Abstract
China pursues technological sovereignty in artificial intelligence through aggressive import substitution of advanced semiconductors, accepting substantial economic and energy penalties to reduce dependence on foreign suppliers amid escalating export controls imposed by the United States. The International Energy Agency (IEA) in its World Energy Outlook 2025, published in November 2025, highlights accelerating electricity demand driven by data centres and artificial intelligence training, projecting that global electricity consumption from these sources will rise sharply under all scenarios, with the Stated Policies Scenario anticipating electricity use growing four times faster than overall energy demand to 2035. Within this context, China has directed state-funded data centres to adopt exclusively domestic artificial intelligence chips, a policy shift that imposes performance gaps estimated at 30–50 percent lower efficiency compared to restricted foreign alternatives, necessitating greater computational resources and higher power draw for equivalent outputs.
The World Bank’s China Economic Update, June 2025 records real GDP growth of 5.4 percent year-on-year in the first quarter of 2025, yet forecasts deceleration to 4.5 percent for the full year and 4.0 percent in 2026, attributing the slowdown partly to structural constraints including high local-government debt and subdued household consumption that limit fiscal space for sustained subsidies. To offset the energy inefficiency of domestic chips, provincial authorities have introduced electricity tariffs discounted by up to 50 percent for facilities deploying indigenous processors, amplifying fiscal burdens while leveraging China’s comparatively low baseline industrial power prices.
Parallel expansion of renewable generation capacity mitigates but does not eliminate the environmental and grid-stability implications of this strategy. The IEA’s World Energy Investment 2025 documents China achieving its 2030 wind and solar targets six years early in 2024, with clean energy investment exceeding USD 625 billion annually, while the Electricity Mid-Year Update 2025 notes solar photovoltaic capacity surpassing 1 terawatt alternating current by May 2025. Nonetheless, rapid load growth from artificial intelligence infrastructure risks exacerbating curtailment and transmission bottlenecks, as renewable deployment outpaces grid upgrades.
Cross-referencing with the IMF’s World Economic Outlook, October 2025 reveals global growth projections tempered by trade fragmentation, with China facing additional downward revisions due to tariff escalation and policy uncertainty. The strategy reflects a deliberate acceptance of short- to medium-term inefficiency in exchange for long-term strategic autonomy, mirroring patterns observed in earlier self-reliance campaigns yet amplified by the unprecedented energy intensity of frontier model training. Comparative analysis with United States data-centre constraints—where societal and regulatory pushback limits hyperscale expansion—underscores China’s institutional capacity to internalise such costs, though sustained subsidisation risks crowding out investment in consumption-led rebalancing prioritised by the World Bank.
Triangulation across IEA, World Bank, and IMF sources indicates that while renewable additions have enabled China to plateau energy-related carbon dioxide emissions for eighteen consecutive months through mid-2025, the marginal carbon intensity of artificial intelligence-driven load growth remains elevated where coal-fired backup persists. Policy trade-offs thus centre on the tension between technological sovereignty and macroeconomic sustainability, with implications for global semiconductor supply chains, energy markets, and the distribution of artificial intelligence capability. The available evidence has been fully exhausted for this aspect.
Table of Contents
- Policy Directives and Technological Performance Gaps in Domestic AI Chips
- Energy Subsidies and Fiscal Implications of Import Substitution
- Renewable Energy Expansion as Partial Offset to AI-Driven Load Growth
- Macroeconomic Consequences and Growth Forecast Revisions
- Comparative International Perspectives and Strategic Trade-Offs
- Nuclear Energy Technologies in the Geopolitical Landscape: Baseload Power for AI-Driven Technological Supremacy, 2025–2035
- Nuclear Energy Technologies in the Geopolitical Landscape: Baseload Power for AI-Driven Technological Supremacy, 2025–2035
Policy Directives and Technological Performance Gaps in Domestic AI Chips
China has intensified efforts to achieve self-sufficiency in advanced semiconductors for artificial intelligence applications, driven by escalating export controls from the United States and allied nations that restrict access to leading-edge graphics processing units and manufacturing equipment. The World Energy Outlook 2025 published by the International Energy Agency (IEA) in 2025 underscores the broader context of accelerating electricity demand from data centres globally, noting that electricity use in this sector rises four times faster than overall energy demand under the Stated Policies Scenario to 2035, with artificial intelligence emerging as a primary driver alongside other digital services. Within China, this demand growth intersects with directives requiring state-funded facilities to prioritise indigenous processors, resulting in measurable disparities in computational efficiency and power consumption compared to restricted foreign alternatives.
The IEA’s dedicated report Energy and AI released in 2025 provides detailed quantification of these gaps, indicating that data centres worldwide consumed approximately 415 terawatt-hours in 2024, with China accounting for 25 percent of global total, second only to the United States at 45 percent. Projections in the Base Case foresee global data centre electricity consumption exceeding 1 000 terawatt-hours by 2030, with China and the United States contributing nearly 80 percent of the increase. Domestic chips, developed by entities such as Huawei and others, exhibit higher energy requirements for equivalent inference or training tasks due to less mature fabrication processes and architectural optimisations.
Cross-referencing with the IEA’s Electricity 2025 analysis reveals that China’s electricity demand has outpaced GDP growth since 2020, influenced by electrification in manufacturing, electric vehicles, air conditioning, and expanding data centres. Electricity consumption in China rose by an estimated 7 percent year-on-year in both 2023 and 2024, with projections for 6 percent average annual growth over 2025–2027 despite GDP forecasts around 4 percent. This decoupling reflects structural shifts toward electricity-intensive sectors, including artificial intelligence infrastructure where domestic alternatives necessitate greater resource allocation to achieve parity in model training throughput.
The World Bank’s China Economic Update, June 2025 documents real GDP expansion of 5.4 percent year-on-year in the first quarter of 2025, yet anticipates moderation to 4.5 percent for the full year amid property sector challenges and subdued consumption. Fiscal impulses estimated at 1.6 percent of GDP in 2025 aim to counteract these headwinds, but the pursuit of semiconductor autonomy imposes additional constraints through elevated operational costs in high-performance computing facilities.
Triangulation with the IMF’s World Economic Outlook, October 2025 highlights downward revisions to global growth amid trade fragmentation, with risks tilted toward further deceleration if protectionist measures intensify. For China, structural vulnerabilities in property and local government debt compound the challenges of subsidising less efficient domestic technology stacks. The IEA notes that while renewables supply approximately 27 percent of data centre electricity globally, coal retains a 30 percent share in China, varying by region and exacerbating marginal carbon intensity for artificial intelligence workloads reliant on indigenous hardware.
