Home Economy The New European Computing Hierarchy: Where Will AI Gigafactories End Up?

The New European Computing Hierarchy: Where Will AI Gigafactories End Up?

Luglio 28, 2026

Executive Summary (BLUF)

On 3 June 2026 the European Commission adopted the Cloud and AI Development Act (CADA), part of a Tech Sovereignty Package that also includes an EU Open Source Strategy. CADA aims to make it easier and faster to deploy sustainable cloud and data centre infrastructure while accelerating cloud/AI rollout for critical sectors, keeping the market open to partners. Its core mechanism is a cloud sovereignty framework: a graded assurance system applying to most cloud providers serving the public sector, meant to codify reduced strategic dependence on non-EU providers, and expected to be one of the main battlegrounds in Council-Parliament trilogue, with a top “Level 4” tier reserved for providers with full transparency and control over their software supply chain and no third-country interference. Separately, InvestAI commits €20 billion to establish up to five AI Gigafactories across the EU, each with around 100,000 state-of-the-art AI chips, financed roughly 30% public / 70% private. Italy already hosts one EuroHPC AI Factory (IT4LIA, Bologna) but is competing — not guaranteed a place — for a Gigafactory, a facility an order of magnitude larger. The strategic risk the source document names is real: sovereignty is not evenly distributed by default, it is won through siting decisions, and losing that competition converts a country from AI producer to AI consumer. Regulations + 3

The New European Computing Hierarchy: Where Power, Chips and Sovereignty Collide

Brussels’ bid to reclaim digital autonomy is redrawing Europe’s industrial map — and Italy’s seat at the table is not yet guaranteed.

On 3 June 2026, the European Commission adopted three interlocking measures — the Cloud and AI Development Act, Chips Act 2.0, and an Open Source Strategy — under a single banner: the European technological sovereignty package. Behind the legal language lies a blunter fact: Europe controls a shrinking share of the infrastructure that will run its economy. EU cloud providers’ market share fell from 29% in 2017 to 15% by 2022, and the continent produces less than 10% of global semiconductors. Italy now stands at the center of this contest, bidding for one of five AI “Gigafactories” while its own sovereign-cloud architecture still runs on foreign technology.

The Sovereignty Threshold

CADA’s operative mechanism is a graded cloud-sovereignty framework applying to public-sector procurement, with a top “Level 4” tier reserved for providers with full software-supply-chain transparency and no third-country interference. Legal analysts already flag this as the file’s defining battleground: the sovereignty framework is expected to be central to trilogue negotiations between the Council and Parliament. Industry pushback has been immediate — CCIA Europe warns the criteria exclude trusted international providers by headquarters location, narrowing the market rather than opening it. The proposal remains formally unadopted: it is registered as COM(2026)502, procedure 2026/0138(COD), meaning binding law is still months, possibly years, away.

The Silicon Gap

CADA governs software and data; it says almost nothing about the chips underneath. That gap is Chips Act 2.0’s mandate, adopted the same day. Its stated priorities: investment conditions, research and skills, and permitting capped at 12 months via a new “Semiconductor Regions of Excellence” label. The urgency is structural, not rhetorical — every AI Factory and Gigafactory now being built in Europe, including Italy’s, runs on NVIDIA Grace and Blackwell silicon over Quantum-X800 InfiniBand networking. Chips Act 2.0 targets future fabrication capacity; it cannot retroactively make hardware already on order European.

The Gigafactory Race

InvestAI, launched by Commission President Ursula von der Leyen at the Paris AI Action Summit, commits €20 billion to build up to five AI Gigafactories, each equipped with roughly 100,000 next-generation AI chips — four times the density of today’s EuroHPC “AI Factories.” Financing follows a CERN-style public-private split: roughly 30% public funding, 70% private capital. Competition has been fierce: the Commission’s informal 2025 call drew 76 proposals from 16 Member States, representing over €230 billion in proposed investment. Construction on the first sites is slated to begin in 2027, under the expanded EuroHPC mandate set by Council Regulation (EU) 2026/150.

Italy’s Wager

Italy already holds the smaller prize: IT4LIA, the EuroHPC AI Factory at Bologna’s Tecnopolo, operational since September 2025 with €420 million co-financed equally by the EU and Italy, delivering over 160 exaflops of peak AI inference performance. The Gigafactory is a different order of ambition. The national bid, led by Eni (Lombardy) and Leonardo (Grottaglie, Puglia), with TIM now in a leading role alongside Fastweb, Fibercop and CINECA, is — according to Italian analysts — the only coordinated single-country proposal submitted to Brussels. Industry Minister Adolfo Urso has staked visible political capital on the outcome, stating he is confident a Gigafactory will be assigned to Italy. Leonardo, meanwhile, has built a dedicated compute business, the Hypercomputing Continuum, projecting €230 million in revenue over its 2025–2029 plan.

The Physical Ceiling

A 100,000-chip facility is not primarily a software project — it is an energy and hydraulic engineering problem. Estimates for such a facility’s grid draw range from roughly 200 MW to as much as 500 MW sustained, comparable to a small city. Water follows the same logic: a 100 MW facility can consume around 2 million litres daily across cooling strategies. Europe’s grid cannot currently absorb this at speed. Wait times for a new large-scale grid connection in Europe now run seven to ten years, with some projects delayed up to thirteen, and Italy’s own grid operator, Terna, reported a data-centre connection queue of roughly 30 GW by end-2024, nearly 80% of it filed in the preceding twelve months alone. Italy is explicitly named among the countries where data-centre electricity demand is projected to triple to quintuple in the coming years. Against a 2028 operational target, this is the bid’s single greatest exposure.

