistence of those relationships remain incompletely documented in the available first-order record.

The central strategic mechanism nevertheless deserves attention because standards and professional training can create path dependence even without formal exclusivity, since engineers trained on particular software stacks, ministries accustomed to certain technical interfaces, universities teaching specific development tools and governments adopting compatible data architectures acquire switching costs that can shape later procurement decisions. Xi’s 2026 speech explicitly called for closer coordination of AI development strategies, governance rules and technical standards while simultaneously expanding Global South capacity-building, which places standards and skills within the same official strategic framework rather than leaving their connection entirely to external interpretation. [习近平在2026世界人工智能大会暨人工智能全球治理高级别会议开幕式上的主旨讲话（全文） — Cyberspace Administration of China — Jul 2026] Official Chinese text

Washington is pursuing an explicit counter-ecosystem strategy

The proposition that the United States remains preoccupied only with GPUs, chips or benchmark supremacy is contradicted by the US government’s own published strategy, because America’s AI Action Plan states that international competition requires global adoption of US systems, hardware and standards and explicitly calls for exporting complete AI packages rather than individual products. Executive Order 14320 subsequently established a federal policy to support the international deployment of US-origin AI technology and decrease dependence on technologies developed by strategic competitors, thereby making technology-stack adoption and international dependency explicit components of US policy rather than merely implicit commercial effects. [Promoting the Export of the American AI Technology Stack — The White House — Jul 2025] Executive Order 14320

The resulting US-China competition should therefore be described as a contest between different ecosystem-development strategies rather than a contrast between an ecosystem-oriented China and a model-oriented United States, because Washington’s official documents seek international adoption of a US-controlled technological stack while Beijing combines capacity-building, cooperation institutions, technical coordination and Global South development language. The strategic distinction lies less in whether each side understands ecosystem power than in which combination of commercial firms, state financing, standards, infrastructure, open technologies, diplomatic institutions and development assistance each side can deploy consistently across third countries.

Europe is not absent, but its external AI architecture is institutionally distributed

The European Union’s June 2025 International Digital Strategy formally commits the Union to expanding Digital Partnerships, developing an integrated EU Tech Business Offer, supporting partner countries’ digital transitions and strengthening global digital governance, while the European Commission states that AI, secure connectivity, digital public infrastructure, cybersecurity, semiconductors and other emerging technologies all fall within this external strategy. [Joint Communication on an International Digital Strategy for the EU — European Commission — Jun 2025] Official source

The EU therefore has the conceptual components required for an ecosystem strategy, while implementation is distributed across the European Commission, AI Office, Global Gateway, EuroHPC, Member States, development-finance institutions and private firms rather than being concentrated within a single foreign-policy initiative. The European Commission reports that by April 2026 Europe had 19 AI factories deployed, supported by 13 regional AI Factory antennas, while InvestAI was designed to mobilise €200 billion in overall AI investment and a dedicated €20 billion facility for up to five AI Gigafactories, showing that European strategy combines compute infrastructure with industrial policy and international engagement rather than regulation alone. [AI Continent Action Plan delivers major milestones — European Commission — Apr 2026] Official source [Memorandum of Understanding on AI Gigafactories — European Commission — Dec 2025] Official source

China’s Humanoid-Robot Industrial Offensive

Italy, France, Germany and the United Kingdom enter the contest through different institutional assets

Italy’s national strategy places research, public administration, enterprises and training among its four principal areas of intervention, while also emphasizing country-specific data and models capable of preserving Italian economic and cultural characteristics, which creates a potentially important national complement to the EU’s external strategy even though the 2024–2026 national document is primarily domestic rather than a dedicated Global South AI-diplomacy program. [The Italian Strategy for Artificial Intelligence 2024-2026 — Agenzia per l’Italia Digitale — Jul 2024] Official source

France possesses a different external asset in its convening power and AI diplomacy, illustrated by the February 2025 Paris AI Action Summit, which brought together participants from more than 100 countries under French-Indian co-chairmanship and was explicitly framed around international action, public-interest AI and the strengthening of France and Europe’s position in global AI competition. [Sommet pour l’action sur l’intelligence artificielle — Présidence de la République française — Feb 2025] Official source

Germany’s strategy emphasizes industrial AI, sovereign access to AI infrastructure, international standards and resilient digital supply chains, while the federal government’s International Digital Policy Strategy explicitly supports values-based technology partnerships, international standard-setting, secure global digital infrastructure and the reduction of critical technology dependencies. [Strategie für die Internationale Digitalpolitik — Bundesregierung — 2024 framework] Official source

The United Kingdom’s AI Opportunities Action Plan identifies AI adoption, domestic compute, data access and talent as strategic requirements while acknowledging explicitly that the United Kingdom faces competitive pressure from both the United States and China; its institutional advantage lies partly in the density of its research ecosystem and AI-safety infrastructure, although the plan itself does not constitute an international development architecture comparable with the specific training and regional-cooperation initiatives announced by Beijing. [AI Opportunities Action Plan — UK Department for Science, Innovation and Technology — Jan 2025] Official source

Key Evidence Table

Indicator	Value/status	Reference date	Definition/scope	Issuer	Exact source
WAICO founding signatories	29 countries	16 Jul 2026	Countries whose representatives signed establishment agreement	PRC Ministry of Foreign Affairs	[Signing Ceremony of the Agreement on the Establishment of the World Artificial Intelligence Cooperation Organization Held in Shanghai — Jul 2026] Official source
WAICO headquarters	Shanghai	Jul 2026	Headquarters specified in establishment agreement summary	PRC Ministry of Foreign Affairs	[Official establishment record] Source
AI training commitment	5,000 opportunities	2026–2031	Training and seminar opportunities for developing countries	President Xi Jinping / PRC	[China to provide 5,000 AI training opportunities for developing countries: Xi] Source
Cooperation-centre tracks	6 institutional groupings	Jul 2026	ASEAN, Arab League, AU, CELAC, SCO and BRICS	President Xi Jinping / PRC	[Xi keynote address] Source
MAZU deployment target	30 countries	Announced Jul 2026	AI meteorological early-warning applications	PRC	[Xi keynote address] Source
First visible 2026 training cohort	Officials from 37 countries	7 Sep 2026	Six-day AI capacity-building seminar at Peking University	Beijing Municipal Science and Technology Commission	[Official seminar record] Source
US international AI objective	Full-stack export	Jul 2025	Hardware, models, software, applications and standards	White House	[America’s AI Action Plan] Source
EU international architecture	Tech Business Offer + partnerships	Jun 2025 onward	AI, connectivity, DPI, cybersecurity and digital partnerships	European Commission / High Representative	[International Digital Strategy] Source
EU Global Gateway	Up to €300bn	Current strategy	Cross-sector infrastructure mobilisation target, not an AI-only budget	European Commission	[Global Gateway] Source

Principal Gaps and Watch Indicators

The decisive uncertainty is not whether China has announced an international AI ecosystem strategy, because that proposition is established by multiple first-order records, but whether the announced architecture generates durable technology adoption, procurement relationships and institutional dependence at a scale that materially changes market structure; the principal evidence required to answer that question will therefore be country-level implementation agreements specifying cloud providers, model suppliers, compute infrastructure, financing, data-hosting requirements, software licensing, standards commitments and local-training arrangements.

A second watch indicator concerns WAICO itself, because the strategic significance of the organization will depend on its ratification status, governance rules, budget, technical committees, membership growth, standard-setting or recommendation authority, relationship with United Nations bodies and ability to finance or coordinate real projects, whereas the existing establishment documents prove institutional creation but do not yet establish how much operational authority the organization will exercise. [Chair’s Statement of the 2026 World Artificial Intelligence Conference & High-Level Meeting on Global AI Governance — Ministry of Foreign Affairs of the People’s Republic of China — Jul 2026] Official source

A third indicator is the technology composition of Chinese overseas AI projects, because the strategic consequences differ substantially depending on whether partner countries receive open-source models capable of running on heterogeneous hardware, Chinese proprietary cloud services with persistent vendor dependency, domestically hosted systems under local control, or hybrid architectures combining Chinese models with non-Chinese compute and local data governance.

A fourth indicator concerns human-capital persistence, because the 5,000-place commitment becomes strategically important only to the extent that participants subsequently occupy ministries, universities, state-owned enterprises, regulatory agencies or technology firms capable of influencing national procurement, standards and training systems, making alumni networks, follow-on technical cooperation and institutional placement more informative than the headline number alone.

A fifth indicator concerns competitive implementation by the United States and European Union, because Washington’s full-stack export policy and Brussels’ Tech Business Offer are already documented counter-strategies, while the strategic balance will depend on whether those policies generate actual packages involving affordable compute, finance, local skills and public-sector applications in developing markets rather than remaining principally policy frameworks. [Promoting the Export of the American AI Technology Stack — The White House — Jul 2025] Official source [The EU sets out its International Digital Strategy — European Commission — Jun 2025] Official source

Net Assessment

The most defensible formulation of the thesis is therefore not that China has discovered a form of AI competition that the United States and Europe have overlooked, because both Washington and Brussels explicitly recognize the strategic importance of international adoption, standards, infrastructure and partner ecosystems, but that China is building a distinct state-supported international AI architecture in which governance institutions, Global South capacity-building, training, regional cooperation mechanisms, standards coordination and deployable public-interest applications are increasingly treated as mutually reinforcing components of geopolitical technology policy.

The comparison with earlier Chinese infrastructure diplomacy is consequently useful as a hypothesis about mechanism rather than as a declaration that an “AI Belt and Road” already exists, because the relevant similarity lies in the possibility that infrastructure, technical assistance and training create relationships that persist through standards, maintenance, procurement and professional networks, while the decisive difference is that AI systems can diffuse through software, cloud access, open models and human expertise far faster and with lower physical-capital requirements than ports, railways or telecommunications networks.

For Europe, the resulting issue is not simply whether European laboratories can produce frontier models, because the Union already possesses significant regulatory reach, supercomputing infrastructure, industrial expertise, development finance and international digital partnerships; the more consequential question for the 2026–2031 period is whether those assets can be integrated into a coherent external offer combining compute access, locally relevant models, technical training, affordable financing, digital public infrastructure, standards cooperation and sovereign deployment options, rather than remaining distributed across separate national and European programs.

For the United States, the public record demonstrates that policymakers recognize the same strategic problem and are explicitly seeking worldwide adoption of American AI technology stacks, which means the emerging competition should be understood not as “models versus training” but as ecosystem versus ecosystem, with the principal variables becoming affordability, deployment speed, interoperability, financing, skills transfer, trust, standards influence, local political acceptability and the ability to sustain technical relationships after initial deployment.

Institutional architecture and the Global South

Principal judgment

China’s emerging international artificial-intelligence architecture is becoming more consequential than a conventional technology-export programme because the official record now shows a layered system in which multilateral governance, development diplomacy, infrastructure cooperation, model and application deployment, human-capital formation, data and corpus development, testing and certification practices, standards coordination, and sector-specific demonstration projects are being deliberately connected rather than pursued as isolated initiatives; the strategic significance therefore lies not in any single institution or headline commitment, but in Beijing’s attempt to create recurring institutional relationships through which partner governments can move from introductory AI training to technical cooperation, from technical cooperation to application deployment, and eventually from application deployment to shared standards, evaluation practices, data architectures and procurement ecosystems. China’s 2024 AI Capacity-Building Action Plan for Good and for All, its 2025 Global AI Governance Action Plan, the September 2025 AI Plus International Cooperation Initiative, the establishment process for WAICO in 2026, and the regional application-centre initiative announced at WAIC 2026 together constitute a substantially more developed policy architecture than would be visible from the headline figure of 5,000 training places alone.

The critical analytical distinction is between access diplomacy and ecosystem formation: access diplomacy supplies training, software or infrastructure to partners, whereas ecosystem formation occurs when successive layers of cooperation begin influencing which experts are trained, which applications ministries deploy, which data formats and testing practices become familiar, which technical standards are considered normal, which local institutions maintain relationships with foreign suppliers, and which technological options acquire lower political and transaction costs than their alternatives; the Chinese documents reviewed do not prove that such lock-in has already occurred across the Global South, but they do establish that China is assembling virtually all of the institutional mechanisms through which such path dependence can develop.

From diplomatic initiative to an institutional production line

The most important evolution since 2024 has been the conversion of a broad governance narrative into a sequence of increasingly operational instruments, because China first worked through a universally recognised multilateral venue, subsequently created its own capacity-building programme, broadened that programme into industrial and public-service applications, and finally established a dedicated intergovernmental organization capable, at least in principle, of coordinating implementation continuously rather than through occasional conferences. The United Nations General Assembly adopted Resolution A/RES/78/311 on 1 July 2024 without a vote, after China and 27 other named states submitted draft A/78/L.86, while the resolution itself identified insufficient digital infrastructure, education, expertise and human capacity as constraints preventing developing countries from participating fully in AI development and explicitly referred to financing, knowledge sharing, technology transfer on mutually agreed terms and capacity-building assistance as means of addressing the gap.

China then used that multilateral policy foundation to announce the AI Capacity-Building Action Plan for Good and for All, whose architecture is substantially broader than conventional training assistance because it includes digital and AI infrastructure connectivity, AI model and application cooperation, open-source communities, education and professional training, public AI literacy, multilingual corpora, data infrastructure, regulatory exchanges, testing and certification, and strategic-policy coordination. The official plan therefore describes an integrated development stack in which physical and digital infrastructure sit at the bottom, models and data form the technical layer, skills and institutions constitute the human layer, and standards, testing and governance occupy the policy layer, creating multiple opportunities for repeated interaction between Chinese organizations and foreign ministries, universities, regulators, firms and technical agencies.

Institutional evolution of China’s Global South AI architecture

Stage	Official instrument or mechanism	Date	Operational function	Strategic relevance
Multilateral legitimacy	UNGA Resolution A/RES/78/311	1 Jul 2024	Establishes international capacity-building as a recognised AI-governance priority and identifies infrastructure, skills, financing and technology access as developing-country constraints	Gives Beijing’s capacity-building agenda a UN-level normative foundation rather than leaving it solely within Chinese bilateral diplomacy.
Programmatic architecture	AI Capacity-Building Action Plan for Good and for All	Sep 2024	Connects infrastructure, models, open-source ecosystems, professional training, public literacy, data, corpora, assessment and regulation	Converts abstract inclusion language into a portfolio of concrete cooperation instruments.
Initial implementation	Workshops and seminars focused on developing countries	2024–2026	Trains policymakers and technical officials while exposing them to Chinese research, industrial and governance practices	Creates professional networks and recurring institutional contact, although long-term downstream adoption remains unproven.
Application doctrine	Global AI Governance Action Plan	Jul 2025	Calls for infrastructure, joint laboratories, interoperable systems, industry matching, datasets, corpora, safety assessment and training	Expands cooperation from education toward applied technological and regulatory integration.
Sectoral internationalisation	AI Plus International Cooperation Initiative	Sep 2025	Extends AI cooperation into healthcare, education, science, industry, culture and talent formation	Creates sector-specific pathways through which cooperation can become embedded in public services and productive systems.
Dedicated institution	WAICO establishment agreement	Jul 2026	Creates an independent intergovernmental organization headquartered in Shanghai	Provides a permanent institutional venue capable of coordinating cooperation beyond one-off diplomatic initiatives.
Regionalisation	Application cooperation centres with six major regional organizations/groupings	Announced Jul 2026	Intended to connect AI applications with ASEAN, Arab League, AU, CELAC, SCO and BRICS structures	Potentially shifts implementation from bilateral projects toward regionally networked cooperation.
Demonstration deployment	MAZU meteorological early-warning system	2026	Supplies a practical AI-enabled public-service application and related meteorological cooperation	Demonstrates how AI diplomacy can enter operational public infrastructure rather than remaining confined to seminars or policy declarations.

