THE GIGAFACTORY PARADOX
Capacity without reproduction is not sovereignty.
Europe is procuring the largest AI computing infrastructure in its history. The question is no longer whether it can build the capacity — it is whether the capacity, once built, can still be reproduced without another actor's permission.
The AI Architecture of Compute Capacity
Location tells us where a capability resides. Architecture tells us upon whom that capability depends.
AXISYNC POSITION PAPER 46 · 2 SEPTEMBER 2026
Europe is preparing to build some of the largest artificial-intelligence computing infrastructures in its history.
The ambition is substantial. Under the European High Performance Computing Joint Undertaking, the European Union has moved from the AI Factory model toward a new generation of AI Gigafactories: very large computing environments intended to train and operate frontier-scale artificial-intelligence systems. On 30 July 2026, EuroHPC formally opened the call for the consortia that will establish and operate them. The tender closes on 12 November. Up to seven Gigafactories are expected to be selected, with decisions anticipated in early 2027 and operations expected within eighteen months of contract signature.
This is important.
Europe is no longer merely debating whether it should possess sovereign AI capacity. It is beginning to procure it.
But procurement introduces a harder question than ambition.
What exactly does it mean to possess capacity?
A machine can stand on European soil. A data center can operate under European law. A consortium can be headquartered inside the European Union. Public institutions can purchase guaranteed access to its computing resources. European researchers, companies and governments can run models through it.
All of those things matter.
None of them, by themselves, prove sovereignty.
A capability is not sovereign merely because it is located inside a sovereign territory.
The deeper question is whether the institution possessing that capability can continue to operate it, modify it, repair it, expand it, substitute critical components and ultimately reproduce it when circumstances change.
That distinction is where the European AI Gigafactory project becomes much more than an infrastructure program.
It becomes an experiment in the architecture of optionality.
From Authority to Capacity
In the previous AXISYNC paper, Authority Without Recall Is Not Control (Position Paper 45), we examined the first layer of a larger institutional spine.
Authority asks:
Who may act?
It concerns delegation, permissions, decision rights and the ability of an institution to withdraw authority once it has been granted.
Capacity asks a different question:
Where does action become possible?
Authority without capacity is ceremonial. An institution may legally possess the right to act and still lack the systems, infrastructure, knowledge, capital, energy or technologies required to exercise that right.
Artificial intelligence makes this distinction particularly visible.
A government may possess the legal authority to develop an AI system. A university may possess the scientific expertise. A company may possess the models, data and customers.
But without access to enormous quantities of compute, those actors may still be unable to act.
Compute therefore becomes more than infrastructure.
It becomes an enabling condition of agency.
And whoever controls an enabling condition can influence the range of actions available to everyone who depends upon it.
That is why the AI Gigafactory debate should not begin with the size of the building.
It should begin with the architecture beneath the building.
The Factory Is Not the Capacity
The word Gigafactory creates a powerful physical image.
Buildings. Processors. Cooling systems. Power substations. Fiber connections. Thousands of machines operating together.
But the European regulation establishing the AI Gigafactory framework already reveals that the real object is considerably more complex. It defines an AI Gigafactory as a state-of-the-art large-scale facility capable of handling the complete lifecycle of very large AI models, from development to large-scale inference — combining AI-optimized computing capacity with supporting data-center infrastructure, high-capacity storage and networking, dedicated secure cloud access environments, specialized AI-oriented support services, and the energy and water systems required to sustain them. The expected scale is on the order of 100,000 advanced AI processors per site, roughly four times the capacity of Europe's largest existing AI Factories.
That description is important because it reveals something easily missed.
The factory is not a single asset.
It is a dependency stack.
The processors depend on memory.
The memory depends on semiconductor manufacturing.
The processors and memory depend on networking.
The network depends on software.
The software depends on compilers, libraries, drivers and development environments.
The entire system depends on enormous amounts of electricity.
Electricity depends on generation, transmission and grid stability.
Cooling may depend on water and specialized equipment.
Operations depend on engineers.
Engineers depend on access to knowledge, tools and replacement components.
Expansion depends on capital.
And almost every layer eventually depends upon another supply chain.