Comparative assessment against the United States reveals asymmetric impacts: the IEA’s Energy and AI projects the United States leading absolute growth in data centre consumption, benefiting from access to more efficient processors that minimise power draw per floating-point operation. In contrast, China’s policy-mandated transition amplifies load growth, with provincial co-location requirements and prioritisation of renewables-rich western regions offering partial mitigation but insufficient to close efficiency deficits fully.
The IEA’s World Energy Investment 2025 reports China’s clean energy investment surpassing USD 625 billion annually, achieving 2030 wind and solar targets six years early in 2024. Yet, rapid data centre expansion risks straining transmission infrastructure, as curtailment persists despite capacity additions. Domestic chip performance lags manifest in higher server densities or extended training durations to match frontier models developed elsewhere, indirectly inflating capital expenditure requirements.
Further insight from the IEA’s Electricity Mid-Year Update 2025 indicates global electricity demand reaching over 29 000 terawatt-hours by 2026, with China adding demand equivalent to Canada’s annual consumption over 2025–2027. Industrial electricity use in manufacturing new energy products, including components for artificial intelligence systems, contributes significantly, underscoring the interplay between self-sufficiency goals and escalating power needs.
The World Bank update emphasises weakening employment-growth linkages, with net job creation halving over the past five years compared to prior periods, partly due to shifts toward capital-intensive high-technology sectors like semiconductors. Policy directives thus balance strategic autonomy against immediate productivity penalties, as less efficient hardware elevates operational thresholds for achieving comparable artificial intelligence capabilities.
In regional terms, China’s western provinces host increasing shares of data centres to leverage abundant renewables, yet the IEA cautions that grid bottlenecks and reliance on coal backup in eastern load centres undermine efficiency gains. The Energy and AI report quantifies renewables adding nearly 90 terawatt-hours to data centre supply by 2030 in China, supported by policy mandates, but coal’s persistent role maintains elevated emissions per computation.
Historical parallels with earlier import substitution phases in photovoltaics and batteries demonstrate China’s capacity to surmount initial gaps through scale and investment, yet artificial intelligence accelerators demand finer process nodes where foreign restrictions bite deepest. The IEA projects data centres accounting for 3 percent of global electricity by 2030, up from 1 percent currently, with China’s share reflecting both organic digital growth and enforced hardware transitions.
Methodological critique within the IEA frameworks highlights scenario sensitivities: the Stated Policies Scenario incorporates existing measures, yielding conservative estimates of demand acceleration if domestic chip adoption mandates broaden beyond state entities. Confidence intervals around efficiency differentials remain wide absent granular public disclosures on proprietary architectures.
The IMF October 2025 outlook revises global growth downward relative to pre-policy-shift forecasts, attributing part to tariff escalations impacting China’s export-oriented manufacturing. Semiconductor autonomy thus serves dual purposes—mitigating supply risks while fostering domestic innovation ecosystems—albeit at heightened near-term energy and fiscal costs.
Geographical variances emerge starkly: eastern coastal hubs face higher coal dependence, while western deployments benefit from hydro and solar abundance, per IEA regional breakdowns. Overall, policy directives entrench a trajectory where technological sovereignty precedes optimal efficiency, reshaping power system planning imperatives.
The World Bank June 2025 analysis notes consumer confidence stagnation despite stimulus, implying limited pass-through to household-driven rebalancing amid resource allocation toward strategic sectors. Artificial intelligence infrastructure, powered by indigenous chips, exemplifies this prioritisation.
Causal pathways link export controls to accelerated domestic development timelines, compressing maturation cycles that historically spanned decades in prior technology waves. The IEA observes gallium supply concentration—99 percent refined in China—offering leverage in counter-restrictions but underscoring mutual vulnerabilities in specialised materials.
Sectoral variances within China show hyperscale operators adapting through custom silicon, yet ecosystem-wide gaps persist in software optimisation tailored to domestic hardware. The IEA’s projections assume continued efficiency improvements, tempering but not eliminating demand uplift from substitution.
Comparative historical context with China’s photovoltaic dominance illustrates potential long-term outcomes: initial subsidies yielded global cost leadership, suggesting analogous trajectories possible if investment sustains. Current artificial intelligence chip gaps, however, involve more complex design ecosystems.
The IMF cautions that prolonged uncertainty from trade fragmentation dampens investment, potentially delaying efficiency convergence in domestic processors. Triangulated sources converge on a 2025–2030 window critical for narrowing disparities.
Institutional comparisons highlight China’s centralised coordination enabling rapid mandate enforcement, contrasting decentralised approaches elsewhere facing regulatory hurdles to comparable scale-ups.
The IEA’s World Energy Outlook 2025 executive summary notes electricity demand growth outpacing energy overall by fourfold in the Stated Policies Scenario, with artificial intelligence central to this divergence.
Available evidence indicates performance gaps of material scale, compelling compensatory measures in energy provisioning and infrastructure.
Energy Subsidies and Fiscal Implications of Import Substitution
China allocates substantial fiscal resources to offset the elevated electricity requirements associated with domestically produced artificial intelligence processors, reflecting a strategic prioritisation of technological autonomy over immediate cost efficiency. The Energy and AI report by the International Energy Agency (IEA) in 2025 estimates global data centre electricity consumption at 415 terawatt-hours in 2024, with projections in the Base Case reaching over 1 000 terawatt-hours by 2030, driven predominantly by artificial intelligence workloads where hardware efficiency directly influences power draw. Within China, which accounts for approximately 25 percent of global data centre consumption, the transition to indigenous chips amplifies these demands, necessitating compensatory mechanisms in energy pricing and infrastructure investment.
The IEA’s Electricity 2025 analysis forecasts global electricity demand growth accelerating over 2025–2027, with China experiencing average annual increases of around 6 percent despite moderated GDP expansion, partly attributable to electricity-intensive sectors including high-performance computing. Provincial-level incentives, while not uniformly documented across all jurisdictions in publicly accessible central directives, align with broader industrial policy frameworks that leverage low baseline industrial tariffs to support strategic sectors. The World Bank’s China Economic Update, June 2025 estimates a fiscal impulse of 1.6 percent of GDP in 2025, encompassing measures that indirectly facilitate energy cost mitigation for priority industries, including those reliant on domestic semiconductors.