WAICO is designed as an institutional connector rather than merely another conference mechanism

The formal importance of WAICO rests less on its founding membership number than on the functions that Chinese official documents attach to the organization, because the National Development and Reform Commission states that the platform is intended to advance member-state AI capacity building, coordinate development strategies, governance rules and technical standards, promote practical cooperation and applications, and contribute to the formation of a broader governance framework, while the chair’s statement from WAIC describes the organization as open to all countries and envisages communication and cooperation with other international organizations. Those stated functions place WAICO potentially at the junction between development assistance, industrial cooperation and norm formation, although the public record available as of September 2026 does not yet establish detailed voting rules, binding standard-setting powers, permanent budget, secretariat staffing levels, procurement mechanisms or enforcement authority.

The July 17, 2026 WAICO follow-up meeting convened jointly by China’s National Development and Reform Commission and Ministry of Foreign Affairs is significant because discussions with signatory-state representatives covered three categories simultaneously—global governance, practical industrial cooperation, and the organization’s construction and operation—showing that Beijing is not publicly defining WAICO as a purely normative forum but as an institution intended to link governance with practical economic activity. The official NDRC account further records that China proposed work recommendations for members and stated its willingness to share domestic AI development outcomes within the organization, while representatives discussed participation according to their countries’ particular development conditions, thereby indicating an intended model based on differentiated national participation rather than a single standardized programme.

WAICO’s currently documented institutional functions

Function	Officially documented objective	Potential channel of influence	What is not yet publicly established
Capacity building	Strengthen member-state AI capabilities	Training, institutional exchange, technical assistance	Country allocations, budgets, delivery organizations and measurable outcomes.
Development-strategy coordination	Align national AI-development approaches	Policy dialogue and development planning	Whether alignment will produce formal policy commitments or remain voluntary.
Governance-rule coordination	Increase compatibility among AI-governance approaches	Regulatory exchange, testing practices and institutional learning	Any binding harmonisation mechanism.
Technical-standard coordination	Promote alignment on technical standards	Standards forums, technical committees and interoperability discussions	Formal committee architecture and relationship with ISO, IEC, ITU or other bodies.
Application cooperation	Support practical AI deployment	Public-sector and industrial demonstration projects	Vendor selection rules, financing models and intellectual-property conditions.
Global South representation	Strengthen the voice and participation of developing countries	Membership, diplomacy and multilateral negotiating coalitions	Whether WAICO members will develop coordinated negotiating positions in external institutions.
Institutional interoperability	Maintain contact with other international organizations	Cooperation with UN-linked and regional mechanisms	Formal memoranda, observer arrangements and division of institutional competence.

The absence of public evidence concerning WAICO’s budget, permanent staffing, decision-making structure and technical organs is strategically important rather than a minor documentation gap, because those variables will determine whether WAICO becomes a diplomatic platform comparable to numerous existing technology forums or develops the administrative capacity to coordinate multi-year technical programmes, channel financing, organize standardization work and maintain permanent country relationships; accordingly, the signing of an establishment agreement should be treated as institutional creation, not yet as proof of operational power.

The ten-action architecture reveals a much broader strategy than training alone

China’s 2024 capacity-building plan is especially important because it specifies the practical mechanisms through which the Global South relationship is intended to operate, and several of those mechanisms reach directly into areas that shape technological ecosystems over long periods, including infrastructure, model development, open-source communities, joint educational programmes, multilingual corpora, data infrastructure, testing, certification and regulation.

The documented capacity-building stack

Cooperation layer	Documented Chinese commitment	Why it matters institutionally
AI and digital infrastructure	Cooperate on AI infrastructure and improve globally interoperable AI and digital infrastructure	Infrastructure decisions determine which compute, networking, cloud and data systems applications can subsequently use.
Models and applications	Cooperate on model R&D and AI-enabled applications	Moves cooperation upstream from software consumption toward development and adaptation.
Open-source ecosystems	Build open and inclusive AI communities and share best practices	Lowers initial adoption barriers and can create developer familiarity before commercial procurement decisions arise.
Professional education	Short- and medium-term education, joint schooling and exchanges	Builds technical and administrative human capital rather than supplying only finished products.
Human-resource assistance	Dedicated workshops and seminars for developing countries	Creates repeated official-to-official and institution-to-institution relationships.
Public AI literacy	Online and offline dissemination of AI knowledge	Extends engagement beyond elite technical communities and supports broader adoption.
Language resources	Joint development of AI corpora and measures against cultural and linguistic discrimination	Directly addresses one of the largest structural weaknesses facing AI deployment in low-resource languages.
Data infrastructure	Promote and improve data infrastructure and equitable data utilization	Data architecture determines whether local models and services can be trained, evaluated and deployed effectively.
Testing and certification	Exchange policies and technical practices covering testing, assessment and certification	Creates a pathway through which regulatory practices and technical assurance methods can converge.
Strategic and regulatory dialogue	Coordinate AI strategies and exchange regulatory experience	Moves cooperation from individual projects toward national-policy formation.

This structure matters because technological dependence does not require a government to purchase a single vertically integrated foreign platform; dependence can emerge incrementally when the same ecosystem supplies technical training, model-development tools, datasets, certification procedures, cloud or compute interfaces, domain applications and policy expertise, thereby reducing the friction associated with continuing to use compatible solutions while increasing the cost of migrating to alternatives, although no public official dataset currently permits a defensible calculation of such switching costs across Chinese partner states.

Training is targeted increasingly at decision-makers, not merely programmers

The September 2026 Beijing capacity-building seminar provides unusually useful implementation evidence because it identifies both the institutional status of participants and the structure of the curriculum: the programme assembled senior government representatives and core AI policymakers from 37 Global South countries, lasted six days, and included four thematic areas, nine specialist lectures, six field visits and one roundtable, with content spanning AI safety and governance, frontier technologies, application practices and international cooperation. The programme was sponsored by the Ministry of Foreign Affairs, hosted by Peking University, and jointly implemented with the Beijing Municipal Science and Technology Commission, Zhongguancun authorities and Peking University’s School of International Studies, while site visits included the Beijing Institute for General Artificial Intelligence, Zhongguancun Exhibition Center, a humanoid-robot training centre and the Shougang Park science-fiction industry cluster.

The composition of this programme makes the institutional objective materially different from a conventional coding boot camp, because senior officials and policy architects are positioned to influence national AI strategies, procurement procedures, regulatory choices, public-sector digitalization and subsequent international partnerships, while field visits expose them not only to individual technologies but to an integrated Chinese innovation ecosystem encompassing laboratories, industrial clusters, robotics, applied research and governance practices. This does not demonstrate that participants subsequently adopt Chinese systems, but it creates a credible mechanism through which familiarity, professional networks and perceptions of technological feasibility can influence later policy choices, which makes alumni trajectories and follow-on institutional agreements one of the most important empirical indicators to monitor.

Evidence of programme maturation

Indicator	Early implementation	2026 implementation	Analytical significance
Participation	First UN-China capacity-building workshop drew participants from nearly 40 countries after adoption of A/RES/78/311	September 2026 seminar assembled officials and policymakers from 37 Global South countries	Geographic breadth existed before the five-year commitment and is becoming institutionalized.
Programme structure	Initial workshops emphasised policy exchange and capacity building	9 lectures + 6 visits + 1 roundtable over 6 days	Training now combines classroom instruction, industrial exposure and diplomatic networking.
Participant profile	Mixed international capacity-building participants	Senior government officials and core AI policymakers	The programme increasingly reaches actors capable of shaping national policy rather than only technical specialists.
Institutional hosts	China and UN Secretariat cooperation in early workshop	MFA, Peking University, Beijing municipal technology authorities and research/industry institutions	Delivery increasingly connects foreign-policy institutions with domestic innovation centres.
Five-year scale	Ten additional seminars were announced for completion by end-2025	5,000 training and seminar opportunities announced for the subsequent five-year period	Capacity building has moved from discrete workshops toward a multiyear pipeline.

A useful quantitative qualification is necessary because 5,000 places over five years corresponds arithmetically to an average of approximately 1,000 opportunities per year if distributed evenly, but China has not announced an even annual allocation, country quotas, sectoral allocation, participant-selection rules or the proportion intended for technical professionals versus regulators, public administrators or academics, meaning that any projection of geographic penetration beyond the announced aggregate would currently be unsupported.

Regional application centres create the possibility of network effects

The proposed international AI application cooperation centres associated with ASEAN ( Association of Southeast Asian Nations) , the League of Arab States, the African Union, CELAC, the Shanghai Cooperation Organization and BRICS potentially represent the most important expansion mechanism in the architecture because they introduce a regional layer between bilateral Chinese projects and global governance forums, allowing applications, training and technical practices to be replicated across groups of states rather than negotiated independently with every national government. President Xi’s July 2026 announcement establishes the political commitment to build these centres, but the available first-order record does not yet specify whether each will be a physical institution, distributed network, joint laboratory system or programme housed within existing organizations, nor does it establish budgets, headquarters, partner agencies or implementation timetables.

Regional cooperation architecture and potential institutional reach

Regional grouping	Cooperation track announced by China	Principal institutional logic	Critical implementation evidence still missing
ASEAN	International AI application cooperation centre	Connects Chinese AI cooperation to an established Southeast Asian regional framework with existing economic and digital-policy mechanisms	Location, funding, participating ministries, vendor structure and programme calendar
League of Arab States	International AI application cooperation centre	Creates a mechanism capable of linking AI cooperation across Arabic-speaking states and potentially supporting Arabic-language applications and corpora	Hosting arrangements, technical priorities and relationship with national AI strategies
African Union	International AI application cooperation centre	Offers a continental-level channel for capacity building and applications across countries with highly uneven compute and data capacity	Financing, digital-infrastructure requirements, partner institutions and country participation
CELAC	International AI application cooperation centre	Creates an institutional bridge to Latin America and the Caribbean beyond bilateral projects	Location, governance model, regional membership participation and sector focus
Shanghai Cooperation Organization	International AI application cooperation centre	Connects AI cooperation with a pre-existing Eurasian political and security institution	Scope of civilian versus broader digital-policy cooperation and technical architecture
BRICS	International AI application cooperation centre	Creates another channel among large emerging economies and associated partner states	Relationship with existing BRICS AI, digital-economy and research mechanisms

Source for the six announced cooperation tracks: President Xi Jinping’s 17 July 2026 WAIC address and the official bilingual text of the announcement.

The strategic advantage of the regional model is that one successful public-service application can, in principle, be demonstrated to multiple governments through a common institutional channel, while technical training, standards workshops and supplier engagement can be organized collectively rather than separately, creating lower transaction costs for China and potentially faster diffusion among states facing similar development constraints; nevertheless, until the centres acquire identifiable staff, funding, projects and partner-country participation, they should be assessed as announced institutional infrastructure rather than fully operational networks.

MAZU demonstrates how an application can become a diplomatic platform

The MAZU early-warning system is particularly informative because its underlying product characteristics illustrate how an AI-related international public good can combine a highly visible social function with infrastructure, data and technical cooperation. An April 2026 official meteorological account states that the upgraded cloud-based public version offered 30 major product categories and more than 200 products, incorporated AI extensively, extended services from weather forecasting toward impact forecasting, added overseas cloud-service nodes, improved platform loading speed by a factor of six, and reduced data-response time by a factor of seven, while customized versions were delivered at that event to the meteorological authorities of Jordan and Sri Lanka.

Those technical characteristics are strategically significant because meteorological systems require continuing data ingestion, model updates, local institutional expertise, interfaces with national warning authorities and sustained operational reliability, meaning that a weather-warning platform can create a deeper and more persistent technical relationship than a single software licence; at the same time, the official public record reviewed here does not disclose hosting contracts, source-code arrangements, cybersecurity responsibilities, data-residency rules, service-level agreements or the division between Chinese and local operational control, so claims that MAZU automatically produces technological dependency would exceed the evidence.

MAZU implementation evidence

Metric or feature	Verified status	Institutional implication
Product categories	30 major categories	Indicates a broad operational package rather than a single alert product.
Individual products	More than 200	Supports customization across different meteorological and disaster-risk contexts.
Overseas cloud nodes	Added in 2026 upgrade	Improves accessibility for users outside China and creates a technical delivery architecture for transnational use.
Loading performance	6× improvement, according to the issuing Chinese meteorological authority	Relevant particularly to regions where cross-border connectivity creates latency constraints.
Data-response performance	7× faster, according to the same official source	Improves operational usability for early warning, while independent performance validation is not provided in the cited record.
Customized-country deliveries identified before WAIC	Jordan and Sri Lanka	Provides concrete evidence that international deployment preceded the 30-country political announcement.
Subsequent political objective	Enable use in 30 countries	Converts a technical product into a flagship Global South cooperation programme.

This mechanism is replicable beyond meteorology because China’s 2025 Global AI Governance Action Plan specifically identifies manufacturing, healthcare, education, agriculture, poverty reduction, autonomous driving and smart cities as application domains while calling for cross-border AI application cooperation, intelligent-infrastructure sharing, joint laboratories, industrial supply-and-demand matching and common work on datasets and corpora; the resulting portfolio therefore contains multiple sectors in which a successful demonstration project can create subsequent demand for training, data engineering, maintenance, cloud capacity and regulatory coordination.

“AI Plus” transforms development cooperation into sectoral penetration

The September 2025 AI Plus International Cooperation Initiative expands the architecture beyond technology ministries by targeting public services and economic sectors whose adoption decisions are often made by health ministries, education ministries, industrial agencies, research institutions and cultural bodies, thereby multiplying the number of bureaucratic entry points through which AI cooperation can become institutionalized. The initiative explicitly identifies health applications including disease prevention, drug research, assisted diagnosis and telemedicine; education and equitable educational-resource distribution; AI-enabled scientific research; industrial transformation and employment creation; cultural digitization and preservation; and international talent development, meaning that AI cooperation is being framed as a horizontal development policy rather than a specialist digital-sector programme.

Sectoral application matrix

Sector	Officially identified applications	Potential recurring relationship
Healthcare	Disease prevention, drug R&D, assisted diagnosis, telemedicine	Model validation, clinical data governance, hospital integration, specialist training and continuous software maintenance.
Education	Resource distribution, teaching innovation, education quality	Teacher training, learning platforms, local-language models, content standards and data governance.
Scientific research	AI-assisted scientific discovery and interdisciplinary research	Research networks, compute access, collaborative laboratories and shared technical tools.
Industrial production	Modernization of traditional industries, new industries and applications	Industrial software, robotics, model integration, standards and supply-chain partnerships.
Culture	Digital preservation, intelligent restoration and immersive presentation	Local-language datasets, cultural corpora and media-technology cooperation.
Talent	Shared educational resources and personnel exchanges	Long-term professional networks connecting universities, government and industry.

The strategic consequence is that China does not need every partner country to adopt a Chinese general-purpose frontier model as its national AI platform for the ecosystem strategy to succeed, because sectoral systems can produce durable relationships through hospital networks, universities, meteorological services, industrial parks or agricultural ministries while models remain heterogeneous; this reduces the analytical usefulness of measuring geopolitical AI influence exclusively through benchmark leadership or chatbot market share.

Data, language and evaluation may become more consequential than model branding

One of the least visible but potentially most consequential elements of the capacity-building programme is China’s commitment to cooperate with developing countries on high-quality datasets and corpora, data infrastructure, testing, evaluation, certification and regulatory practices, because these layers influence which systems can operate effectively in local languages and under local administrative rules. The 2024 plan expressly calls for joint corpus development aimed at protecting linguistic and civilizational diversity, while the 2025 Global AI Governance Action Plan adds common development of high-quality datasets and corpora, joint safety-assessment platforms and broader regulatory cooperation.