The building is therefore only the visible shell around a network of permissions, suppliers, technologies and dependencies.
This produces the first principle of the Gigafactory Paradox:
Installed capacity and controllable capacity are not the same thing.
Europe Has Already Recognized the Difference
Perhaps the most consequential development in Europe's digital policy architecture is that European institutions themselves are beginning to distinguish between location and sovereignty.
The European Commission's proposed Cloud and AI Development Act, published on 3 June 2026, creates four levels of cloud and AI sovereignty.
At the first level, data is processed and stored on infrastructure located inside the European Union.
At the second, providers must demonstrate greater independence from third countries and provide transparency regarding their software supply chains.
At the third, ownership and control from within the European Union become relevant.
And at the highest level, providers must demonstrate full transparency and control over the software supply chain together with protection from third-country interference.
That classification contains a remarkable admission.
Infrastructure located in Europe represents only the first level of sovereignty.
Geography is therefore not the conclusion of sovereignty.
It is merely its beginning.
This distinction should fundamentally change the way Europe evaluates AI infrastructure.
A computer does not become strategically independent when it crosses a border.
A data center does not become sovereign because its address changes.
A cloud environment does not become autonomous simply because European law governs the property on which its servers stand.
Location tells us where a capability resides.
Architecture tells us upon whom that capability depends.
The Silicon Question
Nowhere is the distinction clearer than in the processor layer.
Europe needs frontier-scale computing capacity now. The fastest route toward obtaining that capacity necessarily involves technologies produced by companies that already dominate advanced AI computing.
The European Commission has signed letters of intent with AMD, Nvidia and Qualcomm to support the access of Gigafactory consortia to the hardware they will require. According to the Commission's own description of the arrangement, the letters do not close the door to other suppliers: consortia remain free to procure from hardware providers in Europe or in like-minded countries.
There is nothing inherently irrational about this.
Strategic autonomy should not be confused with technological isolation.
If the strongest available processor architecture comes from outside Europe, refusing to use it solely because of its origin could reduce European competitiveness rather than increase European sovereignty.
The mistake would lie somewhere else.
The mistake would be to install today's strongest architecture in such a way that tomorrow's alternative becomes economically, technically or institutionally impossible.
That is the difference between procurement and architecture.
Procurement asks:
What should we buy?
Architecture asks:
What options remain after we buy it?
A sovereign system does not necessarily require every component to be domestic.
It requires critical dependencies to remain visible, governable and, where necessary, substitutable.
Italy Is Already Showing the Second Path
The emerging Italian AI Factory architecture provides an instructive example.
The new IT4LIA AI supercomputer in Bologna, contracted by EuroHPC in April 2026, is being manufactured by Dell Technologies and integrated by Italy's E4 Computer Engineering on Nvidia's liquid-cooled GB200 NVL4 architecture, combining Grace CPUs, Blackwell GPUs and Nvidia Quantum-X800 InfiniBand networking. It is therefore able to take advantage of one of the strongest currently available AI computing ecosystems.
But the architecture also contains something else.
A dedicated inference partition is being built with accelerators from Netherlands-based Axelera AI together with European-designed SiPearl CPUs. The €290 million system therefore combines leading global technology with a deliberately maintained European technological pathway.
That small architectural decision may ultimately matter more than its current percentage of total compute capacity suggests.
Europe does not need an immediate replacement for every foreign processor.
It needs the ability to prevent the foreign processor from becoming the only conceivable future architecture.
That is optionality.
A second path does not need to dominate the first path in order to matter.
It needs to remain viable.
Because once the second path disappears, negotiation changes.
Pricing changes.
Dependency changes.
And eventually control changes.
Sovereignty Is Not Autarky
This distinction matters because discussions of technological sovereignty frequently collapse into an unhelpful binary.
Either Europe builds everything itself, or Europe is dependent.
That is too crude.
No major technological economy is completely self-contained. Advanced semiconductor systems themselves contain extraordinary networks of international interdependence involving lithography, design software, fabrication, packaging, memory, specialty chemicals, optics, networking and intellectual property.
The objective therefore cannot reasonably be the elimination of dependency.
The objective must be the architecture of dependency.
Some dependencies are beneficial.
Some create efficiency.