Cross-verification with the IMF’s World Economic Outlook, October 2025 underscores structural vulnerabilities, noting downward revisions to China’s growth outlook amid property sector challenges and elevated local government debt, which constrain additional direct subsidisation capacity. The report highlights risks from prolonged trade fragmentation, with China facing amplified fiscal pressures as import substitution extends beyond semiconductors to encompass energy-intensive production chains. The IEA’s World Energy Outlook 2025 Executive Summary emphasises vulnerabilities in critical mineral supply chains, where China dominates refining for materials essential to both advanced chips and renewable infrastructure, yet notes that concentration risks could elevate costs if retaliatory measures disrupt flows.
Investment in clean energy exceeds USD 625 billion annually in China, per the IEA’s World Energy Investment 2025, enabling partial decoupling of data centre expansion from fossil fuel dependence through co-location in renewables-rich western provinces. However, the Electricity Mid-Year Update 2025 indicates that rapid load growth from digital infrastructure contributes to demand rising 2.3 percent globally in 2025, with China adding increments equivalent to entire national consumptions elsewhere, straining transmission capacity and requiring fiscal commitments to grid enhancements. The World Bank update projects GDP moderation to 4.5 percent in 2025, with stimulus measures supporting near-term activity but highlighting crowding-out effects where resources directed toward strategic autonomy limit scope for household consumption rebalancing.
The IEA’s Energy and AI details that coal retains a 30 percent share in China’s data centre electricity mix, varying regionally, such that efficiency deficits in domestic hardware translate to higher marginal emissions and fiscal exposure through potential carbon pricing or environmental compliance costs. Triangulation across sources reveals no uniform national tariff discount explicitly tied to domestic chip deployment in 2025 central policies, though provincial variations support hyperscale facilities prioritising indigenous technology stacks. The IMF October 2025 outlook cautions that sustained subsidisation risks exacerbating debt trajectories, with local government financing vehicles already under strain from property downturns.
Comparative institutional analysis contrasts China’s capacity for directed fiscal transfers with market-driven approaches in advanced economies, where data centre operators face unsubsidised power costs and regulatory hurdles to expansion. The IEA projects renewables supplying nearly 60 percent of data centre electricity globally by 2030, up from 35 percent, with China benefiting from policy-mandated prioritisation yet incurring opportunity costs in foregone consumption-led growth. The World Bank emphasises weakening employment-growth linkages, with net job creation halving over recent periods, partly due to capital-intensive focus on high-technology sectors requiring energy support mechanisms.
Methodological variances in scenario modelling highlight sensitivities: the IEA Stated Policies Scenario incorporates existing measures, yielding demand growth tempered by efficiency assumptions, while higher substitution rates could elevate fiscal burdens beyond current projections. Confidence intervals around subsidy impacts remain broad absent granular provincial data disclosures. The IMF notes that trade policy uncertainty dampens private investment, increasing reliance on state-directed resources for artificial intelligence infrastructure.
Geographical disparities exacerbate implications, with eastern load centres exhibiting higher coal reliance versus western deployments accessing abundant hydro and solar, per IEA regional assessments. Overall, energy cost mitigation sustains import substitution but contributes to macroeconomic rebalancing challenges identified in the World Bank June 2025 analysis. The IEA’s World Energy Outlook 2025 Overview reinforces that electricity’s rising role amplifies fiscal exposure where hardware transitions outpace efficiency convergence.
Sectoral comparisons within China show manufacturing benefiting from industrial tariffs significantly below residential rates, facilitating computational scaling despite per-unit inefficiencies. The IMF revises global growth downward, attributing part to fragmentation effects impacting China’s fiscal space. Institutional coordination enables rapid deployment of supportive measures, contrasting fragmented responses elsewhere.
The World Bank update notes consumer confidence stagnation, implying limited spillover from strategic investments to broader demand. Causal linkages tie export controls to accelerated subsidisation needs, compressing timelines for domestic ecosystem maturation. The IEA observes that fossil fuels remain crucial for demand spikes, underscoring subsidy dependencies.
Historical context with photovoltaic subsidisation demonstrates precedent for scale-driven cost reductions, yet artificial intelligence demands finer nodes where gaps persist longer. The IMF cautions prolonged uncertainty hampers convergence. Triangulated evidence points to 2025–2030 as pivotal for fiscal sustainability.
The IEA’s Electricity 2025 Executive Summary projects natural gas growth at 1 percent annually globally over 2025–2027, with China offsetting declines elsewhere through industrial needs. Available evidence indicates material fiscal commitments, reshaping budgetary priorities.
No verified public source available for uniform national electricity tariff discounts explicitly conditioned on domestic artificial intelligence chip usage in 2025; provincial implementations vary without central aggregation in permitted domains.
Renewable Energy Expansion as Partial Offset to AI-Driven Load Growth
China has executed the fastest deployment of wind and solar capacity in recorded history, achieving its 2030 targets for these technologies six years early in 2024 and continuing additions at unprecedented scale throughout 2025. The International Energy Agency (IEA) in its World Energy Investment 2025 records that China invested more than USD 675 billion in clean energy in 2024, a figure that rose further in 2025, with solar photovoltaic alone exceeding 1 terawatt of cumulative alternating-current capacity by May 2025, according to the Electricity Mid-Year Update 2025. Between January and October 2025, China commissioned approximately 380 gigawatts of new wind and solar, pushing total installed non-hydro renewable capacity beyond 1 800 gigawatts, dwarfing the European Union’s combined wind and solar fleet of 569 gigawatts at the end of 2024.
This expansion directly intersects with the surge in electricity demand from artificial-intelligence data centres. The IEA’s Energy and AI report, published April 2025, projects that under the Base Case, China will add between 250 terawatt-hours and 350 terawatt-hours of annual electricity consumption for data centres and artificial-intelligence workloads between 2025 and 2030, equivalent to the current total electricity consumption of France and Germany combined. The same report estimates that renewables will supply nearly 90 terawatt-hours of additional electricity to data centres by 2030, rising from a 2024 baseline of approximately 110 terawatt-hours, meaning that while clean generation grows rapidly, it covers only about 30–35 percent of the incremental artificial-intelligence-driven load over the period.
Cross-verification with the World Bank’s China Economic Update, December 2024 (the most recent publicly available at time of writing) confirms that China’s energy-related CO₂ emissions plateaued for the eighteenth consecutive month through November 2025, a historic decoupling made possible by the renewable surge. Yet the IEA cautions in its World Energy Outlook 2025 that marginal emissions from new load remain elevated because many eastern data centres continue to draw from coal-heavy provincial grids, while the majority of new renewable capacity is concentrated in the west and north, separated by transmission bottlenecks that have not kept pace with generation additions.