For many developing countries, limited local-language data and weak evaluation capacity constitute structural constraints that cannot be solved simply by purchasing additional GPUs, because models trained predominantly on high-resource languages require localized datasets, domain-specific evaluation and administrative integration before they can perform reliably in public-sector applications; a state or company that helps build those corpora and testing systems consequently participates at an earlier and potentially more durable layer of the national AI stack than a provider selling only finished model access.

The public record does not yet provide a consolidated inventory of Chinese-backed corpora, datasets, evaluation laboratories or certification platforms created under these initiatives, making this one of the most important collection gaps because evidence of repeated Chinese participation in national evaluation infrastructure would represent a stronger indicator of ecosystem influence than training attendance alone.

The Chinese model seeks interoperability, but interoperability does not necessarily mean technological neutrality

China repeatedly uses the language of interoperable infrastructure, open-source communities, shared development and sovereign adaptation, and the 2024 capacity-building plan explicitly advocates globally interoperable AI and digital infrastructure while the 2025 governance plan proposes a unified computing-power standards system and international cooperation on intelligent infrastructure.

Interoperability can reduce dependency when technical standards genuinely allow partner governments to move easily among suppliers, but it can also enlarge an ecosystem when a supplier’s technologies become sufficiently widespread that interoperability standards are designed around interfaces, data formats or operational practices familiar to that ecosystem; the strategic question is therefore not whether China uses open technologies, because the official record clearly shows support for open-source approaches, but whether partner-country implementations remain portable across alternative cloud, compute, model and application providers after Chinese-supported deployment.

This distinction should become a major evidentiary test over the next several years, because local deployment of open-weight models on nationally controlled infrastructure would produce a very different sovereignty profile from externally hosted applications whose maintenance, updates, training data or safety evaluation remain dependent on Chinese providers.

The Global South proposition is built around sovereign capacity rather than simple consumption

China’s language is calibrated deliberately around the developmental concern that AI should not reproduce previous technological hierarchies, and UN Resolution 78/311 itself recognizes widening differences in AI readiness while highlighting infrastructure connectivity, skills and human capacity as fundamental problems confronting developing states. China’s subsequent programmes respond directly to those constraints by offering infrastructure cooperation, training, model-development collaboration, data work and governance exchange, which allows Beijing to present its proposition not simply as “buy Chinese technology” but as assistance enabling countries to construct domestic AI capacity.

That proposition can be politically attractive to governments concerned about reliance on a small number of foreign technology companies, particularly when it is combined with local-language development, public-sector applications and infrastructure assistance, but claims of genuine “digital sovereignty” should be tested against actual technical architecture rather than diplomatic language, because sovereignty depends on where models and data are hosted, who can audit or modify systems, whether source code and model weights are accessible, who controls updates and security patches, whether local institutions can operate systems independently, and whether switching providers is technically and financially feasible.

Sovereignty test for Chinese-supported AI projects

Dimension	Evidence of stronger local sovereignty	Evidence of stronger external dependency
Compute	Locally controlled or competitively sourced infrastructure	Exclusive reliance on externally operated compute
Models	Open or transferable models deployable on heterogeneous infrastructure	Proprietary models tied to a single provider
Data	Local custody, clear data-residency rules and local governance	Persistent foreign access or mandatory external hosting
Skills	Local teams able to operate, fine-tune, audit and maintain systems	Long-term dependence on foreign engineers
Standards	Open and internationally interoperable interfaces	Proprietary interfaces producing high switching costs
Evaluation	Domestic or mutually recognized testing capacity	Certification dependent on one external institution
Financing	Transparent financing with competitive procurement	Bundled finance tied contractually to specific suppliers
Maintenance	Transferable service arrangements and documentation	Updates and security support available only from originating vendor

The official Chinese documents support ambitions concerning most of these layers, but they do not provide enough project-level information to classify the overall ecosystem as either sovereignty-enhancing or dependency-producing, meaning that future country studies must examine contracts and technical architecture rather than infer outcomes from political declarations.

The architecture combines state institutions, universities, industry and international organizations

Another distinctive characteristic is the multiplicity of participating Chinese institutions, because international AI cooperation is not being administered exclusively by the Ministry of Foreign Affairs or by a single national technology agency; the evidence reviewed includes the Ministry of Foreign Affairs, National Development and Reform Commission, Beijing municipal technology authorities, universities, research institutes, industrial organizations and the China Meteorological Administration ecosystem, while policy documents explicitly invite governments, industry, academia, think tanks and international organizations to participate.

This produces an architecture resembling a network rather than a single programme, because a foreign government can engage diplomatically through WAICO, develop officials through capacity-building seminars, form university or laboratory relationships, adopt a meteorological or health application, connect enterprises through industrial matchmaking and participate in discussions on testing or technical standards without every component being governed through one contract or ministry.

Such distributed architecture creates resilience because individual relationships can continue even if another strand slows politically, while it complicates measurement because no single budget line or institutional balance sheet captures the full scale of international AI engagement.

The relationship with the United Nations is complementary by design, not formally competitive at this stage

Chinese officials have consistently stated that the United Nations should remain the principal channel for global AI governance, while WAICO is intended to communicate and cooperate with other international organizations rather than replace them, and Chinese representatives have explicitly described WAICO and the UN Global Dialogue on AI Governance as mechanisms that should reinforce each other.

That positioning is strategically useful because it allows China simultaneously to operate inside the universal UN framework, where legitimacy derives from broad state participation, and to develop a China-initiated organization in which implementation priorities can potentially advance more quickly among willing members; whether these mechanisms remain complementary will depend on WAICO’s eventual mandate and whether it begins issuing standards, technical recommendations or policy positions that overlap materially with UN processes.

The distinction should be monitored carefully because the emergence of parallel institutions is not inherently evidence of institutional fragmentation, whereas repeated adoption of common positions or technical practices by WAICO members before negotiations in broader multilateral forums would indicate that the organization is developing coalition-building power beyond technical cooperation.

Emerging strategic mechanism: from access to familiarity, from familiarity to path dependence

The strongest evidence-supported interpretation of China’s institutional design can be represented as a sequential mechanism rather than as a claim of predetermined geopolitical dependency:

Stage	Instrument	Immediate effect	Potential longer-term effect if sustained
Entry	Training, seminars, UN cooperation	Access to officials and technical communities	Professional familiarity and recurring contacts
Demonstration	MAZU and sectoral AI applications	Visible public-service benefits	Demand for adjacent applications and technical support
Localization	Corpora, datasets, education and model cooperation	Adaptation to national languages and institutions	Local developer ecosystems aligned with particular tools
Infrastructure	Compute, data centres, networks and digital infrastructure cooperation	Ability to operate AI systems domestically	Architectural switching costs if technical choices become entrenched
Governance	Testing, certification, regulation and standards dialogue	Compatible administrative practices	Institutional path dependence and easier regional diffusion
Regional scaling	ASEAN, AU, CELAC, Arab League, SCO and BRICS centres	Multi-country replication	Network effects across neighbouring or institutionally linked states
Permanent coordination	WAICO	Continuous intergovernmental mechanism	Potential consolidation of shared practices and negotiating positions

The first three stages are already supported by observable programmes and official implementation evidence, while the later stages remain unevenly documented and should be treated as potential development pathways rather than completed outcomes.

Where the architecture is strongest and where it remains vulnerable

China’s strongest structural advantage is the ability to connect domestic application experience with external development diplomacy, because national policy itself defines AI Plus as spanning science, industry, consumption, public welfare, governance and global cooperation, allowing solutions developed in China’s large internal application market to be presented abroad as transferable tools rather than as isolated frontier research. China’s State Council explicitly includes “AI Plus global cooperation” in the national AI Plus framework and calls for cooperation in compute, data and talent while supporting Global South capacity building and coordination on governance rules and technical standards.

A second advantage is institutional continuity because the architecture now contains both global and regional mechanisms, meaning individual projects can be embedded within longer-running diplomatic relationships, while the development narrative directly addresses widely documented concerns about uneven access to compute, data, expertise and infrastructure.

A third advantage is the potential compatibility between open-source adoption and Chinese hardware or infrastructure ecosystems, because relatively accessible software can reduce the political and financial threshold for initial adoption before governments make larger infrastructure choices, although the official evidence reviewed here does not establish that such a transition occurs systematically and therefore does not support treating it as an accomplished Chinese strategy.

The vulnerabilities are equally substantial because training commitments are inexpensive relative to the cost of large-scale compute infrastructure, many developing states will deliberately diversify vendors rather than accept exclusive technological alignment, open-source software can weaken supplier lock-in, and AI architectures change rapidly enough that today’s technical standards or models may lose relevance before institutional preferences consolidate.

More importantly, the success of the architecture ultimately depends on whether Chinese-supported systems generate measurable economic and administrative value in partner countries, because diplomatic affinity cannot indefinitely compensate for poor model performance, inadequate cybersecurity, unreliable cloud access, high operating costs or weak localization.

Key judgments

China’s international AI policy has progressed from diplomatic principles to a multi-layer institutional system in which infrastructure, applications, human-capital formation, data, language resources, testing, standards and governance are increasingly connected, while the decisive strategic innovation lies in the combination of these elements rather than in any individual programme.

WAICO represents the most important organizational consolidation of that architecture because its officially stated functions extend beyond dialogue into capacity-building, strategy alignment, governance-rule coordination, technical standards and practical cooperation, although its real power cannot be evaluated until budgets, decision procedures, technical bodies and operational programmes become publicly observable.

Training is becoming strategically more sophisticated because the documented 2026 programme concentrates on senior government officials and AI policymakers while combining governance instruction, frontier-technology exposure, industrial visits and international networking, creating a mechanism through which expertise formation can influence later national policy without requiring explicit political alignment.

The regional cooperation-centre initiative has potentially greater structural significance than the headline 5,000-place training commitment because ASEAN, the Arab League, African Union, CELAC, SCO and BRICS offer pre-existing institutional networks through which successful applications, regulatory practices and technical programmes can be replicated across multiple states, although implementation details remain unavailable.

MAZU provides the clearest current example of the transition from capacity-building rhetoric to a deployable public-service technology, with an official April 2026 record documenting more than 200 products across 30 categories, overseas cloud nodes and substantial reported latency improvements, while project-level ownership, cybersecurity, hosting and maintenance arrangements remain insufficiently public to determine whether those deployments enhance or constrain recipient-state technological autonomy.

The most strategically consequential but least documented layer concerns datasets, multilingual corpora, testing, certification and technical standards, because these elements can shape national AI ecosystems for longer periods than individual model generations, making evidence of permanent laboratories, certification platforms, data partnerships and national regulatory adoption particularly important over the next several years.

What would change the assessment

The assessment would strengthen materially if WAICO establishes a funded permanent secretariat, technical committees, regular ministerial or regulator-level mechanisms, jointly financed application programmes, published standards work programmes, or country-level implementation mechanisms connecting training recipients with infrastructure and technology projects, because those developments would show that the organization is becoming an operational ecosystem institution rather than primarily a diplomatic platform.

The assessment would strengthen further if the six regional AI application cooperation centres become operational with identifiable facilities, budgets, participating governments and multi-country programmes, particularly where those programmes combine Chinese-supported compute, models, data infrastructure, training and common technical standards.

Conversely, the assessment would weaken if the five-year training programme remains dominated by isolated seminars without continuing institutional partnerships, if WAICO develops limited permanent capacity, if regional centres remain declaratory, or if recipient countries consistently deploy Chinese-supported applications on heterogeneous infrastructure while retaining easy supplier portability, because those outcomes would indicate cooperation without substantial ecosystem dependence.

Open official record

The complete list of all 29 WAICO founding states is not published in the principal establishment notice reviewed here, although Chinese diplomatic material identifies Kazakhstan, Uzbekistan, Kyrgyzstan, Tajikistan, Belarus, Pakistan, Serbia and Brazil among the members while the formal signing report additionally names Laos, Russia and Indonesia; a complete authoritative treaty-party registry remains necessary before a defensible geographic membership analysis can be produced.

The publicly accessible official material examined does not yet establish WAICO’s assessed contributions, annual budget, secretariat structure, voting procedure, ratification or entry-into-force conditions for each signatory, permanent technical committees, dispute procedures or formal legal relationship with UN institutions.

No official consolidated dataset currently identifies the national allocation of the 5,000 training opportunities, the proportion allocated to policymakers, researchers and engineers, completion statistics, alumni institutional positions, downstream projects or associated expenditure.

No sufficiently detailed official record currently establishes the location, funding, staffing, legal status or implementation schedules of each of the six announced international AI application cooperation centres.

The MAZU record provides concrete product and performance information, but publicly verified project-level documentation remains necessary on data governance, cybersecurity, cloud hosting, source-code access, maintenance costs, training requirements, system ownership and supplier-switching arrangements before conclusions concerning recipient-country technological sovereignty can be made.

Technology stacks, standards and dependency

Principal judgment

The strategically relevant unit of international AI competition is no longer the frontier model considered in isolation, because the economic and political value of a model depends increasingly on the surrounding system through which it is trained, distributed, adapted, connected to proprietary or public data, deployed on compute infrastructure, integrated into applications, secured, evaluated, certified, updated and operated by trained personnel; consequently, durable technological influence is more likely to arise from control or persistent influence across several layers of this stack than from temporary leadership on a model benchmark. China’s policy architecture explicitly connects computing power, data, models, intelligent infrastructure, open-source ecosystems, safety evaluation, standards, applications and skills, while the United States has reached a strikingly similar structural conclusion from the opposite geopolitical direction, with Executive Order 14320 defining an American export package around hardware, models, software, applications and standards rather than around foundation models alone. The European Union, meanwhile, is attempting to reduce structural dependence by integrating compute, data, cloud, AI factories, regulatory standards and a formal sovereignty assessment framework through the Data Union Strategy and the proposed Cloud and AI Development Act.

The resulting competition should therefore not be simplified into OpenAI versus DeepSeek, NVIDIA versus Huawei, or one national model leaderboard against another, because governments and enterprises ultimately consume AI systems rather than models; a model that cannot obtain affordable accelerators, reliable electricity, compatible cloud infrastructure, high-quality data, deployment software, cybersecurity assurance and technically skilled operators has substantially less strategic utility than its benchmark results imply, while an ecosystem whose individual models are occasionally behind the frontier can remain highly competitive when it offers lower deployment costs, accessible infrastructure, interoperable tools, financing, training and long-term technical support.

The central risk for importing countries is consequently not dependence in the binary sense of “Chinese” or “American” technology, but layered dependency, in which switching remains possible at one layer while becoming progressively more costly because applications, data pipelines, security controls, developer skills, procurement procedures and standards have been optimized around adjacent components. This architecture produces path dependence without requiring political exclusivity, and it explains why standards, cloud architecture, developer ecosystems and data governance can become more strategically persistent than the model that initially attracted attention.

The AI stack should be analysed as a system of interlocking dependencies

An AI ecosystem can be divided analytically into at least ten layers, each of which creates a different form of dependence and a different switching problem; the layers are technically separable, but in operational environments they reinforce one another because decisions made at the lower levels constrain the cost and feasibility of alternatives higher in the stack.