Some accelerate innovation.
Some allow specialization.
Others remove alternatives.
The relevant strategic question is consequently not:
Are we dependent?
Every modern technological system is.
The relevant questions are:
Can we see the dependency? Can we govern it? Can we substitute it? Can we survive its interruption?
Optionality begins when those questions are answered before dependency becomes necessity.
Access Is Not Ownership
The funding architecture of the Gigafactory program introduces another important distinction.
Under the amended EuroHPC regulation, the Union contribution can cover up to 17 percent of the capital expenditure of a Gigafactory's overall computing infrastructure. Participating Member States must at least match that contribution, while the remaining investment and the operating expenditure are borne by the Gigafactory consortium. And the July call is structured accordingly: EuroHPC and the participating states are not buying the facilities. They are jointly procuring guaranteed compute access time from privately led consortia, with public funding acting as the anchor customer that de-risks the private capital behind them.
This is an intelligent mechanism for mobilizing private capital.
It also creates an important architectural question.
What precisely does public funding secure?
Access is valuable.
Guaranteed access can protect researchers, startups, governments and European companies from complete exclusion from scarce computing resources.
But access and ownership are different forms of control.
And ownership itself does not necessarily equal technological control.
An institution may own the company operating the infrastructure while remaining dependent on another party for processors.
It may own the processors while remaining dependent on proprietary software.
It may control the software but depend on external cloud infrastructure.
It may control all of those and still depend on an energy grid that cannot support expansion.
Each layer changes the institution's possible actions.
The architecture therefore needs to be assessed as a whole.
Europe Is Trying to Put Control Into the Governance Layer
The European legislation deserves credit for recognizing part of this problem.
The coordinator of an AI Gigafactory consortium must be incorporated and headquartered inside the European Union and must ultimately be controlled by entities or individuals established within the Union. Hosting agreements must also contain mechanisms enabling European authorities to exercise the scrutiny and control necessary to protect strategic assets, autonomy, interests and security. And the tender itself asks bidders to describe the measures they will take to avoid lock-in to their suppliers.
This is not superficial.
Europe is deliberately inserting sovereignty into the institutional governance layer.
But governance cannot compensate indefinitely for technological dependency beneath it.
A European board of directors cannot manufacture a replacement processor.
European incorporation cannot restore a discontinued software stack.
A European contract cannot create grid capacity.
A European ownership structure cannot automatically reproduce a supply chain.
Legal control and physical capacity are different architectural layers.
Both matter.
Neither can substitute entirely for the other.
The Cloud Problem Beneath the Gigafactory
The same distinction already exists in European cloud infrastructure.
Synergy Research Group reported in 2025 that European providers held approximately 15 percent of their regional cloud market, while Amazon, Microsoft and Google together accounted for roughly 70 percent. European providers had more than tripled their revenues since 2017 — but the overall market had grown by a factor of six.
Europe therefore confronts an unusual situation.
Its digital economy can grow dramatically while its relative control over some of the infrastructure enabling that growth remains limited.
This is why building more data centers alone cannot solve the sovereignty problem.
Europe could triple the amount of computing infrastructure physically located inside its borders and still preserve much of the dependency architecture that existed before the expansion.
More capacity can therefore produce more dependency when capacity is expanded through a non-substitutable architecture.
That is the Gigafactory Paradox.
Scale can increase capability while simultaneously decreasing optionality.
The Test Is Not Whether It Works
Traditional infrastructure policy tends to ask whether a system performs.
Does the system work?
Does it deliver the promised computing power?
Does it remain within budget?
Can researchers access it?
Can companies train their models?
Can governments use it?
Those questions remain necessary.
But strategic infrastructure requires another test.
What happens when something underneath the system stops working?
What happens if access to a processor generation changes?
What happens if export policy changes?
What happens if geopolitical relations deteriorate?
What happens if a supplier changes licensing conditions?
What happens when the software architecture evolves?
What happens if energy requirements double?
What happens if the company controlling a critical technology is acquired?
What happens when the next generation of hardware requires an entirely different infrastructure?
These are not predictions.
They are architecture questions.
And architecture should not depend upon predicting which disruption will occur.
Its purpose is to preserve choices when prediction fails.