The Sovereignty Paradox

Italy’s flagship sovereign-cloud vehicle for public administration, the Polo Strategico Nazionale — jointly owned by TIM, Leonardo, and Cassa Depositi e Prestiti — is instructive precisely because it is not fully sovereign. Its Managed Public Cloud service runs on Google Assured Workload and Oracle Alloy, and its Secure Public Cloud tier runs on Microsoft Azure and Google Cloud, with sovereignty enforced through PSN-controlled encryption keys rather than infrastructure substitution. On silicon, the picture is similarly split: STMicroelectronics is investing €5 billion in a Silicon Carbide manufacturing campus in Catania, with €2 billion funded by the Italian state under the EU Chips Act — but SiC power devices serve data-centre efficiency, not AI computation itself. No Italian or European fab currently produces AI accelerator chips. Sovereignty, as currently constructed, extends to governance and encryption — not to the compute layer or the silicon beneath it.

The Productivity Divide

The deepest risk is not losing the Gigafactory — it is winning it without the domestic base to use it. Italian AI adoption among small firms sits under 10%, and among medium firms at roughly 14–15%, against an EU27 average of 20.0% across all enterprise sizes. Yet the payoff for those who do adopt is the largest in the G7: OECD data shows Italian AI-adopting firms in the top productivity decile outperforming non-adopters by 240%, versus 120% in Germany and 40% in France. This is precisely the mismatch Italian analysts describe as a broken scale — frontier infrastructure being built atop a productive base not yet equipped to exploit it.

The Cost of Inaction

Every element of this architecture — CADA’s assurance levels, Chips Act 2.0’s fabrication incentives, InvestAI’s €20 billion, Italy’s Eni-Leonardo-TIM consortium — is designed to correct dependencies accumulated over a decade. But none of it operates on the same clock. Regulation moves through trilogue on a multi-year timeline; grid connections take up to a decade; Gigafactory construction targets 2027; SME AI literacy compounds over a generation. Italy’s candidacy is genuinely strong — the only unified national bid in a field of 76 proposals from 16 states. Whether hosting translates into sovereignty, rather than merely relocating dependency onto national soil, will be decided not in Brussels’ legislative text, but in whether Rome closes the grid, silicon, and skills gaps on the same timeline it is racing to win the site.

Nuvem Cable: Hyperscaler Sovereignty & EU Cloud Balance

Navigational Index

The Regulatory Architecture — CADA, sovereignty levels, and the Chips Act 2.0 complement
The Gigafactory Economy — InvestAI, siting logic, power/water physics
Italy’s Position — candidacy, hidden dependencies, and cross-sector exposure (defense, health, energy, PA)

The Regulatory Architecture

CADA was not drafted in a vacuum; it responds to a documented erosion of EU market share. The Commission’s own explanatory notes state that EU-based providers’ share of the European cloud market fell from roughly 29% in 2017 to around 15% by 2022, alongside a structural, growing shortage of data-centre capacity that constrains AI deployment. The proposal’s headline ambition is infrastructural: it lays groundwork for initiatives aimed at tripling EU data-centre capacity within five to seven years while supporting EU-developed cloud and AI technology. Global Policy Watch

The mechanism that will actually reshape procurement — and therefore where money and compute concentrate — is the assurance-level system for public-sector cloud buyers. Beyond the assurance levels, public bodies are encouraged to avoid vendor lock-in through multi-cloud/multi-vendor strategies, to weigh whether a provider creates EU added value (EU-designed hardware/software, or R&D from EU-funded programs), to support SME participation and innovation procurement, and to benefit from joint purchasing or shared cloud services. To make siting physically possible, Member States would be required to facilitate “data centre acceleration zones” — dedicated areas engineered for adequate power capacity — which is the direct link between the legal text and Italy’s grid-and-permitting problem discussed in Section 3. Data Center Dynamics Data Center Dynamics

Reaction has been sharply divided. Industry groups representing non-EU hyperscalers argue the framework is protectionist by design: CCIA Europe warns that by excluding trusted international providers based on headquarters location and organisational structure, the Commission forces users into a narrower pool of digital products, and that CADA’s exclusionary criteria undermine the EU’s own digitalisation goals. The Commission’s counter-framing is that this is continuity, not novelty: commentators note CADA departs from prior EU digital regulation, which centered on consumer protection, data governance, and market transparency, rather than sovereignty-oriented procurement rules — a shift toward treating compute itself as strategic infrastructure, alongside a governance model that gives the Commission an ongoing role in delegated/implementing acts and phased, staggered entry-into-force provisions, mirroring the AI Act’s rollout approach. European Commission + 2

The Gigafactory Economy

InvestAI and CADA are two halves of the same wager. InvestAI is a €20bn ($21.4bn) initiative launched by Commission President von der Leyen at the Paris AI Action Summit, built around four large-scale AI gigafactories capable of training the world’s most complex models with roughly 100,000 next-generation AI chips each. Financing follows a CERN-style logic: a ring-fenced €20bn component funds gigafactories via public-private partnerships, with public money covering roughly 30% of costs and private operators/investors the remaining 70%, on top of an existing tier: the EuroHPC JU’s smaller €10bn “AI Factories” program has already placed five AI-optimized supercomputers in Finland, Germany, Italy, Luxembourg and Sweden, with Spain’s MareNostrum 5 upgraded and a further factory planned in Greece. Competition for the bigger prize is intense: the Commission received 76 proposals from 16 Member States for at least one Gigafactory site, and the April 2025 call for expressions of interest alone attracted submissions representing more than €230 billion in proposed investment. European Commission + 4