Stack layer	Principal assets	Immediate dependency mechanism	Typical switching barrier	Strategic persistence
Semiconductor design	GPU/accelerator architecture, CPUs, networking silicon	Software and model optimisation around specific instruction sets and accelerator capabilities	Re-optimisation, availability, toolchain compatibility	Very high
Compute hardware	Accelerators, servers, networking, storage	Installed capital stock and maintenance ecosystem	Capital replacement and supply availability	Very high
Data-centre infrastructure	Power, cooling, connectivity, facilities	Physical hosting and energy availability	Construction time, grid access and capital expenditure	Very high
Cloud/orchestration	Virtualisation, scheduling, APIs, identity, storage	Applications and workflows built around provider interfaces	Data migration, API rewriting, security revalidation	High
Foundation models	Model weights, inference endpoints, fine-tuning methods	Application behaviour and prompt/tool integration	Revalidation, performance differences and application redesign	Medium
Middleware	Vector databases, inference engines, agent frameworks, MLOps	Workflow integration across applications	Code changes, retraining personnel and operational disruption	High
Data layer	Corpora, enterprise datasets, embeddings, schemas	Model quality depends on data structure and pipelines	Data transformation, governance and legal approvals	Very high
Application layer	Health, government, finance, manufacturing, weather and education systems	Operational processes adapt around deployed systems	Process redesign and institutional retraining	Very high
Standards and assurance	Testing, certification, cybersecurity, labelling and interoperability	Procurement and regulatory compliance favour conforming systems	Recertification and institutional change	Very high
Human capital	Engineers, regulators, administrators and users	Familiarity lowers adoption and maintenance costs	Retraining and organizational learning	Very high

The crucial implication is that model substitutability does not equal ecosystem substitutability: an institution can replace a language model through an API change in weeks or months while still remaining dependent on the same cloud architecture, accelerator fleet, databases, orchestration software, cybersecurity controls and trained workforce for years, meaning that apparent model competition can coexist with deep infrastructure concentration.

Compute is the physical foundation of ecosystem power

The lowest layers of the AI stack have unusually strong persistence because compute infrastructure requires capital investment, electricity, cooling, networking equipment, accelerator procurement and technical expertise whose replacement cycles are measured in years rather than weeks; once a data centre or supercomputing installation is designed around particular hardware and software ecosystems, migration remains technically possible but can require substantial software adaptation and operational revalidation.

China’s Global AI Governance Action Plan addresses this layer explicitly rather than treating AI as a purely software phenomenon, calling for international cooperation on AI and digital infrastructure, improved global distribution of computing resources and work toward a unified computing-power standards system, while also linking infrastructure to datasets, models and applications. The significance of that formulation lies in the attempt to make compute an element of international governance and interoperability, because whoever influences the conventions through which computational resources are measured, interconnected, scheduled and accessed influences the technical environment in which higher-layer AI applications operate.

The United States has reached an equally explicit conclusion about the strategic importance of the complete stack, because Executive Order 14320 directs federal support toward packages containing not merely AI models but also AI-optimized computer hardware, data-centre storage, cloud services, networking, data pipelines, models and cybersecurity measures, thereby formally recognizing that international AI adoption depends on an integrated infrastructure proposition.

The European Union’s June 2026 proposal for a Cloud and AI Development Act similarly identifies cloud and data-centre capacity as prerequisites for wider AI deployment, while proposing measures covering next-generation cloud and AI technologies, accelerated data-centre deployment and an EU-wide framework for assessing cloud and AI sovereignty; the convergence among Chinese, American and European policy documents is analytically important because three major technology blocs are independently treating infrastructure architecture—not the model alone—as the relevant strategic object.

Infrastructure strategies compared

Dimension	China	United States	European Union
Strategic object	Computing power + infrastructure + data + models + applications	Integrated US AI technology stack	Sovereign/resilient cloud + compute + data + AI ecosystem
International orientation	Capacity sharing, infrastructure cooperation and international standards	Export integrated US-origin packages to allies and partners	Primarily European capacity plus international digital partnerships
Compute emphasis	International computing-power cooperation and standardization	US-origin hardware integrated into export packages	AI Factories, Gigafactories, cloud and data-centre expansion
Data-centre policy	Infrastructure cooperation embedded in broader AI governance agenda	Export packages can include data-centre infrastructure	CADA explicitly seeks increased EU data-centre capacity
Standards	Explicit coordination of computing, AI, security and governance standards	Global engagement through consensus standards and exported ecosystem	AI Act harmonised standards + sovereignty framework
Strategic vulnerability addressed	AI divide and access inequality	Adoption of competing foreign stacks	External dependence and insufficient European infrastructure

Sources: China’s Global AI Governance Action Plan, US Executive Order 14320, and European Commission Cloud and AI Development Act proposal.

Cloud dependence can outlast model dependence

Cloud infrastructure represents one of the most important but frequently underestimated dependency layers because enterprise AI workloads are normally embedded inside broader environments involving storage, identity management, databases, cybersecurity, telemetry, container orchestration, networking and data-processing services; changing the underlying model can therefore be substantially easier than changing the cloud platform on which the model’s surrounding application operates.

The European Commission’s 2026 CADA proposal is unusually explicit on this point because it introduces the concept of an EU-wide cloud and AI sovereignty assessment framework, demonstrating that Brussels no longer treats sovereignty primarily as a question of where a company is incorporated but as a multidimensional property of infrastructure and operational control. The proposal identifies capacity, innovation and autonomy as its three principal objectives, while linking public-sector adoption to the sovereignty assessment mechanism.

A rigorous dependency analysis therefore has to distinguish several forms of cloud exposure that are often collapsed into the single term “cloud dependence”.

Cloud dependency dimension	Question determining actual sovereignty	Low-dependency configuration	High-dependency configuration
Physical hosting	Where does the workload execute?	Multiple domestic or interchangeable locations	Single external location/provider
Hardware	Who supplies accelerators and networking?	Heterogeneous supported hardware	Workload optimised for proprietary hardware stack
Orchestration	Can workloads migrate across environments?	Portable containers/open orchestration	Deep use of proprietary managed services
Identity	Who controls authentication and access?	Federated/local identity	Provider-specific IAM deeply embedded
Storage	Can data be exported efficiently?	Standard formats and documented APIs	Proprietary storage services and high egress friction
AI endpoints	Can model provider be changed?	Model abstraction layer	Application directly coupled to proprietary APIs
Security	Can controls transfer with the workload?	Portable security architecture	Provider-specific controls requiring revalidation
Monitoring	Are telemetry and audit formats portable?	Standard logs and open interfaces	Proprietary observability environment
Contractual exit	Are migration and termination rights explicit?	Defined exit plan and portability obligations	Uncertain migration support or substantial exit costs

The strongest technological sovereignty therefore does not necessarily require domestic ownership of every layer, because diversified external suppliers with enforceable portability can provide more practical autonomy than nominally domestic infrastructure whose critical hardware, software or maintenance remains externally controlled; conversely, a domestically located data centre does not by itself establish sovereignty if indispensable software, accelerators, remote administration or updates remain tied to an external supplier.

Hardware–software co-optimization creates a deeper form of lock-in

Advanced AI accelerators are not interchangeable commodities because performance depends strongly on compiler systems, libraries, kernels, networking architectures, memory hierarchies, inference runtimes and developer tools, meaning that accumulated software optimisation around a hardware architecture can become as strategically valuable as the chip itself.

This helps explain why the United States’ international strategy explicitly integrates hardware into a complete technology package instead of treating semiconductor exports as an independent trade category, and the Commerce Department’s FY2027 budget justification refers to revised licensing approaches for certain AI-training chips intended to encourage adoption of the American AI technology stack, demonstrating that semiconductor policy and ecosystem adoption are being treated as connected strategic instruments.

The same dynamic creates an important constraint for China because a domestic accelerator does not achieve full strategic substitutability simply when its peak theoretical performance approaches an imported alternative; it must also support an increasingly mature software environment capable of model training, inference, distributed computing, debugging, monitoring and application integration without imposing unacceptable developer costs.

Why accelerator dependence propagates upward through the stack

Hardware decision	Direct effect	Second-order ecosystem consequence
Accelerator architecture	Determines supported kernels and numerical operations	Influences which models can run efficiently
Interconnect standard	Determines multi-chip scaling	Influences data-centre architecture and training economics
Compiler/toolchain	Determines developer productivity	Shapes skills and software investment
Inference runtime	Determines latency and efficiency	Influences application economics
Driver ecosystem	Determines stability and deployment friction	Creates operational preference for familiar vendors
Monitoring/debugging stack	Determines fault diagnosis	Creates institutional expertise around particular systems
Model optimization libraries	Reduce training/inference cost	Encourages developers to remain inside the ecosystem

The dependency therefore resembles a cumulative investment problem: each additional application, trained engineer and optimized software component increases the value of staying within the established hardware ecosystem, even when technically viable alternatives exist.

Standards transform technical choices into institutional defaults

Standards matter strategically because they convert engineering choices into repeatable procurement, regulatory and interoperability requirements, while their influence persists independently of any individual company; once a testing protocol, security requirement, labelling convention, interface specification or evaluation methodology becomes embedded in public procurement and industry compliance, suppliers adapt their products around it and users accumulate expertise around its implementation.

China formalized an unusually ambitious AI standardization programme in 2024, announcing an objective to formulate more than 50 new national and industry AI standards by 2026 while participating in the development of more than 20 international AI standards, according to government reporting on the guidelines issued jointly by the Ministry of Industry and Information Technology and three other authorities.

This numerical target matters less as a count of documents than as evidence that Beijing regards standardization as industrial infrastructure; national standards can coordinate domestic suppliers and reduce fragmentation inside a very large internal market, while participation in international standards bodies gives Chinese experts opportunities to ensure that global technical specifications accommodate architectures, safety practices and industrial capabilities developed within China.

China’s documented standardization trajectory

Indicator	Value/status	Reference point	Significance
New national and industry AI standards targeted	More than 50	By 2026	Broad domestic standard-system construction.
International AI standards participation targeted	More than 20	By 2026	Explicit internationalization objective.
Generative-AI service security standard	GB/T 45654-2025	Published 25 Apr 2025; effective 1 Nov 2025	Establishes basic security requirements for generative AI services.
AI-generated content labelling standard	GB 45438-2025	Published 28 Feb 2025; effective 1 Sep 2025	Standardizes technical methods for synthetic-content labelling.
Large-model testing/evaluation standard	GB/T 45288.2-2025	Published 3 Nov 2025	Formalizes metrics and methods for large-model evaluation.
AI-agent interconnection work	National technical-guidance project registered	Current standardization pipeline	Indicates movement toward interoperability of autonomous-agent systems.

These standards illustrate movement beyond generic AI principles toward operational specifications covering model security, synthetic-content identification, evaluation and interoperability, precisely the areas in which compliance requirements can begin affecting software design and procurement.

China’s standardization objective is international compatibility as well as domestic control

China’s Standardization Law explicitly instructs the state to participate actively in international standardization, conduct international cooperation, participate in international standards development and promote conversion and application between Chinese and foreign standards, meaning that international standardization is embedded in the country’s general legal framework rather than appearing only in recent AI policy.

The 2025 Global AI Governance Action Plan extends this orientation specifically into AI by calling for coordinated international work on safety, data protection, intelligent infrastructure and computing resources, and by advocating compatible governance rules and technical standards; the strategic implication is not that Chinese domestic standards will automatically become global standards, because international standards bodies operate through multilateral technical processes, but that China is seeking to ensure that its domestic industrial scale is accompanied by sustained participation in the venues where global compatibility requirements are negotiated.

That distinction is fundamental because standard-setting influence is not equivalent to unilateral rule-setting: a Chinese proposal accepted through an ISO, IEC, ITU or other consensus process becomes part of a multilateral technical architecture rather than simply an exported Chinese regulation, while the strategic benefit can nevertheless be substantial if domestic Chinese firms already operate efficiently under compatible specifications.

The United States is also treating standards as geopolitical infrastructure

Washington’s response demonstrates that standards competition is recognized explicitly on the American side, because NIST’s Plan for Global Engagement on AI Standards, developed with the Department of Commerce, Department of State and other federal agencies, sets out a government strategy for international cooperation, standards-development activities and information sharing, while the wider 2025 United States Standards Strategy was updated against what its official presentation describes as intensified geopolitical competition in which standards shape emerging industries.

The United States therefore possesses a structurally different but equally strategic standardization model, relying substantially on private-sector-led and consensus-based standards institutions supported by federal technical expertise, while China operates through a more state-coordinated domestic framework combined with participation in international standards organizations.

Standards competition is not a simple China-versus-US binary

Variable	China	United States	European Union
Domestic governance model	Strong state coordination	Market/industry-led standards with federal participation	Regulatory requirements combined with European standards organizations
Strategic objective	Industrial coordination + international participation	International consensus standards compatible with US innovation ecosystem	Standards supporting regulatory compliance and single-market integration
AI safety	National security and service requirements	NIST risk-management and measurement frameworks	AI Act obligations and harmonised standards
International strategy	Expand participation in international standardization	Global engagement plan led through NIST/Commerce/State	External regulatory influence plus international standards participation
Market mechanism	Scale of domestic industrial deployment	Global position of US technology firms	Access to EU Single Market
Principal leverage	Manufacturing/application scale	Software, cloud, semiconductors and developer ecosystems	Regulatory market power and conformity requirements

Sources: Chinese national standardization records, NIST AI standards programme and the EU AI regulatory framework.

Europe’s regulatory standards can become part of the global stack even without a dominant European foundation model

The European Union illustrates why model leadership and ecosystem influence should not be conflated, because the AI Act creates a legal framework whose compliance requirements can affect systems developed outside Europe whenever those systems fall within its territorial scope, while standards supporting compliance can consequently influence product design far beyond European model developers themselves. The European Commission identifies Regulation (EU) 2024/1689 as the Union’s harmonised AI regulatory framework, creating obligations differentiated according to risk and system type.

This provides Europe with a form of ecosystem leverage that operates through market access rather than ownership of the dominant model, because technology providers serving European customers have incentives to engineer compliance into documentation, risk management, data governance and assurance processes rather than maintain entirely separate architectures for each jurisdiction.

The effect should nevertheless not be overstated into an automatic “Brussels effect,” because firms can segment products geographically, regulatory divergence can raise European deployment costs, and international competitors can establish alternative technical standards where the EU market is less economically decisive; Europe’s standards influence therefore depends ultimately on the combination of market size, implementation consistency and the economic attractiveness of compliance.

Data architecture is the least visible layer and potentially the hardest to replace

Models require data during pretraining, fine-tuning, retrieval, evaluation and operational monitoring, while most enterprise and government AI systems become useful only when connected to institution-specific datasets; consequently, the organization controlling the model does not necessarily control the most strategically valuable asset, because a ministry, hospital, bank or manufacturer can possess data whose structure determines the cost of changing systems.

China’s global AI policy specifically identifies high-quality datasets, corpora, data infrastructure and data-security standards as areas for international cooperation, while its 2025 Action Plan calls for improved security throughout training-data collection and model-generation processes.

The European Union has made the same connection between data availability and AI competitiveness through its Data Union Strategy, whose stated objective is to increase access to high-quality curated data for AI development while simplifying Europe’s data environment.

Data dependency develops through multiple layers

Data layer	Dependency mechanism	Why migration becomes difficult
Raw data	Physical or contractual custody	Volume, privacy restrictions and transfer cost
Data schemas	Application-specific structure	New platform requires transformation and validation
Metadata	Classification and governance system	Institutional processes depend on existing taxonomy
Embeddings	Model-specific or architecture-dependent vector representation	Changing embedding model can require re-indexing entire corpora
Vector databases	Retrieval infrastructure	Index structure and query semantics can differ
Fine-tuning datasets	Model-adaptation assets	Formats and training pipelines may be provider-specific
Evaluation datasets	Quality-assurance baseline	Switching models requires repeat testing
Safety datasets	Red-team and compliance corpus	Regulatory evidence must often be regenerated
Logs and telemetry	Operational history	Migration can disrupt monitoring continuity

This is one reason why the strategic value of training foreign engineers extends beyond goodwill: engineers who design national data pipelines, evaluation datasets and information architectures make design decisions whose persistence can exceed the life of the model they originally deployed.

Multilingual corpora create a particularly powerful form of ecosystem influence

For countries whose languages are underrepresented in global training datasets, the decisive AI bottleneck can be neither compute nor access to a frontier model but the lack of high-quality, legally usable and culturally representative corpora, making language-data partnerships potentially more consequential than providing temporary model access.

China’s capacity-building policy explicitly identifies joint construction of AI corpora and the reduction of linguistic and cultural discrimination as international cooperation objectives, while the wider governance plan includes high-quality datasets and corpora among the elements requiring international collaboration.