Capacity Has a Time Dimension
This leads to a distinction that may become increasingly important for policymakers.
There is installed capacity.
And there is reproductive capacity.
Installed capacity tells an institution what it can operate today.
Reproductive capacity tells the institution whether it can recreate, replace, expand or transform the capability tomorrow.
The difference is fundamental.
A country may possess thousands of advanced accelerators without possessing the industrial capability to manufacture a single replacement.
It may operate frontier systems without controlling the tools required to design their successors.
It may possess enormous computing capability while depending on one software environment to make that capability useful.
In such a system, capacity exists.
But its continuation is conditional.
This is why capacity without reproduction is not sovereignty.
Reproduction does not mean that Europe must manufacture every component itself.
It means that Europe must maintain credible pathways by which critical capabilities can be rebuilt, substituted or migrated.
The objective is not self-sufficiency.
It is freedom of movement.
The Second Path Must Be Designed Before It Is Needed
The most dangerous moment to begin searching for an alternative architecture is the moment when the primary architecture becomes unavailable.
At that point, optionality has already been lost.
Second paths therefore need to exist before they are economically obvious.
They may initially appear inefficient.
They may be smaller.
They may cost more.
They may not offer the performance of the dominant architecture.
But strategic optionality frequently looks inefficient during periods of stability.
Its value becomes visible only when conditions change.
This creates a difficult responsibility for governments and corporate boards because conventional efficiency metrics reward concentration.
One supplier is easier than three.
One software environment is easier than two.
One architecture reduces integration costs.
One hyperscaler simplifies procurement.
One standard accelerates deployment.
Efficiency therefore tends naturally toward dependency.
Architecture must provide the counterweight.
A Different Definition of European AI Sovereignty
Europe should therefore resist defining technological sovereignty through flags, headquarters or server locations alone.
A more useful definition would be architectural.
European AI sovereignty exists to the extent that European institutions retain credible freedom of action over the critical systems enabling artificial intelligence.
That means understanding where compute can be obtained.
Who controls access.
Who controls the processors.
Who controls the software.
Who controls the network.
Who controls the energy.
Who can repair the infrastructure.
Who can expand it.
Who can substitute its components.
And who can reproduce the capability after the current generation becomes obsolete.
None of these questions requires Europe to disengage from its allies or global technology companies.
Quite the opposite.
Europe should remain deeply integrated into the strongest global technological ecosystems available.
But integration should be chosen.
Dependency should be visible.
And exit should remain conceivable.
That is the difference between participation and architecture.
The Gigafactory Opportunity
The European Gigafactory initiative should therefore be understood as something larger than an industrial investment program.
Europe is building an architecture that may determine who can develop frontier AI systems on the continent for decades.
The decisions being taken now will establish interfaces, procurement practices, energy commitments, software environments, supplier relationships and capital structures that can become extraordinarily difficult to change later.
That is precisely why the current moment matters.
The concrete has not yet hardened.
The dominant interfaces have not yet become permanent.
The Gigafactory consortia have not yet been selected.
Europe still possesses something enormously valuable.
Design freedom.
The question is whether it uses that freedom merely to acquire capacity or to architect optionality into the capacity itself.
Those are not the same objective.
The Optionality Standard
A genuinely strategic Gigafactory should therefore be judged not simply by how much compute it contains, but by how much freedom survives its construction.
Can major hardware components eventually be substituted?
Can alternative accelerator architectures be introduced?
Can workloads migrate?
Are software interfaces sufficiently open to prevent unnecessary lock-in?
Can critical technologies be repaired or supported without a single external actor?
Is the power architecture expandable?
Can public authorities retain meaningful access during periods of scarcity?
Does the governance model distinguish between ownership, access and technical control?
Are European second-path technologies given enough real workload to mature rather than being kept permanently at experimental scale?
And perhaps most importantly:
If today's preferred supplier disappeared tomorrow, would the institution still possess a path forward?
That is an optionality stress test.
Europe should apply it before awarding the Gigafactories, not after dependency has already been embedded into them.
Architecture Determines Capacity
The temptation in the AI race is understandable.
Count the processors.
Count the megawatts.
Count the billions invested.