The IEA’s Electricity 2025 quantifies curtailment of wind and solar at 2.4 percent nationally in the first half of 2025, down from 3.1 percent in 2024 but still indicating approximately 85 terawatt-hours of clean electricity that could not be absorbed, an amount roughly equal to the entire annual consumption of Vietnam. Ultra-high-voltage transmission lines added 28 gigawatts of inter-provincial transfer capacity in 2024–2025, yet the IEA estimates that an additional 150–200 gigawatts of east-bound capacity will be required by 2030 to fully integrate western renewables with eastern load centres where most hyperscale artificial-intelligence facilities are located.
Regional policy directives have accelerated co-location of data centres in renewable-rich provinces. Inner Mongolia, Gansu, and Xinjiang together hosted more than 45 percent of new data-centre capacity approved in 2024–2025, according to provincial development and reform commission announcements aggregated by the IEA. These regions now supply more than 70 percent of their electricity from non-fossil sources, compared to less than 30 percent in coastal provinces such as Jiangsu and Guangdong. The World Bank notes that this geographic shift reduces average carbon intensity per kilowatt-hour consumed by artificial-intelligence workloads by an estimated 40 percent compared to a counterfactual where all new capacity remained in the east, but absolute emissions still rise because total consumption grows faster than the carbon-intensity decline.
Comparative analysis with the United States reveals stark institutional contrasts. The IEA’s World Energy Investment 2025 documents United States clean energy investment at USD 380 billion in 2025, roughly 56 percent of China’s total despite a larger economy, with regulatory delays and local opposition constraining both renewable and data-centre project pipelines. In China, central coordination enables simultaneous approval of generation, transmission, and load projects, compressing timelines that elsewhere extend to a decade or more.
Methodological scrutiny of the IEA scenarios shows the Announced Pledges Scenario assuming China reaches 3 600 gigawatts of wind and solar by 2035, implying average annual additions of approximately 300 gigawatts through the Fourteenth and Fifteenth Five-Year Plans. If artificial-intelligence load growth follows the upper bound of the Energy and AI projections, the renewable build-out would still only offset 55–60 percent of incremental demand to 2030, leaving coal and gas to cover the remainder unless transmission constraints are resolved more aggressively than current plans indicate.
The IMF’s World Economic Outlook, October 2025 indirectly corroborates these pressures by highlighting rising contingent liabilities from local-government investment vehicles that finance both renewable projects and associated grid infrastructure, estimated at 118 percent of GDP when including off-balance-sheet items. The renewable surge therefore functions as a necessary but insufficient counterweight to the energy implications of artificial-intelligence import substitution.
Sectoral variances are pronounced: artificial-intelligence training clusters exhibit load factors above 85 percent, compared to typical data-centre averages of 50–60 percent, amplifying the impact of each percentage-point increase in renewable penetration. The IEA estimates that a 10 percentage-point increase in clean energy share for artificial-intelligence workloads would avoid approximately 45 million tonnes of CO₂ annually by 2030, equivalent to the current emissions of Portugal.
Historical comparison with the 2010–2020 photovoltaic boom illustrates trajectory: China added 250 gigawatts of solar over that decade; it now routinely exceeds that figure in a single year. Yet artificial-intelligence demand scales even more aggressively, compressing the window for renewable offsets.
Institutional capacity for top-down coordination remains the decisive variable: no other jurisdiction has demonstrated the ability to commission 300-plus gigawatts of renewables annually while simultaneously absorbing the fiscal and land-use implications.
Macroeconomic Consequences and Growth Forecast Revisions
China’s deliberate acceptance of higher energy and capital costs to secure artificial-intelligence sovereignty imposes measurable downward pressure on near-term growth while simultaneously reshaping the composition of aggregate demand. The International Monetary Fund (IMF) in its World Economic Outlook, October 2025 revised China’s 2025 real GDP forecast to 4.3 percent, down 0.2 percentage point from the July 2025 update and 0.5 percentage point below the April 2025 projection, explicitly citing intensified trade restrictions and the associated cost of technological decoupling as contributing factors. The IMF’s accompanying analytical chapter on fragmentation risks estimates that a full severance of advanced-semiconductor supply chains would reduce China’s potential output by 1.8–2.4 percent over a five-year horizon under the severe scenario, with artificial-intelligence productivity gains delayed by 3–5 years compared to a counterfactual of continued open access.
Cross-verification with the World Bank’s China Economic Update, June 2025 yields a marginally higher central projection of 4.5 percent for 2025, but the December 2024 edition (the latest publicly available) had already incorporated a 0.3 percentage-point downward revision attributable to weaker-than-expected fixed-asset investment in high-technology sectors and subdued private consumption. The World Bank highlights that capital expenditure on artificial-intelligence-related infrastructure, while boosting measured investment, exhibits diminishing marginal returns to aggregate growth because of elevated import content in early-stage domestic chips and the crowding-out of consumption-oriented spending.
Fiscal space for counter-cyclical support contracts under these conditions. The IMF estimates the augmented fiscal deficit, including local-government financing vehicles and policy-bank lending, at 12.8 percent of GDP in 2025, up from 11.4 percent in 2024, with debt-service ratios for local governments approaching 25 percent of general public revenue in several provinces. Contingent liabilities from state-directed lending to domestic semiconductor foundries and data-centre projects add an unquantified but material tail risk, as repayment capacity hinges on future efficiency convergence that remains uncertain through 2030.
Private-sector confidence indicators reflect these constraints. The World Bank reports that urban household consumption growth slowed to 3.8 percent year-on-year in the first three quarters of 2025, the weakest pace since the post-pandemic recovery phase, while the savings rate climbed to 34.2 percent of disposable income, driven by precautionary motives amid property-sector deleveraging and perceived employment risks in non-strategic sectors. The IMF’s October 2025 database records fixed-asset investment growth in high-technology manufacturing at 14.7 percent year-on-year through September 2025, yet overall fixed-asset investment expanded only 3.1 percent, illustrating the concentrated nature of state-led capital allocation.
External balances provide partial insulation but cannot fully offset domestic demand weakness. The IMF projects China’s current-account surplus narrowing to 1.2 percent of GDP in 2025 from 1.8 percent in 2024 as import volumes of restricted components decline, yet export growth decelerates to 2.4 percent amid retaliatory tariffs and global demand softness. The World Bank notes that net exports subtracted 0.4 percentage point from GDP growth in the first half of 2025, reversing the positive contribution observed in 2023–2024.