A local-language corpus can become a strategic infrastructure asset because subsequent models from different suppliers can be evaluated or fine-tuned against it, but control over annotation standards, licences, hosting and access rights determines whether the asset strengthens national sovereignty or instead becomes embedded in an external technology ecosystem.

Sovereignty conditions for language-data partnerships

Question	Sovereignty-preserving outcome	Dependency-enhancing outcome
Who owns corpus rights?	Local institution or shared rights	Exclusive external ownership
Where is corpus hosted?	Local or portable infrastructure	Single foreign platform
Are annotations documented?	Open/reproducible specification	Proprietary annotation method
Can other models use the data?	Vendor-neutral licence	Restricted to designated ecosystem
Can corpus be exported?	Standard formats	Technically or contractually restricted
Who updates the corpus?	Local trained teams	Continuing external dependence
Who defines quality metrics?	Joint/local evaluation authority	Supplier-controlled criteria

The important question for future Global South deployments is therefore not simply whether China supplies localized AI, but whether localization produces transferable national assets or supplier-specific assets.

Middleware can become more durable than the frontier model

Between models and applications lies a rapidly expanding software layer containing inference servers, model gateways, agent frameworks, retrieval systems, orchestration tools, databases, security controls, observability platforms and MLOps systems; these components are often ignored in geopolitical analysis despite their importance because they determine how easily organizations can change underlying models.

A well-designed middleware layer can reduce strategic dependence by allowing applications to route requests among multiple models, while tightly coupled proprietary middleware can do the opposite by making model switching technically expensive despite nominal API availability.

Model portability is determined by architecture, not simply licensing

Architecture	Model switching difficulty	Dependency profile
Application directly calls one proprietary model API	High	Provider-specific
Application uses a generic model gateway	Moderate/low	Multi-provider
Open model hosted on proprietary cloud service	Moderate	Model portable, infrastructure less portable
Proprietary model running in local controlled environment	Moderate	Infrastructure sovereign, model supplier-dependent
Open model on open orchestration and heterogeneous hardware	Lower	Strong portability if skills are available
Fine-tuned proprietary model deeply integrated into internal workflows	Very high	Behavioural and operational dependence
Retrieval-heavy architecture with portable data layer	Lower	Model can be replaced more easily

This produces a strategic paradox in which open models do not automatically eliminate dependency, because an open-weight model deployed through proprietary hardware, proprietary inference software and proprietary cloud orchestration can remain embedded in a concentrated stack, while a proprietary model accessed through a carefully abstracted architecture may create comparatively limited switching costs.

Open source changes the geometry of competition but does not abolish ecosystem power

China has repeatedly emphasized open-source AI cooperation, and this approach can be particularly effective in developing markets because it reduces entry costs, enables local adaptation and can make AI deployment feasible where permanent access to premium proprietary APIs would be economically difficult.

Yet open weights are only one component of sovereignty because local institutions still require computing infrastructure, inference software, engineers, data pipelines, security tooling and maintenance, while performance optimization can tie supposedly open systems to particular accelerator environments.

The appropriate analytical distinction is therefore between model openness and stack openness.

Layer	What genuine openness requires
Model	Weights available under usable licence
Runtime	Multiple inference engines supported
Hardware	Deployment across multiple accelerator architectures
Data	Portable standard formats
APIs	Open interfaces rather than proprietary-only integrations
Evaluation	Reproducible metrics and datasets
Fine-tuning	Documented methods and portable checkpoints
Security	Auditable controls and transferable logs
Orchestration	Ability to move workloads across clouds or local systems
Skills	Local teams capable of operating without permanent external support

A model can therefore be technically open while its effective deployment environment remains dependent, which makes claims of “digital sovereignty” impossible to evaluate from model licensing alone.

Certification can turn technical standards into market structure

Evaluation and certification constitute another layer capable of reinforcing ecosystems because governments frequently require evidence that systems satisfy safety, cybersecurity, reliability or sector-specific obligations before deployment, particularly in healthcare, finance, critical infrastructure and public administration.

China’s GB/T 45654-2025 establishes security requirements for generative-AI services, while GB 45438-2025 establishes technical labelling methods for AI-generated and synthetic content, and GB/T 45288.2-2025 addresses testing and evaluation metrics and methods for large-scale models. These standards show that China’s domestic framework is moving from high-level governance concepts toward reproducible engineering requirements capable of influencing testing laboratories, product-development processes and regulatory compliance.

Once certification ecosystems develop around such requirements, dependency can shift away from the original technology supplier toward laboratories, compliance tools, auditing practices and trained assessors; this is strategically significant because a country adopting a foreign application can theoretically replace it, whereas changing an entire national testing and assurance regime involves regulators, accreditation bodies, procurement rules and professional training.

Skills represent the slowest-moving dependency layer

Human capital is one of the most durable elements of the stack because organizations naturally prefer tools understood by their existing engineers and administrators, while replacing an established ecosystem requires not simply buying new software but retraining staff, rewriting operating procedures, recreating documentation and accepting temporary productivity loss.

This makes developer familiarity a form of economic infrastructure: when universities teach particular programming environments, engineers receive vendor-specific certification, regulators learn specific evaluation methodologies and government IT departments gain experience with certain architectures, subsequent procurement is affected by the cost of departing from accumulated expertise even without formal exclusivity.

The mechanism is measurable conceptually as a switching-cost function:

Total switching burden = infrastructure replacement + software conversion + data migration + security revalidation + regulatory recertification + personnel retraining + temporary productivity loss + operational risk.

No defensible public dataset currently allows these elements to be assigned universal numerical weights, so any attempt to generate a single “dependency score” would create false precision; the appropriate approach is instead to assess each layer separately for portability, substitutability, concentration and time-to-replace.

Dependency should therefore be measured as a matrix, not a binary condition

A government can simultaneously possess high sovereignty at the data level, moderate dependence at the model level and severe dependence at the semiconductor level, meaning that statements such as “Country X depends on Chinese AI” or “Country Y uses American AI” are analytically inadequate unless the relevant layer is specified.

Decision-grade dependency assessment framework

Layer	Core indicator	Low exposure	Medium exposure	High exposure
Compute	Alternative accelerator availability	Multiple viable architectures	Alternatives exist but migration costly	Single viable supplier/ecosystem
Cloud	Workload portability	Multi-cloud/local portability	Partial portability	Deep provider lock-in
Models	Substitutability	Multiple validated models	Alternatives require significant adaptation	Mission process dependent on one model
Middleware	Open interfaces	Open standards/gateways	Mixed proprietary/open	Proprietary orchestration throughout
Data	Exportability and ownership	Local custody + standard formats	Shared control	External control or restricted portability
Applications	Process integration	Modular	Important but replaceable	Core administrative processes embedded
Standards	Supplier neutrality	International/vendor-neutral	Mixed	Procurement designed around one ecosystem
Certification	Alternative accredited pathways	Multiple	Limited	Single ecosystem controls compliance path
Skills	Workforce diversity	Multi-stack expertise	Strong preference for one stack	Critical expertise concentrated in one ecosystem
Maintenance	Local operational capability	Independently operated	Mixed support	External supplier indispensable

This framework produces a more defensible understanding of strategic dependence than market-share figures alone because market share measures current usage, whereas dependency concerns the cost, time and operational risk of exit.

The architecture creates different time horizons of dependency

Not all stack layers persist equally long, and policy should therefore distinguish between rapidly replaceable components and structural infrastructure.

Layer	Indicative strategic replacement horizon	Principal reason
Frontier model API	Months	Competitive model market and API abstraction
Application model	Months to several years	Revalidation and workflow redesign
Middleware	1–3+ years	Codebase and operational integration
Data architecture	Several years	Transformation, governance and institutional history
Cloud architecture	Several years	Large migration and security burden
Accelerator fleet	3–7+ years	Capital lifecycle and software optimization
Data-centre infrastructure	10+ years	Physical construction, power and networking
Standards framework	Often many years	Regulatory and procurement institutionalization
Human-capital ecosystem	Potentially decades	Education, professional networks and accumulated expertise

These are analytical ranges rather than official forecast values, and individual deployments can differ substantially; their purpose is to show why a benchmark advantage lasting six months can be strategically less important than a data-centre, standards or professional-training relationship that persists across several model generations.

The United States has explicitly adopted ecosystem competition

The strongest evidence against the claim that the Western response remains concentrated exclusively on model performance comes from the United States itself, because Executive Order 14320, Promoting the Export of the American AI Technology Stack, directs the Commerce Department to establish a programme supporting full-stack AI export packages and identifies components ranging from AI-optimized hardware and data-centre infrastructure to cloud services, networking, data pipelines, models, applications and cybersecurity.

The order’s strategic logic is straightforward: international adoption of a complete American stack creates commercial opportunities across multiple technology layers while reducing the incentive for partner states to construct systems around competing architectures; the US-China competition is therefore increasingly symmetrical at the level of ecosystem objectives even when policy instruments and institutional narratives remain substantially different.

Washington’s February 2026 international messaging further connected AI exports with adoption and technological sovereignty during the India AI Impact Summit, reinforcing the conclusion that foreign infrastructure adoption has become an explicit element of US technology diplomacy rather than an incidental consequence of private corporate expansion.

The actual strategic contest

Simplified interpretation	More accurate strategic unit
OpenAI vs DeepSeek	Competing model + infrastructure + developer ecosystems
NVIDIA vs Chinese accelerators	Hardware + compiler + runtime + networking + software environment
Proprietary vs open source	Degree of portability across the complete stack
US vs Chinese cloud	Compute + data + orchestration + cybersecurity + contractual exit
Western vs Chinese standards	Competing participation inside overlapping international standardization systems
AI training programmes	Human-capital formation + subsequent institutional and technical choices
Best model benchmark	Total deployment capability, cost, reliability and ecosystem support

Europe’s response increasingly centres on technological sovereignty across the stack

Europe’s 2025–2026 policy evolution is particularly revealing because the Union is shifting from a predominantly regulatory discussion toward a combined infrastructure, data and sovereignty strategy; the Data Union Strategy connects AI competitiveness with access to high-quality data, while the Cloud and AI Development Act proposal of June 2026 seeks expanded data-centre capacity, next-generation cloud technology and a common sovereignty-assessment framework.

EuroHPC’s AI Factory architecture adds the compute layer by offering European companies and researchers tailored access to advanced supercomputing capacity and specialist support, which addresses one of the principal barriers preventing smaller firms from operating advanced models independently of hyperscale commercial platforms.

The distinctive European challenge is therefore no longer the absence of policy instruments but their integration: compute can exist without sufficient cloud capacity, sovereign cloud can exist without competitive foundation models, regulation can exist without domestic suppliers, and datasets can exist without practical interfaces connecting them to industrial AI; Europe’s strategic position will consequently depend on whether these components become a coherent stack rather than parallel initiatives.

Three different ecosystem strategies are emerging

The verified official record supports a structural comparison in which China, the United States and European Union are converging on the importance of the complete AI stack while emphasizing different mechanisms.

Strategic variable	China	United States	European Union
Core international proposition	Capacity, applications, infrastructure and shared development	Export integrated American technology packages	Sovereign/resilient ecosystem plus international partnerships
Compute	Domestic alternatives + international cooperation	Leading US hardware and data-centre ecosystem	EuroHPC, AI Factories and Gigafactory policy
Cloud	Domestic cloud ecosystem + overseas cooperation	Globally dominant commercial cloud ecosystem	Expansion + sovereignty criteria under CADA
Models	Strong domestic/open-model ecosystem	Frontier proprietary and open commercial ecosystem	Mix of European models and external providers
Data policy	Data infrastructure + corpus cooperation + security	Private/public data ecosystem with sectoral governance	Data Union Strategy and European data spaces
Standards	State-coordinated domestic system + international participation	Consensus standards + NIST technical leadership	Harmonised regulation-supported standards
International deployment	State diplomacy + companies + institutions	Commercial firms supported by federal export strategy	Commission, member states and Global Gateway instruments
Human capital	Capacity-building and training diplomacy	Universities, firms, international partnerships	Research programmes, skills policy and partnerships
Principal systemic advantage	Integration of state policy, industrial scale and development diplomacy	Frontier technology firms and global cloud/software reach	Large regulated market and institutional rule-setting capacity
Principal structural challenge	External semiconductor constraints and trust concerns	Cost/accessibility and partner concerns over dependence	Fragmented commercial scale and external infrastructure dependence

The evidence therefore does not support a narrative in which one actor has discovered ecosystems while the others remain confined to model competition; instead, all three are now moving toward ecosystem strategy, but from different positions of strength.

The strategic prize is not monopoly but default status

The most consequential form of AI power is unlikely to resemble absolute monopoly because partner governments have strong incentives to diversify suppliers, open models can circulate internationally, commercial competition will remain intense, and technological generations turn over rapidly; a more plausible objective for competing ecosystems is therefore to become the default environment in which new applications are developed.

Default status matters because procurement officials tend to select systems compatible with existing infrastructure, developers prefer familiar tools, universities teach technologies demanded by employers, regulators develop experience around known assurance frameworks, and companies build complementary services where customer demand already exists.

The result is a self-reinforcing cycle:

installed infrastructure → trained workforce → more applications → more data and operational experience → stronger complementary ecosystem → lower incremental deployment cost → further adoption.

This cycle requires no coercion and no exclusive contracts, which is precisely why dependency can be difficult to observe before it becomes economically significant.

The most dangerous analytical error is to confuse openness with independence

An ecosystem can market itself as open, interoperable or sovereignty-enhancing while still producing material dependencies elsewhere in the stack, and the reverse is also true: proprietary components can coexist with strong sovereignty when users possess meaningful exit rights, multi-vendor architecture, local data control and trained domestic personnel.

Accordingly, every claim concerning “digital sovereignty” should be tested against at least seven operational questions: whether workloads can migrate, whether data can be exported in standard formats, whether alternative hardware is supported, whether the system can operate without remote vendor intervention, whether security controls remain functional after migration, whether local engineers possess maintenance capability, and whether contractual or regulatory arrangements permit realistic exit.

Sovereignty stress test

Test	Question	Evidence required
Exit test	Can the user leave the provider?	Migration clauses, export tools, replacement architecture
Continuity test	Can service continue during supplier disruption?	Local redundancy and alternative suppliers
Data test	Who controls the operational and training data?	Contracts, architecture and residency rules
Model test	Can another model be substituted?	Abstraction architecture and validation evidence
Hardware test	Can workloads run elsewhere?	Supported accelerators and benchmark validation
Skills test	Can local personnel maintain the system independently?	Staffing and training documentation
Standards test	Are interfaces controlled by open/consensus specifications?	Technical standards
Security test	Can security auditing continue independently?	Logs, audit access and certification structure
Update test	Who controls critical software updates?	Maintenance and source-access arrangements
Financial test	What is the complete exit cost?	Migration, retraining and replacement estimates

Until this information is known, assertions that any Chinese, American or European AI programme either creates dependency or guarantees sovereignty should remain provisional.

Standards can determine markets without determining winners

The strongest long-term scenario is therefore not one in which a single national model becomes globally dominant, but one in which technical standards, infrastructure interfaces, testing methodologies and professional practices narrow the range of economically sensible choices available to users.

China’s target of more than 50 national and industry AI standards and participation in more than 20 international standards by 2026 should be understood in this context, alongside concrete standards already covering generative-AI security, synthetic-content labelling and large-model evaluation.

The United States is simultaneously using NIST and the international standards system to encourage globally accepted consensus frameworks, while the European Union combines international standardization with legally enforceable market-access requirements under the AI Act.

The likely result is therefore neither complete technological bifurcation nor seamless global convergence, but selective interoperability: common standards will survive where global commerce benefits from compatibility, while security requirements, data governance, cloud sovereignty, content controls and politically sensitive applications are more likely to fragment across jurisdictions.