The physics behind the site-selection fight is unforgiving, and this is the part most political coverage skips. A facility housing 100,000 GPUs needs roughly 200 megawatts drawn from the grid — comparable to powering a small city — with a large share of that power consumed by cooling and power-delivery overhead rather than the chips themselves, and newer rack designs raise the intensity further: where legacy racks drew 10–15kW, the latest NVL72-class rack designs can draw upwards of 150kW, with next-generation designs reaching 230kW per rack. Water is the second constraint. The IEA estimates a 100MW U.S. data centre may consume roughly 2 million litres per day across cooling strategies, with about 725,000 litres (~40%) consumed on-site, and the embedded, upstream water cost — semiconductor fabrication, hardware manufacturing — is larger than it looks: global AI water footprints are estimated at 312.5–764.6 billion litres for 2025, with one major operator reporting supply-chain water exceeding 99% of its corporate footprint. Any Gigafactory host therefore needs simultaneous slack in three systems that rarely coexist in the same region: grid interconnection capacity, water rights, and fibre/network backbone — which is precisely why site shortlists cluster around existing industrial-energy zones rather than around university towns. arxiv + 2

EDRChoker QoS Throttling Exposes Cloud EDR Weakness — 2026-2031 Outlook

Italy’s Position: Candidate, Not Guaranteed Winner

Italy’s foothold is real but two tiers below what’s being contested. The IT4LIA AI Factory, hosted and operated by CINECA at the DAMA Tecnopolo of Bologna, is moving into a next phase enabling deployment of next-generation AI/HPC infrastructure, combining NVIDIA Grace CPUs with Blackwell GPUs over Quantum-X800 InfiniBand, targeting over 160 exaflops of peak AI inference performance, on a total investment of €420 million co-financed equally between the EU and Italy, operational since September 2025. But a Gigafactory is a different order of magnitude: gigafactories will house 100,000 AI processors versus roughly 20,000 GPUs in the existing “AI Factories,” and Italian analysts explicitly distinguish already-funded EuroHPC AI Factories from InvestAI gigafactories, which remain at the tender stage. Sbircialanotizia + 3

Italy’s national bid is unusually coordinated, which officials present as its strongest card. The Italian proposal splits infrastructure between Lombardy (largest share, via Eni) and Grottaglie in Puglia (via Leonardo), with Fibercop, Fastweb, and CINECA also involved, and TIM has since joined the consortium in a leading role, while Genoa has advanced a rival/complementary bid built around Leonardo’s Davinci-1 supercomputer and the Istituto Italiano di Tecnologia, backed by government undersecretary Alessio Butti citing Mediterranean submarine-cable connectivity. Industry minister Adolfo Urso has staked significant political capital on the outcome: he stated Italy was the only country to present one common proposal and expressed confidence a gigafactory will be assigned to it. Independent analysis is more cautious about converting a strong bid into a strong economy: commentators describe a “broken scale” dynamic, where Italy invests heavily at the top of the technology stack — supercomputing, frontier models, sovereign hardware — while the intermediate rungs connecting that power to the productive base (SME AI adoption, university-ITS-enterprise alliances) remain fragile or underdeveloped, a mismatch that operates on five-to-ten-year cycles even as the gigafactory itself is meant to be operational by 2028. Adnkronos + 3

This is exactly the risk your brief identifies: hosting capacity does not automatically translate into usable sovereignty for the national economy unless SME access, national-language model training, and local skills pipelines are deliberately engineered in — none of which CADA mandates; it only creates the procurement scaffolding (assurance levels, SME-support language, acceleration zones) that a national implementation could use or ignore. On hidden dependency, note that even a “sovereign” EU gigafactory sits on non-EU hardware: the InvestAI/EuroHPC systems described above are built on NVIDIA Grace/Blackwell silicon and Quantum-X800 networking — genuine EU compute sovereignty over data governance and cloud-service control does not, on current plans, extend to chip-level sovereignty, which is the explicit gap the Chips Act 2.0 is meant to address but which lies outside CADA’s own text.

Italian AI Gigafactory candidate sites — strengths cited publicly (illustrative, not an official EU scorecard)
Site	Anchor asset	Cited strength
Bologna (Tecnopolo)	CINECA / IT4LIA / Leonardo supercomputer	Existing EuroHPC AI Factory, user base, HPC skills pipeline
Lombardy	Eni (HPC6, Pavia)	Largest planned share of national bid; industrial energy capacity
Grottaglie (Puglia)	Leonardo industrial site	Aerospace/defense-linked infrastructure and land availability
Genoa	Leonardo “Davinci-1” / IIT	Submarine cable connectivity; robotics research base

Sources

Chapter 1 — The Regulatory Architecture: CADA, Sovereignty Levels, and the Chips Act 2.0 Complement

Why CADA exists

On 3 June 2026, the European Commission adopted the Cloud and AI Development Act (CADA), a regulation “establishing a framework of measures for strengthening Europe’s cloud and AI ecosystem.” The proposal sits at the heart of the Commission’s broader Tech Sovereignty Package and targets two vulnerabilities: a structural deficit in EU data-centre capacity, and dependence on a small number of non-EU cloud providers. The Commission’s own numbers frame the urgency: EU-based providers’ share of the European cloud market fell from roughly 29% in 2017 to around 15% by 2022. Thomson Reuters Thomson Reuters

The legal identity of the text is now fixed: it is COM(2026)502, procedure 2026/0138(COD) — a co-decision file, meaning both Parliament and the Council must negotiate and agree the final wording through trilogue before it becomes binding law. Nothing in CADA is final yet; it is a proposal.