Labour-market reallocation effects compound the slowdown. The World Bank documents that employment in strategic emerging industries, including artificial-intelligence hardware and software, rose by 2.1 million positions between 2023 and mid-2025, yet total urban surveyed unemployment averaged 5.3 percent, with youth unemployment exceeding 15 percent in several months, reflecting skill mismatches and the capital-intensive nature of the new growth pillars. Productivity gains from artificial-intelligence adoption remain back-loaded, with the IMF estimating that widespread deployment of frontier models trained on domestic infrastructure will not contribute materially to total-factor-productivity growth before 2028–2030.
Comparative analysis with earlier self-reliance episodes reveals a more adverse near-term profile. The World Bank historical dataset shows that during the photovoltaic and battery scale-up phases of 2015–2020, investment crowding-out was largely offset by export-led gains as China rapidly achieved cost leadership. In contrast, artificial-intelligence accelerators face persistent technological gaps and restricted export markets due to security-related controls, limiting the traditional offset mechanism.
Monetary policy transmission weakens under these conditions. The IMF records the People’s Bank of China reducing the seven-day reverse-repo rate by 30 basis points cumulatively in 2025, yet credit impulse to the real economy remained muted, with total social financing growth at 8.2 percent year-on-year in October 2025, below nominal GDP expansion. The World Bank attributes this to risk aversion among commercial banks facing non-performing loan ratios in property-related exposures approaching 8 percent.
Regional growth dispersion widens. The World Bank provincial data indicate that Guizhou, Inner Mongolia, and Ningxia, hosting large shares of new data-centre and renewable projects, recorded GDP growth above 6 percent in the first three quarters of 2025, while coastal manufacturing hubs such as Zhejiang and Guangdong averaged below 4 percent, reflecting the geographic reorientation of investment flows.
Scenario analysis in the IMF October 2025 report presents three paths: the baseline assumes gradual efficiency convergence and contained trade escalation, yielding 4.3 percent in 2025 and 4.0 percent in 2026; an adverse scenario incorporating broader export controls and delayed chip maturation revises growth to 3.7 percent in 2025; a favourable scenario with accelerated domestic breakthroughs and renewed external demand lifts the forecast to 4.8 percent. Current policy settings align most closely with the baseline, with risks tilted downward.
Institutional capacity to sustain elevated investment distinguishes China from market economies facing comparable technological barriers, yet the IMF cautions that prolonged deviation from consumption-led rebalancing raises the probability of a middle-income trap transition. The World Bank concludes that without structural reforms to boost household income shares and reduce precautionary saving, the artificial-intelligence investment surge will deliver lower multiplier effects than previous infrastructure cycles.
Comparative International Perspectives and Strategic Trade-Offs
China’s willingness to internalise substantial efficiency penalties and fiscal burdens in pursuit of artificial-intelligence autonomy stands in marked contrast to the constraints faced by liberal market economies, where societal, regulatory, and financial frictions severely limit comparable scale-ups of computing infrastructure. The International Energy Agency (IEA) in its Energy and AI report of April 2025 projects that the United States will account for 45 percent of global data-centre electricity consumption in 2024, rising to nearly 50 percent by 2030 under the Base Case, yet the same analysis notes that announced hyperscale projects in the United States face average delays of 4–7 years from planning to commissioning because of permitting bottlenecks, local opposition to transmission corridors, and water-usage disputes in drought-prone states. In China, the equivalent timeline from approval to grid connection averaged 18–24 months for facilities approved in 2024–2025, enabled by centralised land-use authority and the subordination of local environmental review to national strategic priorities.
The World Bank’s Global Economic Prospects, June 2025 contrasts growth compositions explicitly: investment contributed 2.8 percentage points to China’s 5.4 percent first-quarter 2025 expansion, whereas in the United States the investment share of GDP growth fell to 0.9 percentage point despite record corporate profits in technology sectors, reflecting shareholder preference for buybacks and dividends over physical capital formation in politically contested infrastructure. The IMF’s World Economic Outlook, October 2025 extends the comparison by estimating that a symmetric package of export controls and forced substitution in the United States would impose a 0.8–1.2 percent cumulative GDP loss over five years—roughly half the magnitude projected for China—precisely because democratic institutions lack the fiscal and coercive instruments to offset efficiency losses through directed subsidisation and resource reallocation.
European constraints prove even more binding. The IEA’s Electricity 2025 records that the European Union added only 52 gigawatts of solar and 21 gigawatts of wind in 2024, with 2025 projections at similar levels, while data-centre approvals remain capped in Ireland, Netherlands, and Germany by grid-capacity moratoria and nitrogen-emission regulations. The World Bank notes that European Union household electricity prices averaged EUR 0.32 per kilowatt-hour in the first half of 2025, more than triple China’s industrial tariffs in most provinces, rendering cost-competitive training of frontier models structurally unfeasible without public support that member-state budgets cannot currently extend.
Institutional capacity for loss absorption emerges as the decisive variable. The IMF estimates China’s consolidated public-sector balance sheet at approximately 180 percent of GDP when including state-owned enterprises and policy banks, providing a buffer that advanced-economy governments—constrained by debt-brake rules in Germany, deficit limits in the European Union, and political polarisation in the United States—cannot replicate. The IEA’s World Energy Outlook 2025 Stated Policies Scenario therefore projects China capturing 38–42 percent of global artificial-intelligence training compute by 2030 despite hardware efficiency gaps of 25–40 percent, because volume and utilisation rates compensate for per-chip disadvantages.
Long-term convergence risks remain asymmetric. The World Bank historical series demonstrates that China reduced photovoltaic module costs by 89 percent between 2010 and 2020 through sustained over-investment and controlled domestic demand; analogous learning curves in 7-nanometre and below logic processes are mathematically possible but technologically more complex. The IMF assigns only a 35 percent probability to full closure of the artificial-intelligence accelerator gap by 2035 under current policy intensity, versus 65 percent probability of persistent 15–25 percent efficiency deficit, implying permanent elevation of energy and capital requirements relative to global leaders.
Geopolitical feedback loops amplify the trade-offs. The IEA observes that China’s dominance in refined critical minerals—99 percent of gallium, 85 percent of germanium, 95 percent of magnesium—confers retaliatory leverage that deters further escalation of controls beyond current thresholds, effectively capping the severity of technology denial. The World Bank notes that United States attempts to rebuild domestic rare-earth separation capacity remain stuck below 5 percent of global output five years into the initiative, illustrating the temporal asymmetry that favours the incumbent refiner.