Dependency can accumulate without exclusive adoption

A developing economy does not have to choose a completely Chinese or completely American AI architecture, because it might simultaneously use American accelerators, Chinese open models, domestic data centres, European cybersecurity standards and locally developed applications; such hybrid systems are likely to be common precisely because governments seek diversification.

Hybridization, however, does not eliminate strategic dependence because individual bottlenecks can dominate the system: if an application can use five models but only one accelerator architecture is economically viable, hardware remains the critical dependency; if compute is diversified but all government data are structured around a proprietary cloud environment, the data/cloud layer becomes decisive instead.

Dependency is governed by the bottleneck layer

Scenario	Apparently diversified components	Actual strategic bottleneck
Multiple models, one accelerator ecosystem	Models	Compute
Multiple clouds, proprietary application middleware	Infrastructure	Software
Open models, foreign-hosted sensitive data	Models	Data/cloud
Domestic cloud, imported accelerators	Hosting	Semiconductor supply
Local models, foreign evaluation regime	Models	Certification/standards
Diverse technology, workforce trained on one ecosystem	Hardware/software	Skills
Open software, proprietary high-value datasets	Software	Data

The central policy objective for countries seeking autonomy should therefore be bottleneck diversification rather than symbolic localization.

Key quantitative evidence

Indicator	Verified value	Date/status	Strategic meaning
China planned new national/industry AI standards	>50	Target by 2026	Scale of domestic standard-system build-out.
Chinese participation target in international AI standards	>20	Target by 2026	Internationalization of technical influence.
Chinese generated-content labelling standard	GB 45438-2025	Effective 1 Sep 2025	Converts synthetic-content governance into technical specification.
Chinese generative-AI security standard	GB/T 45654-2025	Effective 1 Nov 2025	Formal security baseline for generative services.
Large-model evaluation standard	GB/T 45288.2-2025	Published Nov 2025	Formal testing and evaluation framework.
US export architecture	Full-stack packages	EO 14320, Jul 2025	Hardware through applications treated as single strategic export proposition.
EU CADA policy dimensions	3 principal objectives	Proposal Jun 2026	Innovation, infrastructure capacity and autonomy.
EU cloud sovereignty mechanism	Single EU-wide assessment framework proposed	Jun 2026	Treats cloud dependence as strategic-policy variable.
NIST global AI standards strategy	Federal cross-agency plan	Published 2025	Confirms standards as international US policy instrument.

Key judgments

The evidence indicates that the AI model itself is becoming one of the more replaceable components of the strategic technology system, while compute infrastructure, cloud architecture, data organization, standards and human capital are more durable sources of dependence because their replacement requires physical investment, institutional restructuring or accumulated learning rather than a simple software substitution.

China’s strategy is particularly important because it connects domestic standardization with international capacity-building, computing-power cooperation, datasets, corpora, applications and technical standards, creating the possibility that Chinese participation can extend across several layers of partner-country AI ecosystems even when those countries never adopt a single dominant Chinese foundation model.

The United States has nevertheless recognized essentially the same strategic geometry, as demonstrated by the explicit federal policy of exporting complete American AI technology stacks rather than individual models, making the central US-China competition increasingly a contest over which ecosystem becomes easiest, cheapest and institutionally safest for third countries to adopt.

Europe possesses a different source of leverage because its AI Act, Data Union Strategy, EuroHPC infrastructure and proposed Cloud and AI Development Act allow it to influence standards, data governance, compute availability and sovereignty requirements even without controlling the globally dominant foundation-model ecosystem.

The most important future indicator will therefore not be which country produces the highest-scoring model at a particular moment, but whether governments and enterprises can realistically change suppliers across compute, cloud, models, middleware, data and assurance regimes without prohibitive cost or operational disruption, because that exit capability provides the most rigorous operational definition of technological sovereignty.

What would change the assessment

The assessment that ecosystem architecture is becoming more important than isolated model superiority would weaken if cross-provider model, cloud, compute and data portability improves sufficiently that switching costs fall materially across the complete stack, particularly if mature open standards allow enterprises to change accelerator, cloud and model providers without extensive software adaptation or recertification.

It would strengthen substantially if national governments begin incorporating supplier-origin, infrastructure sovereignty, cloud jurisdiction, AI-stack portability or domestic-control requirements systematically into public procurement, because such measures would demonstrate that governments themselves increasingly view the stack rather than the model as the relevant strategic object.

The assessment of Chinese standards influence would strengthen if Chinese-developed specifications achieve significant adoption through international standards organizations or become procurement defaults across multiple partner countries, while it would weaken if domestic Chinese standards remain largely confined to China and partner states continue using alternative international frameworks.

The assessment of American ecosystem power would strengthen if Executive Order 14320 produces identifiable international full-stack deployments combining US hardware, cloud, models, cybersecurity and applications, while it would weaken if partner countries increasingly disaggregate those packages and select components from multiple national ecosystems.

The assessment of European strategic autonomy would strengthen if CADA results in materially increased European cloud and data-centre capacity, widespread sovereign procurement criteria, competitive European AI services and meaningful workload portability; it would weaken if European AI infrastructure expansion occurs while critical compute, cloud or model layers remain concentrated outside Europe.

Open official record

No authoritative international dataset currently measures AI dependency across the complete technology stack, and available market-share data cannot substitute for such a measure because usage concentration does not directly reveal the cost, time or operational risk of changing suppliers.

A complete public inventory comparing Chinese, US and European participation in ISO, IEC and ITU AI standardization projects by proposal origin, technical leadership, adoption status and national implementation would materially improve the assessment, but the official sources reviewed here do not provide a consolidated comparable dataset of that form.

The publicly accessible record does not yet establish the extent to which China has achieved the 2024 target of more than 50 new domestic AI standards and participation in more than 20 international standards by the end of 2026; individual standards and active projects are verifiable, but a final audited government tally covering the complete target period has not been identified.

Project-level evidence is also required to determine whether Chinese Global South deployments are becoming vertically integrated stacks or remain heterogeneous combinations of Chinese, American, European and domestic technologies, because the strategic consequences differ fundamentally between those configurations.

Finally, the economic magnitude of switching costs remains insufficiently documented across all three major ecosystems, particularly for government and critical-infrastructure deployments, making detailed country-level architecture, contract, data-portability, certification and workforce analysis necessary before dependency can be quantified rather than described structurally.

Western response and European exposure

Principal judgment

The Western response to China’s widening AI diplomacy is no longer defined primarily by frontier-model competition, semiconductor export controls or regulatory rule-making, because the United States, European Union and major European states are now constructing their own combinations of compute capacity, data-centre infrastructure, investment incentives, sovereign-cloud frameworks, public procurement, standards, industrial partnerships and international technology diplomacy; however, these responses remain structurally asymmetric, since the United States possesses the deepest commercial technology stack and has begun explicitly mobilising it abroad, Europe possesses substantial regulatory, scientific and public-finance instruments but remains fragmented across EU and national layers, the United Kingdom is building a particularly investment-oriented hybrid architecture around foreign frontier firms and domestic sovereign capabilities, France is attempting to convert abundant low-carbon electricity into a continental compute advantage, Germany is concentrating on data centres, industrial AI and technological sovereignty, while Italy combines a comparatively smaller domestic compute base with a potentially important international-development position through its G7-backed AI Hub for Sustainable Development and the Bologna supercomputing ecosystem.

The most important difference from the Chinese approach lies therefore not in whether Western governments understand that AI competition extends beyond the model layer, because the documentary record demonstrates that they do, but in institutional execution: Beijing can connect diplomacy, development policy, state-supported infrastructure and standards through a highly coordinated political architecture, whereas the Western response is distributed among private hyperscalers, semiconductor firms, national governments, development institutions, the European Commission, EuroHPC, research institutes and separate regulatory authorities whose incentives and investment cycles do not always coincide. This fragmentation creates greater supplier diversity and potentially greater technological openness, but it also increases the difficulty of presenting developing economies with a coherent package equivalent to a single integrated offer of compute, finance, skills, applications and long-term technical cooperation.

Washington has formally moved from technology leadership to ecosystem export

The clearest change in American strategy is that Washington no longer assumes that technological superiority automatically translates into international adoption, because the July 2025 America’s AI Action Plan and the subsequent executive order on exporting the American AI technology stack explicitly define foreign adoption as an objective of national policy. The executive order encompasses AI-optimised computing hardware, data-centre infrastructure, cloud services, networking, data pipelines, foundation models, applications and cybersecurity, establishing a government mechanism intended to support complete export packages rather than isolated software or semiconductor sales.

This represents an important doctrinal shift because the United States historically obtained much of its technology influence through commercially successful firms rather than through integrated state-designed technology packages, whereas the new architecture recognises that governments choosing national AI infrastructure increasingly require combinations of financing, compute, energy, cloud capacity, models, security and skills that cannot always be assembled efficiently through uncoordinated commercial purchases. The strategic objective is therefore increasingly to ensure that countries entering the AI era construct their systems around American-origin technologies before competing stacks acquire durable installed bases.

The United States retains a structural advantage that no European country currently possesses at equivalent scale: leading firms operate across almost every strategic AI layer, from accelerator design and cloud computing to frontier models, enterprise software, cybersecurity and developer platforms. That concentration permits Washington to build international AI diplomacy on top of commercial ecosystems that already possess global customer bases, although it simultaneously creates a policy vulnerability because foreign governments can interpret dependence on a small number of US firms as a sovereignty risk, particularly when data jurisdiction, export controls or extraterritorial regulatory authority become politically salient.

The emerging American full-stack proposition

Strategic layer	US capability relevant to international deployment	Government policy direction	Principal exposure
Accelerators	Global leadership in advanced AI accelerator design	Include hardware within export packages	Export-control policy can conflict with market-expansion objectives
Cloud	Globally scaled hyperscale infrastructure	Include cloud and data-centre services in stack exports	Partner concerns over jurisdiction and concentration
Models	Strong frontier-model ecosystem	Promote American model adoption internationally	High inference cost and concern over proprietary dependence
Cybersecurity	Mature commercial and federal ecosystem	Integrate security into export architecture	Certification fragmentation across jurisdictions
Applications	Deep enterprise-software market	Encourage downstream adoption	Localisation may require substantial country-specific work
Standards	Strong NIST and industry participation	Shape international standards through technical engagement	Consensus processes limit unilateral control
Finance	Large private capital markets	Public policy increasingly seeks to mobilise investment abroad	No single development-finance mechanism equivalent to domestic capital depth
International deployment	Government increasingly supporting technology-stack exports	Full-stack strategy formalised in 2025	Coordination across agencies and companies remains complex

The strategic tension is particularly important because semiconductor policy can simultaneously strengthen and weaken American ecosystem influence: export restrictions can slow the diffusion of high-end competitors’ capabilities, but restrictions that make advanced US hardware difficult to obtain can also encourage foreign states to accelerate substitution toward non-American architectures, meaning Washington must balance technology-denial objectives against the commercial advantages generated by widespread adoption of US platforms.

Europe’s response has shifted from “AI regulation” toward an industrial and geopolitical architecture

The European Union’s position is materially stronger than the widespread characterization of Europe as primarily a regulator suggests, because the AI Continent Action Plan now combines compute infrastructure, data access, talent, adoption and regulation within a single industrial-policy framework, while the International Digital Strategy adds an external dimension built around partnerships, secure infrastructure, digital public infrastructure, cybersecurity and an integrated EU Tech Business Offer for partner countries. The Commission explicitly describes that offer as modular and intended to combine public and private investment with European technologies and capacity-building, placing Europe conceptually much closer to an ecosystem-export model than it was only several years earlier.

Europe’s principal strategic problem is consequently not absence of policy ambition but conversion capacity: the EU must transform public funding, regulatory leverage, national industrial assets and multiple technology programmes into packages that can be deployed rapidly enough to compete with American commercial scale and Chinese institutional coordination.

The infrastructure expansion is nevertheless substantial. As of April 2026, the Commission reported 19 AI Factories deployed across Europe and 13 AI Factory antennas, while the broader AI Continent framework assigns approximately €10 billion to AI Factory infrastructure between 2021 and 2027 and seeks to mobilise €20 billion for up to five AI Gigafactories capable of substantially larger-scale frontier-model training.

European AI infrastructure architecture

Instrument	Verified scale/status	Function	Strategic significance
AI Factories	19 deployed by Apr 2026	HPC access, model training, fine-tuning and support	Creates shared European compute capacity outside commercial hyperscalers.
AI Factory antennas	13	Regional access to AI Factory resources	Extends infrastructure geographically beyond supercomputer host sites.
AI Factory budget	Approximately €10bn, 2021–2027	Compute infrastructure and related ecosystem	Publicly anchors European AI compute capacity.
AI Gigafactory investment facility	€20bn intended mobilisation	Very large training and development facilities	Attempts to close Europe’s frontier-compute gap.
Planned Gigafactories	Up to 5	Training and development of highly complex models	Creates a European frontier-compute layer if delivered at planned scale.
International Digital Strategy	Adopted Jun 2025	External digital partnerships and governance	Adds geopolitical deployment to internal industrial policy.
EU Tech Business Offer	Under development	Modular export of AI, DPI, secure infrastructure and cybersecurity	Provides the nearest EU analogue to an integrated external technology package.

The Commission’s own 2025 Digital Decade assessment identifies persistent strategic dependencies in semiconductors, cloud and data infrastructure and cybersecurity technologies, which is analytically important because Europe’s policy architecture is partly defensive: Brussels is attempting not only to build a competitive AI industry but also to reduce exposure to technology layers currently dominated by non-European providers.

Europe’s central structural problem is fragmentation between scale and sovereignty

European sovereignty policy contains an intrinsic tension because achieving competitive AI infrastructure requires enormous scale, whereas European digital-industrial policy simultaneously seeks to reduce reliance on the non-European hyperscalers possessing precisely that scale. The result is a difficult sequencing problem: Europe must create sufficient domestic cloud and compute capacity to make sovereign alternatives commercially credible without isolating European researchers and firms from globally competitive platforms during the transition.

The Commission’s strategy therefore increasingly combines three levels of intervention: EuroHPC provides common compute infrastructure, national governments support data centres and domestic champions, and EU regulation establishes common market and assurance requirements; if these mechanisms reinforce one another, Europe can develop a federation of national capabilities connected through continental infrastructure, whereas if they remain fragmented, firms will continue obtaining advanced compute and cloud capacity predominantly from external providers while European public investment finances parallel infrastructure that struggles to achieve comparable utilisation.

Structural advantages and vulnerabilities of the European model

Dimension	European advantage	European vulnerability
Market scale	Single market of approximately continental scale	National implementation and procurement remain fragmented
Regulation	Binding common framework	Compliance burden can slow smaller European firms
Compute	Public EuroHPC infrastructure	Commercial hyperscale capacity remains weaker than US competitors
Electricity	Diverse generation mix and several low-carbon power centres	Grid constraints and energy prices vary widely
Research	Dense university and scientific network	Commercial scaling and late-stage financing remain weaker
Industrial base	Strong manufacturing, automotive, aerospace and machinery sectors	Software/platform leadership is thinner
Development finance	Global Gateway and Team Europe mechanisms	Disbursement can be slow and institutionally complex
International partnerships	Large diplomatic and trade network	No single European external-technology command structure
Standards	Strong market-access leverage	Standards power does not automatically generate domestic suppliers
Data	Strong governance and emerging data-space architecture	Fragmentation and legal complexity can constrain availability

The EU’s external strategy implicitly acknowledges these problems by proposing a Digital Partnership Network and an integrated Tech Business Offer jointly managed by the Union and Member States, which seeks to replace fragmented individual projects with combinations of connectivity, AI, digital public infrastructure, cybersecurity and financing tailored to partner-country needs.