Count the models trained.
Count the data centers built.
Those numbers describe scale.
They do not necessarily describe control.
The more consequential measurement may be what cannot easily be counted:
How many credible paths remain?
Because technological sovereignty is not achieved when everything comes from Europe.
It is achieved when Europe's freedom to choose does not disappear because one critical dependency becomes unavailable.
That is the deeper opportunity contained inside the Gigafactory program.
Europe can build enormous computing capacity.
It can also build something rarer.
An architecture in which capacity does not quietly become captivity.
The Verdict: Applying the Doctrine to the Program
A doctrine that cannot deliver a verdict is commentary.
So the question should be put directly. Measured against the four questions of the optionality doctrine — can the dependency be seen, governed, substituted, and survived — is the European AI Gigafactory program set up correctly?
The honest answer is layered, because the program is layered. Read the capacity stack from the top down.
Site — answered. A Gigafactory may be built in one Member State, across several sites, or across borders. Eighteen Member States have signed the joint procurement agreement. Location is the one layer Europe fully controls, and the program treats it accordingly.
Governance — answered. The coordinator must be incorporated, headquartered and ultimately controlled inside the Union. The hosting agreement carries scrutiny and control rights. High-risk vendors are excluded from the supply chain. This is the layer where Brussels has done the most deliberate work, and it should be recognized as such.
Access — answered for the contract term, open beyond it. Public funding buys guaranteed compute access time under a framework contract. That protects European users while the contract runs. What the public sector holds when the framework expires — against privately owned facilities built on privately owned processors — is not specified in anything published so far.
Capital — partly answered. The Union may contribute up to 17 percent of computing-infrastructure capital expenditure, with Member States matching. Private capital therefore carries the majority of the investment and, with it, the practical decisions about expansion and the next generation. And the public share itself is not yet fully in hand: a senior Commission official acknowledged at the launch that only about €1 billion of the Union contribution is committed under the current budget, with the remainder depending on the 2028–2034 Multiannual Financial Framework, which has not been agreed. Public leverage over reproduction rests on money not yet appropriated.
Silicon — partly answered. The Commission has secured hardware access through letters of intent, preserved consortia's freedom to source from European or like-minded suppliers, and made avoidance of supplier lock-in a tender criterion. Those are the right instruments. But letters of intent coordinate supply; they are not orders, guarantees or substitution plans. Anti-lock-in is a criterion to be evaluated, not a threshold that must be met. And we located no requirement that any share of a Gigafactory's workload run on a European second-path architecture — the mechanism that, at IT4LIA, actually keeps the alternative alive. Optionality at this layer is permitted. It is not yet required.
Network — open. The public program architecture is silent on whether the interconnect fabric — the layer that binds a hundred thousand processors into one machine — can be substituted independently of the processor vendor. In current frontier systems it frequently cannot.
Software — open, and the widest gap in the stack. The Cloud and AI Development Act is the instrument that reaches the software supply chain: its second and fourth assurance levels demand transparency and ultimately control over it. But CADA governs cloud providers seeking public-sector contracts, it is a proposal still in negotiation between Parliament and Council, and its adoption is not expected before the Gigafactory awards. The Gigafactory tender, by contrast, has published no comparable standard for software portability. Yet software is where substitution actually fails. A processor can be replaced. A decade of libraries, compilers, drivers and trained engineering habits built around one vendor's environment cannot be replaced by a procurement clause.
Power — open. The regulation requires that a Gigafactory be supported by environmentally sustainable energy and water systems. It treats energy as a condition of sustainability, not as a reproducible capacity. Grid capacity, generation and expansion headroom belong to Member States and utilities the program does not govern. Whether a Gigafactory can double its power draw when the next processor generation demands it is a question nobody in the procurement chain is required to answer.
Reproduction — open. The program procures installed capacity. It contains no mechanism addressing what happens when the installed generation becomes obsolete — who funds replacement, on what architecture, under whose permission. Europe is funding a second path in silicon, including EuroHPC's support of up to €61.6 million for Axelera's next-generation chiplet development, but that path is not at Gigafactory scale and is not connected to the Gigafactory procurement by any obligation.
Put the layers together and the verdict is this.