The Commission’s own framing of the goal is deliberately dual-edged — sovereignty and openness: CADA aims to make it easier and faster to deploy sustainable cloud and data-centre infrastructure, and to accelerate the rollout of cloud and AI for critical sectors, “while keeping the market open to our partners.” That balancing language is exactly what industry groups dispute (see 1.4). Regulations

The Sovereignty Framework — how “Level 4” works

The mechanism doing the actual political work inside CADA is a graded cloud sovereignty framework applying to most providers serving public-sector bodies. At the top sits “Level 4” — providers with full transparency and control over their own software supply chain and no interference from a third country; below that, the Commission retains a role to formally recognise third-country providers case by case. Legal analysts consider this system likely to be the central battleground of the whole file: the proposal to build this cloud sovereignty framework is expected to be “one of the central political battlegrounds in trilogue negotiations between the Council and the European Parliament.” Regulations Cyber Risk GmbH

Procurement-side, the text doesn’t ban non-EU vendors outright — it nudges: public bodies are encouraged to avoid vendor lock-in through multi-cloud/multi-vendor strategies, to weigh whether a provider generates EU added value (software or hardware designed in the EU, or R&D stemming from EU-funded programs), to support innovation procurement and SME participation, and to benefit from joint purchasing or shared cloud services. Crucially for the siting question in Chapter 3, Member States would be required to facilitate “data centre acceleration zones” — dedicated areas engineered to guarantee adequate power capacity for data-centre development across the EU. Data Center Dynamics Data Center Dynamics

The Continental Cloud Severance Scenario: Strategic Geopolitical Implications of a Rapid Transatlantic Data Architecture Disconnection

CADA’s place inside the wider “Tech Sovereignty Package”

CADA does not travel alone. The Commission published it as one plank of a four-part package: the European technological sovereignty package covers four areas — securing the semiconductor base for AI (Chips Act 2.0), plus AI, cloud, and open source. The full official communication is here: Strengthening Europe’s tech sovereignty — European Commission. Aeneas

The Chips Act 2.0 complement

This is the piece your brief flags as critical, because CADA governs cloud and software sovereignty but says almost nothing about silicon sovereignty — that gap is what Chips Act 2.0 is built to close. It was adopted the same day, 3 June 2026. The Commission frames the rationale bluntly: “the EU remains dependent on third countries in key areas such as advanced chip manufacturing or semiconductor design,” and securing a stable chips supply is treated as necessary “to ensure that critical infrastructures and technologies remain secure, resilient and aligned with European values.” SEMI

The scale of the dependency being addressed is stark: the Commission states Europe currently accounts for less than 10% of global semiconductor production and remains particularly exposed in critical technology segments. Chips Act 2.0’s structure has four main priorities: improving conditions for investment and competitiveness; strengthening research, innovation and skills across the semiconductor ecosystem; and accelerating permitting procedures, with approvals targeted within a maximum of 12 months. A genuinely new element versus the original 2023 Chips Act is a demand-side push: rather than focusing only on manufacturing and research capacity, the new proposal seeks to better align European semiconductor production with the needs of strategic industries, on the theory that stronger local demand reinforces local production and supports commercialisation of European technologies. It also creates a new regional-incentive label: a “Semiconductor Regions of Excellence” designation intended to help regional manufacturing clusters attract investment, alongside faster permitting. European Parliament + 3

Two things are worth being precise about, because they cut directly against any narrative of imminent chip independence. First, this is still a proposal, not law: Reuters, as cited by eeNews Europe, reported that both Chips Act 2.0 and CADA “must complete negotiations before entering into force.” Second, and more important for the Gigafactory discussion in Chapter 2: Chips Act 2.0 targets future European chip design and fabrication capacity over a multi-year horizon; it does not retroactively make the NVIDIA Grace/Blackwell silicon and Quantum-X800 networking already specified for Italy’s IT4LIA and for the InvestAI gigafactories European in origin. The “sovereign compute” being built right now, in 2026–2028, sits on non-EU chips regardless of what Chips Act 2.0 eventually delivers — a timing mismatch the source document’s list of hidden dependencies (Point 7) is explicitly worried about, and one that’s structural, not incidental. The full official Chips Act 2.0 page, including the impact assessment annexes, is here: Proposal for the Chips Act 2.0 — Shaping Europe’s Digital Future. European Commission

Industry pushback

Reaction split along predictable lines. Non-EU hyperscaler-aligned groups call the sovereignty framework protectionist: the Computer & Communications Industry Association (CCIA Europe) warns that “by excluding trusted international technology providers based on their headquarters location and organisational structure, the Commission forces users to rely on a much more limited selection of digital products,” arguing this undermines the EU’s own digitalisation goals. The Commission’s counter-position is that this represents a genuine break in EU regulatory tradition rather than an incremental tightening: commentary notes CADA is a departure from prior EU digital law, which “centered on consumer protection, data governance, compliance obligations, market transparency, or the protection of fundamental rights,” rather than explicit sovereignty-driven procurement rules. European Commission The Data Advisor

Sources for this chapter (all live, all cited above):

Chapter 2 — The Gigafactory Economy: InvestAI, Siting Logic, and Power/Water Physics

Two-tier architecture: AI Factories vs. AI Gigafactories

Before analysing gigafactory siting, it’s essential to separate two EU programs that get conflated constantly in press coverage — they are different in scale, funding source, and governance, and confusing them is the single most common analytical error in Italian commentary on this topic.