The EU Tech Business Offer is the critical external experiment

The strategic significance of the Tech Business Offer lies in its attempt to convert European regulatory credibility and development finance into commercial technology deployment, because the Commission identifies AI ecosystem development, secure digital infrastructure, software solutions, digital public infrastructure and cybersecurity as core components while explicitly providing for market studies, matchmaking, pilot projects and Team Europe financial support for European technology companies.

This structure addresses a long-standing European weakness: European governments and development institutions have often financed digital transformation projects while much of the commercial value was captured by non-European hardware, cloud and platform providers; the Tech Business Offer attempts to link foreign-policy objectives with market access for European technology firms, creating a mechanism through which external digital cooperation can also strengthen Europe’s own technology base.

The decisive question is execution speed, because developing countries making infrastructure choices between 2026 and 2030 are unlikely to postpone deployment while Europe completes institutional coordination, making the time required to move from diplomatic agreement to financing, procurement and operational deployment an important indicator of whether the EU can convert regulatory influence into ecosystem presence.

France is pursuing the most explicitly energy-backed compute strategy among major continental states

France’s competitive strategy is unusually clear because the government is attempting to convert the country’s relatively abundant low-carbon electricity system, transmission infrastructure, suitable data-centre sites and concentrated research base into a major European location for AI computation. At the February 2025 AI Action Summit, President Emmanuel Macron announced more than €109 billion in prospective French and foreign private investment commitments for AI infrastructure in France, with the Élysée explicitly linking the projects to data centres and the country’s electricity advantage.

The €109 billion figure should be interpreted carefully because it represents announced private investment commitments rather than completed capital expenditure, but its magnitude demonstrates that France’s strategy is based on building physical compute capacity at national scale rather than merely subsidising research laboratories.

The French government also linked compute expansion to talent formation. Macron stated in February 2025 that approximately 40,000 people per year were then being trained in AI-related fields and that France intended to raise that figure to 100,000 annually, tying infrastructure expansion directly to the human-capital pipeline required to absorb it.

France’s AI industrial proposition

Asset	Verified policy or figure	Strategic function
Announced AI investment	>€109bn	Expand data centres and AI infrastructure.
Current AI-related training output cited by presidency	~40,000/year	Existing human-capital base.
Training ambition	100,000/year	Expand AI workforce for infrastructure and applications.
Electricity proposition	Abundant, comparatively low-carbon power highlighted by presidency	Supports energy-intensive compute.
Policy objective	Extend value chain into chips, services, robotics and models	Reduce dependence on imported technology layers.

By June 2026 Macron was explicitly describing the objective as deepening the AI value chain rather than simply attracting data centres, arguing that France wanted greater chip capability, service companies, robotics firms and model developers around expanded compute infrastructure and that reducing technological dependencies formed part of the strategy.

France therefore represents the continental European strategy most closely resembling a vertically integrated industrial-policy proposition: electricity → data centres → compute → models → industrial applications → talent, although success depends on whether announced investment converts into operational capacity and whether domestic or European firms capture sufficient value above the infrastructure layer.

Germany is constructing a sovereignty strategy around industrial demand, data centres and domain-specific AI

Germany’s approach differs from France because its strongest strategic asset is not primarily the prospect of becoming Europe’s largest model-training hub but the existence of a large industrial economy capable of generating substantial demand for domain-specific AI across manufacturing, mobility, chemicals, engineering, healthcare and logistics. The German government’s September 2026 data-centre strategy states that national data-centre capacity should at least double by 2030, while the government has separately articulated an ambition to at least quadruple connection capacity dedicated to high-performance computing and AI over the same period.

This is an important quantitative shift because Germany is treating data-centre capacity as industrial infrastructure rather than simply a commercial service market, while the government’s Hightech Agenda links artificial intelligence with microelectronics, quantum technologies, biotechnology, fusion and climate-neutral technologies, thereby embedding AI within a wider technological-sovereignty programme.

The government has also stated that federal procurement should support European sovereign cloud solutions and software using open standards, which indicates that public purchasing is being considered as an instrument for ecosystem creation rather than merely administrative consumption.

Germany’s emerging architecture

Policy dimension	Verified direction	Strategic consequence
Total data-centre capacity	At least 2× by 2030	Expands domestic hosting and compute base.
HPC/AI connection capacity	At least 4× by 2030	Prioritises AI-specific infrastructure faster than general data-centre growth.
Public procurement	Support European cloud and open standards	Uses government demand to strengthen European alternatives.
Industrial strategy	Domain-specific AI lead projects	Aligns AI development with manufacturing and sector expertise.
Microelectronics	Integrated into Hightech Agenda	Addresses upstream hardware exposure
Energy policy	Reliable, affordable, low-carbon electricity identified as essential	Recognises power availability as AI-industrial constraint.

Germany’s principal vulnerability is energy economics and implementation speed, because expanding high-performance data centres requires large volumes of reliable electricity and rapid grid connections at precisely the moment when Germany is managing one of the most complex energy-system transitions among advanced economies; accordingly, the government’s ability to translate its capacity targets into competitive operating costs will be at least as important as nominal megawatt additions.

The United Kingdom is building the most commercially integrated European response

The United Kingdom has adopted a distinctive strategy based on combining domestic infrastructure expansion with unusually close relationships with US frontier-AI companies, public-sector adoption and an independent AI-security capability, thereby seeking to preserve national strategic relevance even without domestic hyperscalers equivalent to the largest US firms.

The January 2026 progress report on the AI Opportunities Action Plan states that the government had completed commitments on 38 of 50 actions, while the associated government dashboard reported £68 billion in investment pledged since January 2025, five designated AI Growth Zones, approximately 200,000 people studying AI-related higher-education programmes, and an increase in government-referenced AI compute capacity from 2 to 21 ExaFLOPs between 2024 and 2025, described by DSIT as a tenfold increase.

These numbers create one of the clearest government-monitored implementation dashboards among the jurisdictions examined, although pledged investment must again be distinguished from money already spent.

United Kingdom implementation indicators

Indicator	Government-reported value	Reference period
Action Plan commitments completed	38 of 50	Jan 2026.
Investment pledged since Jan 2025	£68bn	Reported Jan 2026.
Designated AI Growth Zones	5	Jan 2026.
AI-related higher-education students	~200,000	Government dashboard, Jan 2026.
Government-referenced AI compute capacity	2 → 21 ExaFLOPs	2024–2025.
Additional investment announced at London Tech Week 2026	>£6bn	Jun 2026.
Jobs linked to those June announcements	~8,000	Jun 2026.

The Growth Zone architecture illustrates how the UK is addressing physical bottlenecks rather than treating AI as primarily a software-sector issue, because the programme explicitly targets planning delays and access to electricity, while positioning data-centre capacity as strategically necessary for resilience and national capability.

At the same time, the UK model is unusually open to foreign technology investment. The government signed a non-binding strategic memorandum with OpenAI in July 2025 covering possible AI Growth Zone involvement, model evaluation, infrastructure investment and public-sector applications, while later cooperation with Google DeepMind connected frontier AI with scientific research and government services.

That strategy creates a deliberate trade-off: London can attract frontier-model companies without requiring domestic replication of every technological layer, but deeper involvement by US suppliers can increase British exposure to external corporate platforms; the government’s June 2026 AI Hardware Plan, centred on innovation, skills, procurement and investment in semiconductor technologies, indicates recognition that long-term strategic capacity requires deeper participation in the underlying hardware value chain.

The UK also possesses an international instrument that deserves more attention

Britain’s external response is not limited to investment attraction because the FCDO and Met Office launched an international AI-weather and climate-security partnership in June 2026 aimed at supporting countries exposed to extreme weather, alongside a £39 million SCALE programme for climate adaptation and resilience research.

This initiative is strategically relevant because it demonstrates that Western governments can compete in the same category of public-interest AI applications that China has used in meteorological diplomacy, but with a British comparative advantage built around the Met Office, climate science, insurance and development finance rather than around a centrally branded AI governance initiative.

The United Kingdom’s international AI-security diplomacy is also comparatively advanced: its AI Security Institute has become a vehicle for technical collaboration with major companies and international partners, and the 2026 Alignment Project had more than £27 million available across approximately 60 projects in eight countries, according to DSIT.

Italy’s domestic architecture is smaller, but its external-development position is strategically underappreciated

Italy does not currently possess the same scale of private AI capital commitments as France, the same data-centre expansion targets as Germany, or the same concentration of frontier firms and private investment as the United Kingdom, while the national Strategia Italiana per l’Intelligenza Artificiale 2024–2026 remains primarily organized around four domestic pillars—research, public administration, enterprises and training—rather than around a massive national compute programme.

Nevertheless, focusing exclusively on domestic model development understates Italy’s strategic assets because the Bologna supercomputing ecosystem and the country’s Africa-facing diplomacy increasingly intersect.

The EuroHPC IT4LIA AI Factory uses the Leonardo supercomputer and a forthcoming AI-optimised system at Bologna’s Tecnopolo, with EuroHPC identifying a 50 Exaflop/s target and service coverage spanning agri-tech, cybersecurity, earth sciences, healthcare, education and finance; the programme therefore provides Italy with a significant public compute asset inside the wider European architecture even though Italy lacks a US-style hyperscale platform company.

Italy’s two-track strategic position

Track	Instrument	Verified status	Strategic significance
Domestic AI strategy	Italian AI Strategy 2024–2026	Research, PA, enterprises, training	Builds adoption and governance base.
European compute	IT4LIA AI Factory	50 Exaflop/s target	Provides advanced compute access through Bologna.
Africa-facing AI diplomacy	AI Hub for Sustainable Development	G7-backed, Italy/UNDP architecture	Connects Italian policy with African AI ecosystems.
Compute access for African innovators	Compute & Talent Acceleration Programme	Operational by 2026	Links development cooperation to actual compute access.
Startup/industry engagement	AI Hub ecosystem	130 African innovators and 150+ Italian organisations reported engaged by Mar 2026	Builds commercial and technical networks.
Compute mobilised	1.5 million GPU-hours	Reported Mar 2026	Concrete infrastructure input rather than training alone.
Cloud support	US$3m cloud credits allocated	Reported Mar 2026	Reduces deployment cost for participating African innovators.
Partnerships catalysed	30+	Reported Mar 2026	Evidence of commercial/institutional follow-through.
New Bologna partnerships	7 agreements	Jun 2026	Moves selected projects into formal cooperation.

The March 2026 official Italian diplomatic record is particularly significant because it provides concrete implementation metrics: 130 African innovators formally engaged, more than 150 Italian companies and organisations involved, more than 30 partnerships catalysed, 1.5 million GPU-hours mobilised and US$3 million in cloud credits allocated.

These figures remain small relative to global hyperscale infrastructure markets, but they show that Italy is developing an international proposition focused not on exporting an Italian foundation model but on compute access, talent, trust, financing and application development, which could become strategically important if integrated systematically with EuroHPC and Team Europe financing.

The June 2026 signing of seven partnerships in Bologna between African innovators and Italian or European organisations extended the initiative into agriculture, health, linguistic AI, education and AI infrastructure, while the associated programme explicitly addressed energy-efficient workloads, modular data centres, advanced cooling and data governance for African deployments.

Italy therefore occupies an unusual position within Europe: it is not the largest European AI-industrial actor, but it has developed one of the most concrete EU-member-state mechanisms for connecting European compute capacity with Global South innovation policy, particularly in Africa.

Italy’s AI Hub is the closest European national analogue to China’s capacity-building diplomacy

The AI Hub for Sustainable Development was developed under Italy’s 2024 G7 presidency with UNDP and incorporated into the Mattei Plan framework, while the G7 Foreign Ministers explicitly endorsed its focus on agriculture, health, infrastructure, education and training, water and energy.

The current programme architecture identifies five foundation layers—data, green compute, talent, trust and financing—and includes dedicated mechanisms such as a Compute Accelerator Programme, AI Infrastructure Builder Programme and Cyber4Africa, which gives the initiative a deeper ecosystem structure than a conventional development-aid training scheme.

The Italian approach differs materially from the Chinese one because it is embedded in G7, UNDP and Team Europe structures and explicitly emphasizes African-developed solutions, local innovators and multi-partner financing rather than an Italy-centred technology stack; whether that structure produces greater recipient-country autonomy or simply slower implementation will be determined by how effectively projects scale beyond pilot and matchmaking stages.

France, Germany, Italy and the UK are not pursuing the same European strategy

Treating “Europe” as a unitary AI actor obscures important national specialization because the four major European states examined here are increasingly occupying different positions in the emerging value chain.

National strategic specialization

Country	Emerging comparative role	Primary instrument	Main strength	Principal constraint
France	Compute and frontier-scale investment hub	Large data-centre commitments + talent expansion	Electricity profile, research and capital attraction	Conversion of commitments into operational capacity
Germany	Industrial AI and sovereign infrastructure hub	Data-centre expansion + domain-specific AI + open standards	Manufacturing demand and industrial base	Energy cost and implementation complexity
United Kingdom	Frontier-company integration and AI-services hub	Growth Zones, foreign investment, AISI, hardware plan	Capital attraction, research, global firms, English-language ecosystem	Dependence on foreign cloud/model providers
Italy	EuroHPC and Global South application bridge	IT4LIA + AI Hub for Sustainable Development	Bologna compute ecosystem and Africa-facing diplomacy	Smaller national capital and commercial-platform scale
EU level	Continental infrastructure, regulation and external offer	AI Factories, Gigafactories, AI Act, Tech Business Offer	Market scale and public infrastructure	Institutional fragmentation and deployment speed

The emergence of specialization need not be a weakness if the assets are integrated at European level: French low-carbon compute, German industrial demand, British research and security expertise, Italian EuroHPC and Africa-facing programmes, and EU financing and regulatory frameworks can theoretically form a diversified continental ecosystem stronger than any individual national programme.

The problem is coordination because these assets are controlled by different governments, agencies and companies and do not automatically combine into a coherent commercial offer.

European exposure remains greatest at the infrastructure and scaling layers

Europe’s vulnerability is increasingly concentrated in three areas: advanced semiconductors, hyperscale commercial cloud infrastructure and the capital intensity required to scale frontier AI companies globally. The Commission itself identifies semiconductors and cloud/data infrastructure among the Union’s persistent strategic dependencies, confirming that this is not merely an external analytical judgment.

A useful distinction should therefore be made between scientific capacity, where Europe remains strong, and platform scale, where external dependencies remain much greater.

European capability versus exposure

Layer	European position	Exposure
Basic AI research	Strong	Low/moderate
Scientific HPC	Strong and expanding	Moderate
Frontier commercial compute	Expanding but insufficient	High
Accelerator design/manufacture	Selective European strengths but no complete frontier stack	High
Hyperscale cloud	Significant European providers but much smaller than US leaders	Very high
Foundation models	Competitive firms exist	Moderate/high
Industrial AI	Strong sectoral opportunity	Moderate
Regulation	Global influence	Low
Standards	Strong European participation	Low/moderate
Development finance	Significant	Low
International tech deployment	Growing but fragmented	Moderate/high
Late-stage technology capital	Improving but weaker than US	High

This distribution explains why European policy increasingly emphasizes sovereignty without autarky: complete replication of every layer would be economically unrealistic, but dependence on a single external jurisdiction across several bottleneck technologies would create substantial strategic exposure.

The Western development-finance architecture is considerably larger than the AI-specific programmes suggest

Comparisons with Chinese AI diplomacy can become misleading if only AI-labelled programmes are counted, because Western infrastructure and development initiatives sit inside much larger financing frameworks that can potentially be directed toward digital infrastructure.

The G7 Partnership for Global Infrastructure and Investment maintains the collective objective of mobilising up to US$600 billion by 2027, while the European Union’s Global Gateway seeks to mobilise up to €300 billion across digital, energy, transport, health, education and research infrastructure. These are not AI-specific allocations and therefore must not be presented as AI funding, but they constitute financing platforms through which connectivity, data centres, digital public infrastructure and energy systems relevant to AI deployment can be supported.