Europe has set the program up correctly at the top of the stack and left it unanswered at the bottom.
The layers Europe legislated first — location, governance, access — are the layers where sovereignty is most visible and least decisive. The layers where reproduction actually lives — silicon share, interconnect, software, power headroom, next-generation replacement — are the layers the program permits but does not require.
This is not a loophole in the sense of bad faith. It is a sequencing gap. The sovereignty framework that would reach the software layer arrives after the concrete is poured. The private capital that will decide expansion is committed before the public capital that is supposed to anchor it. The second path is funded, but not connected.
Against the doctrine's four questions:
Can Europe see the dependency? Yes — CADA's four levels are the proof.
Can Europe govern it? Yes, at the governance layer — coordinator control and hosting rights are real.
Can Europe substitute it? In principle. Sourcing freedom exists; no obligation converts it into a maintained alternative.
Can Europe survive its interruption? Not demonstrated. Nothing published requires any consortium to show what happens to its Gigafactory if its primary supplier, software environment or interconnect became unavailable.
Two of four answered. That is a program that has understood the problem and not yet finished solving it — which is exactly the position in which architecture still has leverage.
The remedy is not a different program. It is three obligations inside this one, before award: a required and growing share of real workload on a European second-path architecture; a software-portability standard applied to the Gigafactories now rather than inherited from CADA later; and a supplier-withdrawal scenario that every consortium must answer as a condition of selection.
Those are not protectionist measures. They do not exclude anyone. They make the option Europe has already paid to keep open into an option it is obliged to keep alive.
More Choices, or Fewer
A capability borrowed from favorable circumstances can feel indistinguishable from a capability we control — until the circumstances change.
The same is true for institutions.
Prosperity conceals dependency.
Stability conceals fragility.
Abundance conceals the absence of alternatives.
Europe is fortunate because the AI Gigafactories are not yet finished.
The choices are still open.
The question confronting European leaders is therefore not whether Europe can build enough machines.
It almost certainly can.
The question is whether, after spending tens of billions of euros and embedding the technological architecture of the next generation, Europe will have more choices than it has today — or fewer.
Because capacity is not measured only by what an institution can operate.
It is measured by what the institution can still choose when circumstances change.
Paper 45 established that formal authority is insufficient without recall. This paper establishes that installed capacity is insufficient without reproduction. The two conclusions belong to the same spine, and they point to a third layer beneath both: whether the institution still understands why it made its decisions, and whether that knowledge survives the architecture that generated it.
Architecture determines optionality.Optionality determines leverage.Leverage determines control.
And when the architecture of capacity removes the ability to reproduce the capacity itself, sovereignty may remain visible on the building while disappearing underneath it.
Capacity without reproduction is not sovereignty.
RHODES OBSERVATION
Capacity is not lost when it is imported.
It is lost when it can no longer be reproduced.
STOIC REFLECTION
Seneca was asked whether the wise man, being self-sufficient, has any need of friends. His answer drew a line that has outlived its occasion: the wise man is self-sufficient not because he wants no one, but because he could lose anyone and remain himself. He wants friends. He does not need them.
That is the difference between chosen interdependence and dependency. An institution may draw on every ecosystem in the world and surrender nothing of its sovereignty — provided that the withdrawal of one does not remove its ability to continue, substitute, or choose another path.
Sovereignty is not the refusal of help. It is the retained ability to stand when help is withdrawn.
Sources
Procurement architecture
EuroHPC Joint Undertaking, The EuroHPC Joint Undertaking launches the AI Gigafactories Call, and call for tenders EUROHPC-2026-CEI-AIGF-01, 30 July 2026 — supports the opening of the call on 30 July 2026, the 12 November 2026 submission deadline, the selection of up to seven AI Gigafactories, selection in early 2027 with operations expected within 18 months, the structure of the procurement as a joint purchase of compute access time by EuroHPC and participating states, public funding acting as anchor customer, and the expectation of more than €20 billion in private investment.
European Commission, EU launches AI Gigafactories call to boost Europe's computing capacity and unlock more than €30 billion in investment, 30 July 2026 — supports the letters of intent signed by the Commission with AMD, Nvidia and Qualcomm to support consortia's access to hardware, the statement that consortia may also procure from hardware providers in Europe or like-minded countries, and the tender's attention to measures avoiding supplier lock-in.