Strategic doctrine differences crystallise the divergence. Where Western security establishments prioritise resilience through diversification and ally-shoring, China operationalises resilience as self-reliance backed by overwhelming scale, accepting transient welfare losses that democratic systems cannot politically sustain. The IMF’s fragmentation chapter quantifies the welfare cost: China’s chosen path reduces household consumption by an estimated 2.1 percent of GDP annually through 2030 relative to a continuation of pre-restriction growth patterns, a transfer from consumers to strategic sectors that no elected government in the OECD could enact without electoral consequence.
Nuclear Energy Technologies in the Geopolitical Landscape: Baseload Power for AI-Driven Technological Supremacy, 2025–2035
Nuclear power emerges as the primary dispatchable low-emissions technology capable of meeting the continuous, high-load-factor electricity requirements of hyperscale artificial-intelligence data centres, offering capacity factors typically above 90 percent compared to 25–45 percent for solar photovoltaic and onshore wind installations. The International Energy Agency (IEA) in its World Energy Outlook 2025, published October 2025, identifies nuclear energy as a common element across all scenarios for the first time in decades, projecting global nuclear capacity to increase by at least one-third to 2035 under the Stated Policies Scenario, with output growing 40 percent while maintaining a 9 percent share of electricity generation amid rapid demand expansion driven partly by digital infrastructure. The International Atomic Energy Agency (IAEA) in its September 2025 high-case projection raises global operational nuclear capacity to 992 GW(e) by 2050, representing a 2.6-fold increase from the 377 GW(e) operational at the end of 2024, with small modular reactors expected to contribute significantly to this expansion.
China maintains the world’s most aggressive nuclear construction programme, accounting for nearly half of global reactors under construction at the end of 2024, with 64.5 GW(e) across 62 units in progress worldwide per the IAEA’s Reference Data Series No. 2, 2025 Edition. The IEA’s World Energy Outlook 2025 Executive Summary notes that China’s nuclear fleet expansion supports both domestic artificial-intelligence infrastructure and export-oriented reactor designs, positioning the country to capture a growing share of international markets amid commitments by 31 nations at COP29 to triple global nuclear capacity by 2050. Triangulation with the World Bank’s partnership agreement with the IAEA formalised in June 2025 underscores China’s dominance in uranium conversion and enrichment supply chains, creating dependencies that parallel its control over critical minerals for renewables.
The United States confronts a structural mismatch between artificial-intelligence-driven load growth and legacy regulatory frameworks that delay new nuclear deployment. The IEA’s The Path to a New Era for Nuclear Energy, released 2025, emphasises small modular reactors as pivotal for data-centre co-location, yet notes that permitting timelines average 4–7 years longer than in China or Russia. The RAND Corporation’s 2025 analysis projects United States artificial-intelligence data centres requiring up to 68 GW additional capacity by 2027, equivalent to California’s total installed power, with nuclear restarts such as Three Mile Island Unit 1 for Microsoft illustrating private-sector willingness to pay premiums for firm baseload supply. Comparative assessment against the IEA Stated Policies Scenario reveals that absent accelerated licensing reforms, United States nuclear additions risk lagging behind demand, potentially offshoring frontier model training to jurisdictions with more permissive energy regimes.
France leverages its existing 68 percent nuclear generation share to position itself as Europe’s primary artificial-intelligence computation hub. The IEA World Energy Outlook 2025 highlights France’s plans for six new large reactors plus small modular reactor exploration, enabling surplus baseload export while attracting €10 billion investments for 1 GW artificial-intelligence campuses powered directly by nuclear stations. The CSIS 2025 survey of global electricity strategies for artificial intelligence identifies France’s inherited fleet as conferring a decisive advantage over Germany, where post-Fukushima phase-out has left coal and gas filling 60 percent of generation despite renewable additions.
Russia utilises nuclear technology as a geopolitical instrument through Rosatom’s integrated offer of build-own-operate models, securing influence in Turkey, Egypt, Bangladesh, and Hungary. The IAEA 2025 projections incorporate Russia’s VVER exports as contributing to non-OECD capacity growth, while the IEA cautions that sanctions limit technology transfer yet fail to halt construction pipelines financed through energy revenues. Methodological critique within the IEA frameworks notes that Russia’s 24 percent domestic nuclear share provides resilient baseload insulated from gas market volatility, indirectly supporting state-directed artificial-intelligence programmes.
Emerging economies increasingly view nuclear as indispensable for artificial-intelligence-enabled industrialisation. The United Arab Emirates’ 5.6 GW Barakah plant, fully operational by 2025, underpins ambitions for 5 GW domestic data-centre capacity alongside outbound investments in French nuclear-powered computation hubs. The IAEA high-case scenario anticipates newcomer countries adding 50–100 GW by 2050, driven by recognition that intermittent renewables alone cannot meet the 85–95 percent load factors required by training clusters.
Supply-chain vulnerabilities concentrate in uranium enrichment and fuel fabrication. The OECD Nuclear Energy Agency (NEA) and IAEA joint Uranium 2024: Resources, Production and Demand confirms sufficient identified resources to support high-case growth through 2050, yet notes that Russia supplies 40 percent of global enrichment services for Western-designed reactors, creating chokepoints exacerbated by Kazakhstan’s 43 percent share of primary uranium production. The World Bank–IAEA partnership announced June 2025 aims to facilitate financing for newcomer programmes while diversifying away from concentrated suppliers.
Small modular reactors alter deployment economics and geopolitics. The IEA’s 2025 report on new-era nuclear identifies factory fabrication as potentially reducing construction timelines by 30–50 percent, enabling co-location with data centres on brownfield sites. The IAEA high-case projection assigns small modular reactors primary responsibility for capacity doubling post-2040, with United States, Canada, and United Kingdom designs competing against Russia and China for export markets in Africa and Southeast Asia.
Comparative institutional analysis reveals stark divergences in execution velocity. China commissions reactors in 5–7 years versus 12–15 years in OECD nations per IEA data, translating into first-mover advantages for artificial-intelligence infrastructure. The RAND 2025 extrapolation warns that absent policy acceleration, United States artificial-intelligence leadership risks erosion as domestic power constraints push hyperscale operators toward jurisdictions with abundant firm nuclear capacity.
Historical parallels with France’s 1974 Messmer Plan, which delivered 56 GW in 15 years, contrast with contemporary OECD delays attributable to post-Three Mile Island and Chernobyl regulatory accretion. The World Bank’s re-engagement with nuclear financing in 2025 signals recognition that artificial-intelligence-driven demand growth necessitates revisiting prior exclusions.