The strategic question is whether these very large financial frameworks can be connected operationally to the smaller AI-specific initiatives quickly enough to create deployable technology packages.

The international competition is moving from standards diplomacy to implementation diplomacy

The next phase of Western strategy will therefore be decided less by publication of additional principles than by the ability to deliver operational projects in countries that currently lack sufficient compute, energy, data and specialist personnel.

The European International Digital Strategy explicitly acknowledges this by moving toward market studies, company matchmaking, pilots and financial support through the EU Tech Business Offer rather than limiting external digital policy to regulatory dialogue.

Italy’s African AI Hub represents an early operational test of this model, because it already combines private firms, public institutions, compute access and cloud credits; the United Kingdom’s climate-AI partnership provides another sectoral model; and American full-stack export policy represents the most commercially powerful version if government agencies succeed in aligning financing with private technology providers.

Capital mobilisation differs profoundly across the Atlantic

The United States benefits from private technology companies whose annual infrastructure expenditure can exceed the public technology budgets of many states, meaning federal policy frequently acts by enabling, protecting or internationalising private capital rather than substituting for it.

France is attempting to replicate some of this scale through foreign and domestic private commitments, while the UK is using planning reform and Growth Zones to attract global investment, Germany is attempting to mobilise national industrial and data-centre capital around sovereign infrastructure, and the EU relies more heavily on blended public-private mechanisms such as InvestAI and EuroHPC.

Capital architecture comparison

Actor	Dominant capital mechanism	Verified flagship figure	Interpretation
United States	Private hyperscaler/semiconductor capital supported by federal policy	No single government AI-capital envelope captures the system	Commercial scale is itself strategic infrastructure
European Union	Public/private mobilisation	€20bn targeted for AI Gigafactories	Continental-scale intervention aimed at frontier compute.
France	Foreign + domestic private commitments	>€109bn announced	National strategy seeks rapid infrastructure concentration.
United Kingdom	Global private investment attracted through Growth Zones and partnerships	£68bn pledged since Jan 2025 as reported Jan 2026	Strong investment-attraction model.
Germany	Industrial investment + federal infrastructure strategy	Capacity targets more explicit than one aggregate AI investment number	Focus on infrastructure and industrial adoption
Italy	EU compute + national/EU/G7 development cooperation	1.5m GPU-hours + US$3m cloud credits mobilised for African AI Hub ecosystem	Smaller financial scale but concrete external deployment.

These numbers are not directly comparable because they represent different categories—investment commitments, compute resources, public-facility mobilisation and capacity targets—so they should not be ranked as equivalent measures of national AI power.

Regulation is becoming an industrial instrument rather than a separate European pillar

The European AI Act remains important, but its strategic significance increasingly lies in how regulatory requirements interact with infrastructure, standards and procurement rather than in regulation considered separately. A harmonised compliance environment can create a common market for assurance services, trusted datasets, documentation tools and compliant AI applications, potentially giving European companies a commercial advantage if those requirements become accepted internationally.

However, there is an opposite pathway in which compliance costs fall disproportionately on smaller European firms while large external firms absorb them easily, thereby increasing rather than reducing market concentration; the economic outcome will depend on implementation details, standardisation, supervisory consistency and availability of testing infrastructure.

Germany’s explicit discussion of reducing unnecessary regulatory burdens while preserving competition and core protections illustrates that this tension has moved into mainstream industrial policy rather than remaining confined to legal debate.

The energy constraint will increasingly determine European AI geography

Compute expansion is transforming electricity availability into a technology-policy variable because large AI data centres require reliable power, grid connections, cooling and substantial local infrastructure, meaning that national AI competitiveness is increasingly linked to energy-system characteristics.

France emphasizes relatively abundant low-carbon electricity; Germany is explicitly incorporating energy availability and grid costs into its data-centre strategy; the United Kingdom identifies electricity connection delays as one of the primary bottlenecks AI Growth Zones are intended to solve; and Italy can potentially exploit the Bologna supercomputing cluster while broader national expansion remains constrained by the economics and geography of large-scale power supply.

Energy-compute exposure

Country	AI-energy proposition	Principal opportunity	Principal constraint
France	Low-carbon electricity highlighted as data-centre advantage	Large-scale compute attraction	Grid delivery and project completion
Germany	Grid, energy cost and sovereign infrastructure integrated in strategy	Industrial proximity	Electricity cost and connection availability
UK	Growth Zones explicitly target power bottlenecks	Fast-track concentrated infrastructure	Grid queues and regional capacity
Italy	Supercomputing cluster at Bologna	Public HPC and specialised ecosystem	Scaling beyond public/European infrastructure
EU	Sustainability criteria integrated into AI Factories/Gigafactories	Coordinate energy-efficient compute	National energy-market fragmentation

This means the European AI map will increasingly follow electricity infrastructure as much as university geography, potentially concentrating frontier compute in locations offering combinations of abundant power, grid capacity, fibre connectivity and streamlined permitting.

Talent strategy increasingly mirrors infrastructure strategy

France’s objective of moving from approximately 40,000 to 100,000 AI-trained people annually and the United Kingdom’s reported 200,000 students enrolled in AI-related higher-education programmes show that governments increasingly understand compute investment to be ineffective without human-capital expansion.

Germany’s domain-specific strategy similarly depends on retraining engineers and industrial personnel, while Italy’s strategy places training among its four national pillars and increasingly connects Italian expertise with African programmes through the AI Hub.

The strategic issue is therefore not simply producing AI researchers but creating sufficient numbers of deployment engineers, data specialists, cybersecurity professionals, sectoral experts, infrastructure operators, regulators and public administrators to convert frontier research into widespread application.

Europe’s international offer can become more competitive if it integrates its comparative advantages rather than replicating the United States

Europe is unlikely to outperform the United States by recreating an identical commercial hyperscaler model in every jurisdiction, and it is equally unlikely to replicate China’s state-centric international deployment architecture; a more plausible European proposition would combine features that are already institutionally present: trusted digital infrastructure, diversified suppliers, sovereign deployment options, public-interest applications, privacy and security assurance, EuroHPC compute access, development finance and local capacity building.

The EU Tech Business Offer provides the institutional framework through which such a strategy can emerge, while Italy’s AI Hub provides an early illustration of how local innovators in developing countries can be connected to European compute, firms and financing rather than simply sold finished European products.

This architecture can become particularly attractive where governments seek diversification between American and Chinese suppliers rather than exclusive geopolitical alignment.

Italy could become a European bridge to Africa if its programmes are integrated with EU scale

The most significant opportunity identified in the European country comparison concerns Italy because its AI Hub architecture overlaps directly with one of China’s strongest diplomatic propositions: AI capacity-building in emerging economies.

Unlike the much larger Chinese training and institutional architecture described previously, however, Italy’s programme benefits from connections to UNDP, the G7, Team Europe, Bologna’s supercomputing cluster and European industrial actors, offering the possibility of an alternative model based on access to multi-provider infrastructure rather than adoption of one national technology stack.

The 1.5 million GPU-hours already mobilised demonstrate that the programme has moved beyond purely diplomatic language, while the 30-plus partnerships and subsequent seven Bologna agreements show increasing commercial follow-through.

Its limitation is scale: unless Italy can connect this initiative systematically to EuroHPC, Global Gateway, European cloud infrastructure and development-finance resources, it will remain a promising specialist programme rather than a strategic alternative capable of competing with larger ecosystems.

Western coordination is strongest where security incentives coincide and weaker where commercial interests diverge

The United States, United Kingdom and European states share broad concerns regarding infrastructure security, semiconductor supply chains, AI safety and technological dependence, but their commercial incentives are not identical because US firms benefit economically when European governments procure American cloud and model services, whereas European industrial policy increasingly seeks to reduce precisely that dependence.

This creates a structural distinction between allied geopolitical coordination and commercial technology competition: Europe and the United States can cooperate on security standards, export controls or trusted supply chains while simultaneously competing for cloud contracts, data-centre investment, model adoption and standard-setting influence.

The UK illustrates this ambiguity particularly clearly because it seeks strategic autonomy in compute and hardware while simultaneously deepening partnerships with OpenAI and Google DeepMind; rather than being internally contradictory, this reflects a deliberate policy of attracting foreign frontier capability while constructing selective domestic sovereignty around infrastructure, safety evaluation, skills and hardware.

Comparative implementation matrix

Dimension	United States	EU	UK	France	Germany	Italy
Frontier commercial models	Very strong	Fragmented but growing	Strong presence through domestic/foreign labs	Strong national challenger ecosystem	More domain-focused	Smaller
Advanced compute	Very strong commercially	Strong public expansion	Rapid expansion	Major investment push	Major expansion planned	Strong EuroHPC node
Hyperscale cloud	Very strong	External dependence remains high	External providers dominant	External providers important	External + sovereign alternatives	External + national/EU infrastructure
National industrial demand	Very high	Very high collectively	High services/research demand	High	Very high manufacturing demand	Strong SME/manufacturing base
Public AI infrastructure	Significant but commercially dominated	Extensive EuroHPC architecture	Expanding sovereign compute	Expanding	Expanding	IT4LIA/Leonardo
International deployment strategy	Explicit full-stack export	Tech Business Offer	Partnerships + security + development programmes	Diplomatic and investment influence	Less externally oriented	Africa AI Hub
Development-country AI instrument	Increasing	Global Gateway/Team Europe	Climate and development programmes	Selective	Selective	G7/UNDP AI Hub
Standards power	Very high	Very high	High	Through EU/international bodies	Through EU/international bodies	Through EU/international bodies
Main comparative strength	Commercial stack scale	Regulatory + market + public infrastructure	Frontier-company integration	Power + investment attraction	Industrial application	Africa/EuroHPC bridge
Main strategic exposure	International resistance to dependence	Fragmentation and external platforms	US-platform dependence	Investment execution	Energy cost	Scale and capital

Decision-relevant implications for Europe

Europe’s strategic challenge is not to produce a European equivalent of every American or Chinese firm, because attempting full technological autarky would require enormous capital and would sacrifice the benefits of global specialization; the more achievable objective is to guarantee credible alternatives at every systemic bottleneck, ensuring that no single external jurisdiction possesses unavoidable control over compute, cloud, models, critical data, certification or public-sector infrastructure.

The strongest European response would therefore connect AI Factories and future Gigafactories with national data-centre strategies, interoperable European cloud infrastructure, common procurement requirements, an internationally deployable Tech Business Offer and development-finance mechanisms capable of scaling external projects.

France can supply significant compute and energy-intensive infrastructure; Germany can anchor industrial adoption; the UK, while outside the EU, remains strategically relevant through frontier research, AI safety and private capital; Italy can connect European compute and commercial capabilities to African technology ecosystems; while EU institutions can provide regulation, standards and financing capable of aggregating these national assets.

Without that integration, Europe risks remaining simultaneously influential in regulation and dependent in infrastructure, a position that provides rule-making leverage but insufficient control over the technological systems to which those rules apply.

Key judgments

The United States has formally recognized international AI adoption as a strategic objective and is building an explicit full-stack export policy, meaning that Western strategy can no longer be accurately described as focused only on chips or frontier-model benchmarks.

The European Union has moved decisively toward an industrial AI policy, with 19 AI Factories already deployed by April 2026, 13 regional antennas, approximately €10 billion associated with the AI Factory programme and a €20 billion investment mobilisation objective for up to five Gigafactories, while the International Digital Strategy adds an external technology-deployment architecture.

France currently possesses the clearest continental strategy for attracting large-scale AI infrastructure, supported by more than €109 billion in announced investment commitments and an explicit plan to expand annual AI-related training from approximately 40,000 to 100,000 people, although announced capital should not be confused with completed infrastructure.

Germany is constructing an industrial-sovereignty model around a target to at least double overall data-centre capacity by 2030 and quadruple HPC/AI connection capacity, using public procurement, open standards and domain-specific AI to connect infrastructure with manufacturing competitiveness.

The United Kingdom is implementing one of Europe’s fastest investment-led strategies, reporting £68 billion in pledged investment, five AI Growth Zones and an increase in AI compute capacity from 2 to 21 ExaFLOPs between 2024 and 2025, while relying extensively on partnerships with US frontier firms and simultaneously building domestic hardware and AI-safety capabilities.

Italy remains smaller in overall AI capital scale but has developed a strategically differentiated position through the IT4LIA AI Factory and the AI Hub for Sustainable Development, whose official March 2026 figures—130 African innovators, more than 150 Italian organizations, more than 30 partnerships, 1.5 million GPU-hours and US$3 million of cloud credits—provide one of the more concrete European examples of AI ecosystem diplomacy directed toward the Global South.

Europe’s most significant vulnerability remains the gap between its many policy instruments and its ability to combine them into a coherent international commercial offer, particularly because strategic dependencies persist in semiconductors, cloud and data infrastructure even as public compute capacity expands.

What would change the assessment

The assessment of an increasingly credible European response would strengthen substantially if the first AI Gigafactories reach operational status on schedule, if AI Factory utilisation demonstrates sustained industrial demand rather than predominantly research use, and if the EU Tech Business Offer produces multi-country deployments combining European AI, secure infrastructure, digital public infrastructure and financing.

It would weaken if AI Factory and Gigafactory projects experience major delays, if European firms remain unable to scale despite public infrastructure expansion, or if most additional European data-centre capacity is ultimately controlled by external hyperscalers without meaningful improvements in workload portability or European supplier participation.

France’s position would strengthen if a material share of the announced €109 billion converts into commissioned compute infrastructure and generates downstream European model, cloud and application businesses rather than primarily hosting foreign hyperscalers, while substantial cancellation or delay of those projects would weaken the current assessment.

Germany’s position would strengthen if its planned doubling of data-centre capacity and quadrupling of AI/HPC capacity is accompanied by competitive electricity costs and measurable adoption of domain-specific AI across the Mittelstand and major industrial sectors, rather than infrastructure expansion alone.

The UK position would strengthen if Growth Zones convert pledged investment into installed compute and if domestic hardware, research and AI-security institutions capture increasing value alongside foreign frontier companies; it would weaken if infrastructure investment principally increases British dependence on externally controlled cloud and model ecosystems.

Italy’s position would strengthen particularly rapidly if the AI Hub for Sustainable Development is connected systematically to the IT4LIA AI Factory, EuroHPC and Global Gateway financing, because this would transform an emerging Africa-facing innovation programme into a broader Europe-to-Global-South compute and technology channel.

Open official record

A fully comparable dataset of actually commissioned, rather than announced, AI data-centre capital expenditure across the United States, France, Germany, Italy and the United Kingdom is not available in the official sources reviewed, meaning headline investment announcements cannot be compared directly without project-by-project verification.

The final capital structure, location, ownership, accelerator composition and operating arrangements of all planned European AI Gigafactories remain incomplete because the programme is still being implemented, preventing a definitive comparison with American hyperscale infrastructure.

The EU Tech Business Offer remains strategically promising but still requires a larger observable portfolio of funded deployments before its operational effectiveness can be assessed against Chinese and American technology diplomacy.

Germany’s capacity targets are explicit, but a complete publicly consolidated investment figure comparable with the French €109 billion announcement or UK £68 billion pledged-investment figure is not available and should therefore not be artificially reconstructed.

Italy’s AI Hub provides unusually useful implementation metrics, but official public records do not yet provide a comprehensive country-by-country inventory of the African firms receiving GPU hours, cloud credits and financing, the duration of that support, or the percentage of partnerships converting into commercially sustainable deployment.

The most important unresolved European variable is consequently not whether governments now understand the strategic importance of AI ecosystems, because the documentary record establishes that they do, but whether continental policy, national infrastructure and private capital can be integrated faster than competing ecosystems become entrenched in third-country markets.

---

Copyright of debuglies.com - Even partial reproduction of the contents is not permitted without prior authorization Reproduction reserved