Reuters, EU aims for seven AI gigafactories with €10 billion plan in race with US, China, 30 July 2026 — supports the €10 billion public envelope, the increase from five to seven planned Gigafactories, and the letters of intent with AMD, Nvidia and Qualcomm.
The Next Web, Europe opens bidding for seven AI 'gigafactories' in a €30 billion push, 30 July 2026 — supports the statement that only about €1 billion of the Union share is committed under the current budget, with the remainder dependent on the 2028–2034 Multiannual Financial Framework, and the senior Commission official's description of the figure as a "best estimate" rather than money in hand.
Regulation and policy
Council Regulation (EU) 2026/150 amending Council Regulation (EU) 2021/1173 on establishing the European High Performance Computing Joint Undertaking, adopted 16 January 2026 — supports the definition of an AI Gigafactory as a large-scale facility handling the complete lifecycle of very large AI models, composed of AI-optimized computing capacity, supporting data-center infrastructure with high-capacity storage and networking, secure cloud user access environments and specialized AI support services, supported by energy and water systems; the scale of more than 100,000 advanced AI processors per site; the 17 percent ceiling on the Union contribution to computing-infrastructure capital expenditure with at least matching contributions from participating Member States; the requirement that the AI Gigafactory Coordinator be incorporated and headquartered in the Union and ultimately controlled by Union-established entities or persons; and the scrutiny and control provisions of the hosting agreement. Council of the European Union, press release Artificial intelligence: Council paves the way for the creation of AI gigafactories, 16 January 2026 — supports the adoption date and the exclusion of high-risk vendors from Gigafactory supply chains.
European Commission, Proposal for a Regulation — Cloud and AI Development Act (CADA), COM(2026) 502, 3 June 2026 — supports the four Union assurance levels of the EU Cloud Sovereignty Framework, from Level 1 (processing and storage on infrastructure located in the EU) through Level 2 (demonstrated independence from third countries and software-supply-chain transparency) and Level 3 (EU ownership and control) to Level 4 (full transparency and control over the software supply chain and no third-country interference). Jones Day, European Commission's Proposed Cloud Sovereignty Framework, June 2026 — supports the statement that CADA remains subject to European Parliament and Council negotiation, with final adoption targeted for late 2027, i.e. after the expected Gigafactory awards.
Silicon and the second path
EuroHPC Joint Undertaking, EuroHPC JU Signs Contract to Boost AI Capabilities with IT4LIA AI Factory, 22 April 2026, and E4 Computer Engineering announcement, April 2026 — supports the IT4LIA system in Bologna hosted by CINECA, manufactured by Dell Technologies and integrated by E4 Computer Engineering on Nvidia's liquid-cooled GB200 NVL4 architecture with Grace CPUs, Blackwell GPUs and Quantum-X800 InfiniBand networking; the dedicated inference partition featuring Axelera AI accelerators and SiPearl CPUs; and the €290 million budget co-funded equally by EuroHPC through the Digital Europe Programme and Italy's Ministry of University and Research.
EuroHPC Joint Undertaking, DARE (Digital Autonomy with RISC-V for Europe) project, March 2025 — supports the grant of up to €61.6 million to Axelera AI for development of a high-performance AI chiplet, cited as evidence that a European second path in accelerators is funded but not yet at Gigafactory scale.
Cloud market structure
Synergy Research Group, European Cloud Providers' Local Market Share Now Holds Steady at 15%, July 2025 — supports the approximately 15 percent regional share held by European cloud providers, the roughly 70 percent share held by Amazon, Microsoft and Google combined, the more-than-tripling of European providers' revenues between 2017 and 2024, and the six-fold growth of the European market to €61 billion over the same period.
Stoic source
Seneca, Moral Letters to Lucilius, Letter 9, On Philosophy and Friendship — supports the distinction between the wise man's self-sufficiency and his desire for friends: he can do without them, but does not wish to.
AXISYNC Partners LLC / axisyncpartners.net | Architecture of Decision Sovereignty / The Gigafactory Paradox | September 2026
Comments