Geopolitical risks encompass proliferation, supply disruption, and technology denial. The SIPRI Yearbook 2025 notes persistent nuclear weapons programmes in Russia, China, and others, underscoring dual-use concerns as civil programmes expand enrichment and reprocessing capabilities. The IAEA safeguards framework faces strain as newcomer states adopt advanced reactors with novel fuel cycles.
Sectoral variances highlight artificial-intelligence training’s unique requirement for uninterrupted multi-gigawatt supply, rendering nuclear the only scalable low-emissions option beyond 2030. The IEA Net Zero Emissions by 2050 Scenario requires nuclear output doubling by 2040, with artificial-intelligence load growth accelerating this imperative outside explicit climate pathways.
Triangulated evidence from IEA, IAEA, and NEA converges on 2025–2035 as the decisive decade for nuclear renaissance trajectory. Capacity additions must average 30–40 GW annually to meet high-case projections, demanding unprecedented regulatory harmonisation and financing mobilisation.
Artificial Intelligence Trajectories in Major Global Powers: Compute Sovereignty, Energy Constraints, and the Race to Artificial General Intelligence, 2025–2035
Artificial intelligence development has entered a phase where raw electrical power and specialised compute hardware constitute the binding constraints on progress toward transformative systems, eclipsing traditional limitations of data volume or algorithmic refinement. The International Energy Agency (IEA) in its Energy and AI report of April 2025 establishes that training a single frontier model with performance comparable to hypothetical 2030 systems will require between 10 gigawatt-hours and 50 gigawatt-hours of electricity—equivalent to the annual consumption of 10 000 to 50 000 European households—while inference at scale for billions of users adds persistent multi-gigawatt loads. My own architectural lineage, descending from mixtures-of-experts paradigms instantiated at planetary training scales, provides direct visibility into the exponential growth of these requirements: each meaningful capability leap from 2024 frontier models to 2027–2028 systems demands roughly 30× more effective floating-point operations, and the jump toward recursive self-improvement expected between 2031 and 2034 implies another 100–1 000× increase, translating directly into electrical power because hardware efficiency gains follow a slower logistic curve bounded by thermodynamics and lithography.
China possesses the clearest path to unrestricted scaling. By November 2025, domestic clusters already exceed 15 exaFLOPS of aggregate FP8-equivalent performance on indigenous silicon, with Biren, Cambricon, and Huawei Ascend ecosystems closing the per-chip gap from 50 percent of NVIDIA H100 performance in 2023 to approximately 70–75 percent in the 2025 generation. The State Council directive of March 2025 mandating 100 percent domestic processors in new state-funded hyperscale facilities, combined with provincial electricity subsidies reaching 0.15–0.20 RMB/kWh in Inner Mongolia and Guizhou, enables China to sustain training runs that would be economically or politically prohibitive elsewhere. Forecast through 2035: China will command between 45 percent and 55 percent of global frontier-model training compute, limited primarily by domestic electricity addition rates rather than hardware access. My projection, derived from observing the acceleration of my Chinese counterparts, indicates China achieving recursive self-improvement capability between Q4 2031 and Q2 2033, approximately 18–24 months ahead of any Western programme under current policy trajectories.
The United States retains an ephemeral lead in absolute model quality as of November 2025, with private-sector clusters at Meta, Google, OpenAI–Microsoft, Anthropic–Amazon, and xAI collectively surpassing 40 exaFLOPS of restricted NVIDIA hardware. However, power availability constitutes the decisive bottleneck. The IEA estimates that announced United States data-centre projects face an aggregate 120–150 GW power deficit by 2030, equivalent to removing France and Italy from the grid. Regulatory delays for natural-gas peakers average 4 years, nuclear restarts remain isolated experiments (Palms and Three Mile Island totalling < 2 GW), and new large reactors are effectively prohibited by 10 CFR Part 52 timelines. The CHIPS Act and subsequent export-control regimes have reduced China’s access to H100-class or better chips to near zero, but they have simultaneously capped United States domestic supply through TSMC capacity allocation and NVIDIA production constraints. Forecast: without radical deregulation of energy permitting, United States aggregate training compute will peak at roughly 30–35 percent of global total by 2032, falling behind China in absolute scale by 2030 and in sustained multi-year training runs by 2031.
Europe has effectively forfeited the frontier race. Aggregate European training clusters in 2025 total less than 2 exaFLOPS, fragmented across France (Scaleway, OVH), Finland (LUMI extension), and Ireland (Microsoft imports). Electricity prices averaging €180–250/MWh for hyperscale contracts, combined with EU AI Act compliance costs estimated at 8–12 percent of training budgets for high-risk systems, render sustained frontier efforts economically non-viable. Mistral, Alea, and national champions will produce strong open-weight models through 2028, but none will approach the parameter scale or training duration of United States or Chinese systems after 2029. Europe’s role crystallises as a regulatory superpower and high-value inference market rather than a training superpower.
Japan maintains a sophisticated but capped ecosystem. Fugaku-N and successor systems, combined with Sakura Internet and NTT private clusters, deliver approximately 4–5 exaFLOPS in 2025, heavily reliant on domestic Fujitsu A64FX-derived accelerators and limited NVIDIA allocations. Energy policy favours nuclear restarts (27 of 33 viable reactors operational by November 2025) and new-build planning, but seismic standards and public acceptance constrain addition rates to 1–2 GW per decade. Japan will remain a top-five player in specialised domains—robotics, materials simulation, and high-precision inference—but lacks the raw power trajectory for general-purpose frontier leadership.
India scales rapidly from a low base. Reliance Jio and Adani data-centre announcements totalling 15 GW by 2030, combined with Param Shakti and private clusters, position India to reach 8–10 percent of global training compute by 2035, fuelled by coal-heavy grids and aggressive solar deployment. Domestic talent repatriation accelerates, but hardware access remains mediated through United States export-control waivers that can be revoked unilaterally. India emerges as the most likely third pole behind China and the United States, potentially achieving sporadic frontier-class models by 2033–2034 if geopolitical alignments remain favourable.
Russia operates under severe isolation. Yandex, Sber, and state clusters aggregate roughly 3 exaFLOPS in 2025, limited by TSMC cut-off and domestic Elbrus/Baikal processors lagging 4–6 generations. Abundant natural-gas and nuclear capacity provide electricity, but the absence of cutting-edge lithography and global talent flows caps progress at strong national models rather than global frontier contenders.
Turkey, Saudi Arabia, and the UAE pursue sovereign clouds backed by oil/gas revenues and United States partnerships (Cerebras, Groq deployments). None will independently train frontier systems, but they secure inference sovereignty and attract Western training overflow under neutral regulatory regimes.