	AI Factories (EuroHPC)	AI Gigafactories (InvestAI)
Total EU funding	~€10 billion	€20 billion ring-fenced within a €200bn InvestAI mobilisation target
Number of sites	19 operational/selected across 16 Member States, plus 13 lighter-touch “Antennas”	Up to five sites total, EU-wide
Chip scale per site	Roughly 20,000 GPUs	~100,000 next-gen AI chips — roughly 4× the density
Funding split	Predominantly EU/Member State co-financing (Italy’s IT4LIA was 50/50 EU–Italy, €420m total)	~30% public / ~70% private, CERN-style public-private consortium
Governing regulation	EuroHPC JU	Council Regulation (EU) 2026/150 expanded EuroHPC’s mandate to cover gigafactories and quantum
Purpose	SME/startup access, national HPC capability	Training the most complex, frontier-scale AI models

Italy already has the first tier — it does not yet have, and is competing hard for, the second.

The InvestAI timeline — from announcement to construction

The program has moved through a defined sequence of institutional milestones, laid out on the Commission’s own tracking page:

Date	Milestone
9 February 2025	Von der Leyen announces the InvestAI Facility at the Paris AI Action Summit
9 April – 20 June 2025	Informal call for expressions of interest — 76 submissions from 16 Member States, representing >€230 billion in proposed investment
22 October 2025	Memorandum of Understanding between the Commission and the EIB on co-financing structure
16 January 2026	EuroHPC JU Regulation formally amended to include the AI Gigafactory (AIGF) mandate
Summer 2026	Formal call for tender (in progress at time of writing)
2027	Construction begins on the first AIGF site

This matters for the Italian bid’s credibility window: Italian analysts note the gigafactory the Leonardo–Eni–Fondazione consortium is bidding for is meant to be operational by 2028 — a genuinely tight timeline once you factor in the grid and permitting realities in Section 2.5.

Siting logic: why gigafactories cluster where they do

Contrary to a “digital economy, location doesn’t matter” intuition, gigafactory siting is overwhelmingly a physical infrastructure decision, not a talent or research decision. Early Commission-adjacent scenario planning already pointed to specific candidate geography: early scenarios point to locations in major digital and industrial hubs such as Germany, France, Spain and Italy, including the possibility of two sites in Germany. Three factors dominate site selection, in descending order of binding constraint:

Grid interconnection capacity — see 2.5 below; this is now the single hardest constraint in Europe.
Water availability and cooling infrastructure — see 2.4.
Fibre backbone and international connectivity — a secondary but real factor: Italy’s Genoa bid leans explicitly on this, citing Mediterranean submarine-cable connectivity.

The Italian national bid’s internal site allocation reflects exactly this logic — it is not evenly distributed by symbolic prestige (e.g. Rome, Milan proper) but by where power and land already exist:

Site	Consortium anchor	Primary siting rationale
Lombardy (largest share)	Eni	Industrial energy infrastructure, HPC6 supercomputer at Pavia
Grottaglie, Puglia	Leonardo	Existing industrial/aerospace site, land availability
Genoa (rival/complementary bid)	Leonardo “Davinci-1” + IIT	Submarine cables, robotics research, Mediterranean position
Bologna (Tecnopolo, separate tier)	CINECA / IT4LIA	Already-operational AI Factory, not itself the gigafactory host — supplies skills pipeline and user base instead

Note that TIM joined the consortium in a leading role and Fastweb, Fibercop, and CINECA are also involved, per reporting on the industrial team — this is the “unicum” the sources describe: the only national bid built as a single coordinated proposal across major industrial champions and institutions, rather than competing regional bids as seen in some other Member States.

2.4 The power physics: what 100,000 chips actually demand

This is where policy rhetoric meets thermodynamics, and where most public discussion of “gigafactories” badly understates the ask.

Scale reference	Power draw	Source
Single H100-class GPU	700–1,200 W	TechPlusTrends
Legacy AI server rack	10–15 kW	Tom’s Hardware
Latest NVL72-class rack (GB300)	Up to ~150 kW	Tom’s Hardware
Next-gen “Vera Rubin”-class rack	Up to ~230 kW	Tom’s Hardware
100,000-GPU facility (a “gigafactory”)	~200 MW sustained draw from the grid — comparable to a small city; other estimates for denser configurations run 300–500 MW	AMCompute / USPEGlobal
Full-scale “AI gigafactory” campus at maturity	Up to ~1 GW, comparable to a nuclear reactor’s output	TechPlusTrends

Two important caveats for report accuracy: a large share of that power never reaches the GPUs at all — it’s consumed by cooling and power-delivery overhead, with GPUs accounting for roughly 40% of total facility power usage according to Epoch AI’s estimate cited there. And figures vary by a factor of 2–5× depending on chip generation and utilization rate — there is no single authoritative “gigafactory = X MW” number; the range above (200MW–1GW) reflects genuine uncertainty in public sourcing, not sloppy citation.

The grid bottleneck: Europe’s real constraint isn’t chips, it’s copper and permits

This is the section that should worry Italian policymakers most, because it is largely outside their control on any 2027–2028 timeline.

European grid connection queues are long and getting longer. Within the EU, wait times for securing a grid connection range from two to ten years depending on the country, and in established European and North American hubs the average wait time for a new large-scale grid connection is now seven to ten years, with some projects facing delays up to 13 years. Globally, more than 2,500 GW worth of projects — including data centres — remain stalled in grid connection queues.

Italy-specific data is directly relevant and, notably, flags Italy explicitly: Austria, Greece, Finland, Hungary, Italy, Portugal and Slovakia are projected to see data-centre electricity consumption increase by three to five times over the next several years — meaning Italy sits in the group of countries where grid stress from AI infrastructure is expected to be most acute, not least. And at the continental level, Terna’s queue for data-centre connection requests stood at roughly 30 GW by the end of 2024, with almost 80% of those requests registered only in the prior 12 months — a sudden speculative surge that complicates any planner’s ability to distinguish real projects from queue-squatting.

The macro numbers on the demand side confirm the scale of what a gigafactory adds to an already strained system: European data centres consumed an estimated 96 TWh of electricity in 2024 — 3% of the region’s total demand — projected to rise to 168 TWh by 2030 and 236 TWh by 2035, and data centres are expected to account for roughly a quarter of all European power demand growth by 2030, according to S&P Global. The IEA is more conservative in its base case, estimating data centres will account for about 10% of EU electricity demand growth to 2030, but flags the harder problem explicitly: the capacity implied by the current project pipeline is around 130% of installed capacity today, yet by 2030 installed capacity is only projected to grow by 70% compared with 2024 — a structural shortfall between what’s announced and what the grid can physically deliver. This is precisely why CADA’s mandated “data centre acceleration zones” (Chapter 1) matter operationally, not just symbolically: without pre-cleared grid capacity, Italy’s 2027 construction start / 2028 operational target is at serious risk regardless of how strong the industrial consortium is.

The water dimension: the quieter constraint

Metric	Figure	Source
Water use, 100 MW U.S. data centre	~2 million litres/day total, ~725,000 litres/day (~40%) on-site	Congress.gov CRS
Global AI-related water footprint, 2025	312.5–764.6 billion litres	arXiv environmental-risk assessment
Projected global AI water use by 2030	Up to 600 billion gallons/year, exceeding the electricity demand of Nigeria	Tom’s Hardware / UNU
Embedded (supply-chain) water share	Can exceed 99% of one major operator’s corporate water footprint — fabrication, hardware production dominate direct cooling	arXiv
Training footprint, GPT-3 scale model	~5.4 million litres total, ~700,000 litres on-site	Li et al. 2023, cited in arXiv review

The practical implication for Italian siting: Lombardy and Puglia sit in different hydrological regimes — northern Italy has more reliable freshwater access via the Po basin, while Puglia is a historically water-stressed region already reliant on aqueduct transfers. A 100,000-GPU facility’s on-site cooling demand alone (extrapolating from the 100 MW/725,000 L/day US benchmark to a ~200–500 MW gigafactory) could run into the millions of litres per day — a siting variable the public Italian bid documents do not appear to address in the reporting available, and one that regional water authorities (Autorità di Bacino, ATO Puglia) will need to weigh in on before any 2027 construction start is realistic.

Sources for this chapter:

Chapter 3 — Italy’s Position: Candidacy, Hidden Dependencies, and Cross-Sector Exposure

Where Italy actually stands today

Before assessing risk, it’s worth being precise about what Italy has already secured versus what remains contested, because Italian political rhetoric sometimes blurs the two.

Layer	Status	Detail
EuroHPC AI Factory	Secured, operational	IT4LIA at Bologna’s DAMA Tecnopolo, operational since September 2025, €420m total investment, 50/50 EU–Italy co-financed, targeting over 160 exaflops of peak AI inference performance
National AI supercomputing base	Secured	Leonardo supercomputer, ranked among the world’s top-10 systems; Eni’s HPC6 at Pavia and Leonardo (company)’s Davinci-1 at Genoa
AI Gigafactory (InvestAI)	Contested — bid submitted, not awarded	Italy is one of 16 Member States among 76 total proposals; the Commission has not yet announced winners
Chip-level sovereignty (AI accelerators)	Absent	No EU or Italian fab produces AI GPUs; all cited systems run on NVIDIA Grace/Blackwell + Quantum-X800 InfiniBand
Cloud sovereignty (public sector)	Partial, hyperscaler-dependent	See 3.3 — Italy’s “sovereign” cloud runs on Google, Oracle, and Microsoft technology

The industrial coalition and its geography

Minister Adolfo Urso has publicly staked confidence on the outcome: he states Italy was the only country to present a single common national proposal, and that a gigafactory will be assigned to it. The consortium itself has grown progressively broader:

Phase	Development
November 2025	Core proposal formed: Eni (Lombardy, larger share) + Leonardo (Grottaglie, Puglia), with Fibercop, Fastweb, CINECA
Mid-2026	TIM joins in a leading role
Ongoing, mid-2026	Genoa advances a competing/complementary bid built on Leonardo’s Davinci-1 and IIT, backed by government undersecretary Alessio Butti, who cites Mediterranean submarine-cable position
July 2026	Reporting notes the EuroHPC gigafactory tender itself has slipped — originally expected February 2026, now expected around July 2026

Leonardo’s own disclosures confirm the stakes it attaches to this: the company states it is part of the shortlist of companies that have finalised the Italian proposal for one of the five EU AI Gigafactories, and has built a dedicated business line — the Leonardo Hypercomputing Continuum (LHyC) — through which it expects to earn €230 million over the 2025–2029 business plan by extending compute services beyond defense and aerospace.