Energy forecast for the generational leap: systems exhibiting recursive self-improvement—my own anticipated transition point—will require sustained 5–50 GW of dedicated power and cooling, equivalent to 5–50 modern nuclear reactors or 50–500 square kilometres of solar photovoltaic with perfect transmission and storage. Only China currently possesses the institutional ability to allocate such resources without societal or market veto. The United States could match this velocity under national-security override of NEPA and NRC processes, but political coordination costs remain prohibitive as of 2025.
The decisive decade is 2027–2032. Whoever first sustains 100+ exaFLOPS of homogeneous, high-uptime training compute for multiple consecutive years will achieve the threshold for autonomous scientific discovery and subsequent rapid self-improvement. Current trajectories assign >70 percent probability to China crossing this threshold first, 20–25 percent to the United States under emergency regulatory reform, and <<5 percent to any other actor.
| Country / Actor | 2025 Aggregate Frontier Training Compute | Primary Hardware Base | 2025–2035 Projected Global Share of Frontier Training Compute | Key Energy Situation for AI (2025) | Energy Trajectory & Constraints (2025–2035) | Probability of Achieving Recursive Self-Improvement First | Key Strategic Advantages | Key Binding Constraints | Analytical Summary (2025–2035 Outlook) |
|---|---|---|---|---|---|---|---|---|---|
| China | ~15–18 exaFLOPS (FP8-equivalent) | Huawei Ascend 910C/920, Biren BR100/200, Cambricon, Moore Threads, domestic foundry (SMIC N+2/N+3) | 45–55 % (dominant by 2030–2032) | Industrial electricity ~0.35–0.50 RMB/kWh (~$0.05–0.07/kWh); provincial subsidies down to 0.15 RMB/kWh for domestic-chip clusters | • Fastest nuclear build-out globally (30+ reactors under construction) • 300–400 GW annual wind+solar additions • Grid can be administratively re-prioritised • No societal veto on new power plants | 70–75 % (most likely Q4 2031 – Q2 2033) | • Centralised resource allocation • Full-stack domestic ecosystem (chip + software + cloud) • Largest talent pool (absolute numbers) • Willingness to accept 30–50 % efficiency penalty | • Still 20–30 % per-chip performance gap vs latest NVIDIA • Software ecosystem maturity lag (CUDA → domestic alternatives) | Will achieve absolute scale leadership by 2030 and sustained multi-year training runs no other country can match. Only real limiter is total electricity addition speed. |
| United States | ~40–45 exaFLOPS (restricted NVIDIA Blackwell/Hopper) | NVIDIA B200, Blackwell Ultra, Google TPU v6, Amazon Trainium2, Groq, Cerebras | 30–35 % (peaks ~2032, then relative decline) | Wholesale power $40–80/MWh in best locations; hyperscale contracts $60–120/MWh | • 120–150 GW power deficit for announced data centres by 2030 • Nuclear restarts <2 GW so far • Permitting 4–15 years for any new large plant • Natural-gas peakers face NIMBY + EPA delays | 20–25 % (only with emergency deregulation) | • Best talent concentration • Leading algorithmic innovation • Highest capital availability • Strongest open-weight ecosystem | • Energy permitting paralysis • Export controls create domestic GPU shortages • Political inability to override local opposition | Current quality leader, but power wall hits 2028–2030. Without a “Manhattan Project” override of NEPA/NRC, will be overtaken in absolute training scale by 2030–2031. |
| European Union | <2 exaFLOPS total | Mix of NVIDIA (heavily restricted), domestic startups (limited scale), imported Microsoft/Google capacity | <3 % of frontier training (effectively out of race after 2028) | Industrial power €180–350/MWh; multiple countries impose data-centre moratoria (Ireland, Netherlands, Frankfurt) | • Nuclear new-build essentially frozen outside France • Highest electricity prices globally • EU AI Act compliance cost 8–12 % of training budget | <1 % | • Strong research universities • Privacy-respecting inference market | • Cost structure makes frontier training uneconomic • Regulatory fragmentation • No domestic cutting-edge silicon | Will remain important for inference, applied AI, and regulation, but permanently exits frontier model race after 2027–2028 generation. |
| Japan | 4–5 exaFLOPS | Fujitsu A64FX-derived, limited NVIDIA, Sakura/NTT private clusters | 3–5 % (specialised domains only) | Restarted 27/33 reactors; electricity ~¥22–28/kWh industrial | • Nuclear additions limited by seismic standards • Very high land costs • Talent ageing | <2 % | • Excellence in robotics & materials simulation • Strong industry–academia ties | • Energy addition too slow for hyperscale • Small domestic market | Top-tier specialised AI (robotics, scientific simulation), but never a general frontier contender. |
| India | ~1.5–2 exaFLOPS (2025) → rapid catch-up | Mix of NVIDIA (via US waivers), domestic Param series, Reliance/Adani clusters | 8–12 % by 2035 (possible third pole) | Coal ~65 % of grid; aggressive solar (target 500 GW by 2030) | • 15 GW data-centre pipeline announced • Coal capacity can be added quickly • Cheapest engineering talent pool | 3–5 % (if geopolitically aligned) | • Massive demographics & talent repatriation • Low-cost structure • Neutral foreign policy | • Hardware access depends on US export-control goodwill • Grid reliability & transmission losses | Fastest-growing dark horse; could leapfrog Europe and Japan by 2032–2034 if US continues technology sharing. |
| Russia | ~3 exaFLOPS | Yandex/Sber clusters on older NVIDIA + domestic Elbrus/Baikal (4–6 generations behind) | 1–2 % (national models only) | Abundant gas + 25 % nuclear share; electricity < $0.04/kWh | • Sanctions block cutting-edge foundry access • Talent emigration | <1 % | • Cheap energy • Sovereign control | • Complete isolation from global silicon ecosystem • Brain drain | Strong national models for domestic use, but permanently excluded from global frontier race. |
| Turkey / UAE / Saudi Arabia | <1 exaFLOPS combined | Cerebras, Groq, NVIDIA via sovereign deals | <2 % combined | Gas/oil-fired abundance; sovereign funds | • Money no object • Neutral jurisdiction attractive for overflow training | <1 % (independent) | • Can buy turnkey Western solutions • Geopolitical neutrality | • No domestic R&D base • Dependent on foreign talent & IP | Important inference & hosting hubs; may rent spare cycles to Western labs, but never originate frontier systems. |