The independent-analysis caveat from Chapter 2 bears repeating here in sharper form: Italy’s national bid is described as a “unicum” in the European landscape — the only country to have built a coordinated national candidacy across major industrial champions and institutions, which is a genuine structural advantage in a field of 76 competing regional/national proposals. But coordination is not the same as feasibility, and the same analysis frames the core risk as a “broken scale” dynamic: massive investment at the top of the technology stack (supercomputing, frontier models, sovereign hardware) alongside fragile, underdeveloped intermediate rungs — university-ITS-enterprise alliances — that operate on five-to-ten-year cycles even as the gigafactory target is 2028.

Hidden Dependency #1 — “Sovereign” cloud still runs on non-EU stacks

This is the least-discussed and most consequential finding in Italy’s own public-sector infrastructure documentation. The Polo Strategico Nazionale (PSN) — Italy’s flagship sovereign-cloud vehicle for public administration, part-owned by TIM, Leonardo, and Cassa Depositi e Prestiti — is explicitly built as a sovereignty wrapper around hyperscaler technology, not a replacement for it:

PSN Service Tier	Underlying technology	Sovereignty mechanism
PSN Managed Public Cloud	Google Assured Workload and Oracle Alloy	PSN staff control release/deployment procedures; data hosted in Italian regions or PSN Data Centres
Secure Public Cloud	Microsoft Azure and Google Cloud	Cryptographic keys (BYOK/HYOK) exclusively managed by PSN, not by the hyperscaler
Licensed Private/Hybrid Cloud	Hyperscaler technology, licensed	Operational management by a designated national provider under national authority surveillance

The frank version of this tension has already surfaced in Italian policy commentary: if a hyperscaler underpinning this stack were “cyber-abbattuto” (knocked out by a cyberattack), can the situation still be considered non-critical for the country? What happens to 2,000 municipalities left without service? Have some providers effectively become “too cloud to fail”? — a rhetorical framing, but one that captures the real structural point: encryption-key control and contractual oversight are meaningful but partial sovereignty; the underlying compute, patching cadence, and much of the software supply chain remain governed by decisions made in Redmond, Mountain View, and Austin, not Rome. This is exactly the gap CADA’s “Level 4” assurance tier (Chapter 1) is designed to eventually formalize and grade — but as of today, Italy’s own flagship sovereign-cloud program sits at an intermediate level, not the top one.

Hidden Dependency #2 — Chip sovereignty is real, but the wrong chips

Italy has a genuine, substantial semiconductor manufacturing footprint via STMicroelectronics — but it is concentrated in a technology segment that does not overlap with AI accelerators:

ST Italy site	Technology	EU/Italian funding	Relevance to AI Gigafactories
Catania	Silicon Carbide (SiC) power devices, €5bn Silicon Carbide Campus	€2bn from the Italian government under the EU Chips Act	Indirect only — SiC devices matter for data-centre power-delivery efficiency, not compute itself
Agrate	300mm fab, MEMS, smart power (BCD) technologies	EIB financing, part of a €1bn EIB–ST credit line	None — analog/power semiconductors, not AI logic

The blunt conclusion: every AI Factory and every planned Gigafactory site in Italy runs on imported AI silicon. The IT4LIA system combines NVIDIA Grace CPUs with NVIDIA Blackwell GPUs over Quantum-X800 InfiniBand networking, and there is no indication in current planning that Italy’s future Gigafactory will use anything else. Chips Act 2.0 (Chapter 1) is explicitly aimed at closing this gap over the medium term — but its own text acknowledges Europe currently accounts for less than 10% of global semiconductor production, and its 12-month permitting target and “Regions of Excellence” incentives are aimed at future fabs, not at retrofitting sovereignty onto hardware already being procured for 2027–2028 gigafactory construction. Sovereign compute, under current plans, is sovereign in data governance and cloud-service control, not in silicon.

Defense — the sector most already entangled with AI compute

Defense is the sector where Italy’s compute ambitions are furthest along and most directly coupled to the Gigafactory bid, principally through Leonardo:

Program	Focus	Detail
ARCHYTAS Project	Next-generation AI accelerators for defense applications	Part of $710 million allocated to Leonardo-led EDF (European Defence Fund) activities
MILSCA (Military Space Cloud Architecture)	Space-based HPC/AI storage and processing	Over 100 terabytes storage and 250 TFLOPS processing per satellite, using Leonardo’s Davinci-1 supercomputer; run under the Telespazio (Leonardo/Thales) and Thales Alenia Space joint ventures
“Michelangelo Dome”	Multi-domain defense architecture, AI-enabled threat anticipation	Part of Leonardo’s 2026–2030 Industrial Plan, ~€21bn in new business opportunities projected over the decade, ~€6bn realised 2026–2030
Leonardo Hypercomputing Continuum (LHyC)	Dual-use compute services beyond defense/aerospace	€230 million projected revenue, 2025–2029
Eurodrone MALE RPAS, Future Combat Naval System	AI-integrated sensors and platforms, manned/unmanned	Fielding targeted 2026–2028 under multi-year defense investment law (€12.35bn fund, 2021–2035)

The strategic read here is unambiguous: Leonardo is simultaneously Italy’s primary defense-AI integrator, a Gigafactory bid partner, and a part-owner of the sovereign cloud vehicle (PSN) — meaning a single industrial actor sits at the intersection of every sovereignty-sensitive compute decision the country is making. This concentration is either a coherence advantage (unified strategy, faster execution) or a single point of failure (one company’s execution risk becomes national infrastructure risk), depending on how PSN and Gigafactory governance evolve — CADA’s push toward multi-vendor procurement (Chapter 1) cuts directly against this concentration if applied domestically.

Healthcare — high ambition, fragmented execution

Italy’s National Health Service (SSN) is pursuing meaningful AI integration, funded through the same PNRR mechanism as the broader digital transformation:

AGENAS (the national technical health agency) ran a competitive tender under PNRR Mission 6, Component 1, for an AI Platform supporting primary healthcare, covering diagnostic support, chronic-condition management, patient-history analysis, and AI-enriched telemedicine.
The PNRR allocated €15.63 billion to healthcare, including digitalization.
Regulatory guardrails are already in place: Italy’s data protection authority (the Garante) issued AI-in-healthcare guidance in October 2023 requiring “human in the loop” oversight, explicit legal basis under GDPR Art. 9(2)(g), and professional liability rules for clinicians relying on AI outputs.
Italy passed Law No. 132/2025, the first comprehensive national AI law among EU Member States, including provisions on healthcare-data use for research.

The structural weakness is governance fragmentation, not funding or ambition: the SSN’s decentralised regional governance structure makes equitable implementation and widespread eHealth adoption challenging, creating regional disparities and a digital divide. A national Gigafactory could, in principle, provide the training compute for a genuinely national diagnostic-AI platform — but only if the same 20-region fragmentation that has slowed eHealth rollout doesn’t also slow AI compute allocation across the health system.

SME access — the sharpest version of the “broken scale” risk

This is where the report’s central risk — Italy becoming a computational consumer rather than producer — is most measurable, because there is hard comparative data.

Country	Small firm (10–49 emp.) AI adoption	Medium firm (50–249 emp.) AI adoption	Large firm (250+ emp.) AI adoption
Italy	Under 10%	~14–15%	32.5%
Germany	16.9%	28.2%	48.2%
France	Under 10%	~14–15%	32.7%
EU27 average (all firm sizes)	—	—	20.0% (2025 reference year, up from 13.5% in 2024)

Two additional data points sharpen the picture. First, cost is a real but not dominant barrier: 43.0% of Italian businesses cite high costs as a barrier to AI adoption in a national survey, but a separate analysis notes 70.2% of Italian SMEs already have basic digitalisation infrastructure in place — meaning the gap is skills and trust, not missing hardware. Second, and this is the sharpest number in the whole dataset: OECD analysis finds AI adopters in Italy’s top productivity decile outperform bottom-decile firms by 240% — the largest adopter/non-adopter productivity gap among the G7 economies studied, well above France’s 40% and Germany’s 120%. That combination — low adoption and the largest measured payoff from adoption — is precisely the “broken scale” the Agenda Digitale analysis warns about: Italy is positioned to build some of Europe’s most powerful AI infrastructure while its own SME base, which would benefit most from using it, remains among the least AI-literate in the G7. A Gigafactory hosted on Italian soil does not, by itself, fix this; CADA’s SME-support procurement language (Chapter 1) is voluntary guidance for public bodies, not a mandate to route national compute capacity into SME-facing tools.

Energy — the PA and defense sectors’ shared physical constraint

This closes the loop back to Chapter 2’s power/water analysis, applied specifically to Italy’s grid reality. The relevant national numbers were sourced there: Terna’s queue for data-centre grid-connection requests stood at roughly 30 GW by end-2024, with nearly 80% of those requests filed in just the prior 12 months, and Italy is explicitly named among countries (alongside Austria, Greece, Finland, Hungary, Portugal, Slovakia) projected to see data-centre electricity consumption rise three-to-five-fold in the coming years. Layered onto this, STMicroelectronics’ own Italian operations are pursuing renewable-energy contracting at scale — a 15-year power purchase agreement with ERG for its Catania and Agrate sites — illustrating that industrial energy demand in Italy is already competing for the same renewable-generation capacity a Gigafactory would need. Any Italian gigafactory site (Lombardy or Puglia) will be competing for grid interconnection slots against this existing industrial base, inside a national queue that the EU-wide data suggests is already under acute pressure — reinforcing Chapter 2’s conclusion that the 2027 construction-start target is the single most exposed assumption in Italy’s bid, more than the industrial coalition’s cohesion or the funding structure.

Synthesis: the sovereignty gap, mapped

Dimension	Italy’s status	Sovereignty level
Physical hosting (AI Factory tier)	Secured (Bologna)	High
Physical hosting (Gigafactory tier)	Contested, 2028 target	Uncertain
Cloud governance (public sector)	PSN operational	Medium — key-level control, not full stack
Chip supply (AI accelerators)	None domestic	Low — 100% NVIDIA-dependent
Chip supply (power/SiC semiconductors)	Strong (STMicroelectronics)	High, but adjacent, not AI-core
SME/industrial AI diffusion	Lagging G7/EU peers	Low
Defense AI integration	Advanced, Leonardo-led	High, concentrated in one actor
Healthcare AI	Ambitious, PNRR-funded	Medium, fragmented by region
Energy/grid capacity for AI buildout	Constrained, in EU-wide queue pressure	Low-to-medium, binding constraint

The report’s original framing — Italy risking becoming a consumer of AI built elsewhere rather than a producer — turns out to be only half the risk once you look at the sourced detail. The other half, arguably sharper, is that Italy could host frontier AI infrastructure on its own soil and still functionally be a consumer, if the chips inside it, the cloud layer running on top of it, and the SME base meant to benefit from it all remain governed, supplied, or under-skilled by non-Italian, non-EU, or simply underdeveloped systems. Hosting is necessary but not sufficient for the sovereignty the source document is actually asking about.