Sovereign infrastructure arguments tend to stop at the data centre door. Everything past that door was designed in the United States and packaged in Taiwan, and the quantity available to you was decided by an allocation process you were never part of.
The sovereignty argument still holds, but the supply chain is the layer where the honest answer is partial control. A procurement team that knows which parts of it are fixed can spend its effort on the parts that move.
The binding constraint is packaging
The mental model of a chip shortage is a full fab and a queue for wafer starts. The actual chokepoint sits further downstream, at the advanced packaging step that bonds compute dies to high-bandwidth memory stacks. Without it, a successfully fabricated die is not a functional accelerator.
Concentration at that step is severe. Allocation trackers estimate that a single buyer holds around 60% of TSMC’s CoWoS capacity against roughly a million wafers of annual demand, with the top three customers accounting for more than 85% between them and advanced-packaging lead times running 52 to 78 weeks.
Those figures come from analyst trackers rather than company disclosure, and the trackers themselves caution that individual customer wafer counts are contested, so treat the percentages as indicative and the concentration as real.
High-bandwidth memory is a parallel constraint with its own concentrated supplier base, so scarcity can reappear from a direction unrelated to logic capacity.
What that changes about what you can buy
Two consequences follow for anyone specifying a cluster.
Generational availability is decided upstream, and not just for current production: the same trackers report that the largest buyer has booked more than half of the packaging expansion still being built, with leading-edge logic capacity committed years out.
The practical options are the previous generation, a long wait, or a provider who already holds allocation.
A queue measured in quarters also means the accelerator available when a project starts may differ from the one written into the plan. That argues for specifying a performance envelope rather than a part number, and for treating delivery dates as a contract term rather than a footnote.
Previous-generation accelerators are frequently the rational choice regardless of supply, particularly for inference and fine-tuning where a newer part’s memory headroom goes unused.
The comparison of training GPUs works through where that holds. Scarcity turns a defensible engineering decision into an obvious one.
Where Europe’s position helps and where it does not
The sovereignty conversation usually flattens into whether Europe can build its own accelerators, which obscures where actual advantage sits. Europe holds a genuine chokepoint one layer up, in the machines that pattern the wafers.
ASML is the sole supplier of EUV lithography systems worldwide and holds roughly 83% of the lithography market overall, which means no leading-edge logic process at any foundry, anywhere, runs without Dutch equipment.
That is why export policy from the Netherlands functions as a geopolitical instrument rather than a technical footnote, and it is a more durable position than any fab Europe could fund.
What Europe does not hold is volume. The European Court of Auditors concluded that the bloc is very unlikely to meet the Chips Act target of 20% of global production value by 2030, citing the Commission’s own forecast of a rise from 9.8% to around 11.7%, and noting that meeting the target would require roughly quadrupling production capacity.
The financing asymmetry explains why. Of the €86 billion attributed to the Chips Act through 2030, the Commission manages €4.5 billion, while the largest global manufacturers budgeted €405 billion across three years.
The Technological Sovereignty Package is the policy response, pairing a Chips Act 2.0 proposal with a Cloud and AI Development Act.
The second matters more for procurement, because it proposes a single EU-wide framework for assessing how sovereign a cloud or AI service actually is, graded from basic requirements about where data sits through to control over the software supply chain.
Both remain proposals moving through the legislative process, so neither changes what you can buy now, though the assurance framework is worth reading early if you expect to sell into the public sector.
Export controls bite the supporting hardware first
The politics are the part Europe can least insulate, and the exposure runs both ways.
Restrictions on lithography sales and servicing constrain what European equipment makers ship east, while licensing regimes for rare earths and compound-semiconductor feedstocks constrain what flows west.
The mechanism is routinely overstated, and being precise about it changes the procurement response. Leading-edge logic is silicon so the accelerator die itself is not where the exposure sits.
It runs through compound semiconductors such as gallium arsenide and gallium nitride, through wafer-processing chemistry, and through the electromechanical parts bolted onto finished systems, meaning power delivery, optics, magnets and cooling.
A cluster does not stop being buildable because of a licensing dispute, but its power and interconnect bill of materials becomes harder to schedule.
Treat these controls as a lead-time and price risk on the supporting hardware, and set procurement windows accordingly.
What to put in the contract
Hardware nationality is the wrong question. A handful of contract terms carry most of the risk the supply chain actually creates:
- Allocation rather than availability – Ask what capacity a provider already holds, for which generations, with delivery dates attached. A quote is not an allocation, and the difference between them is a year.
- Generation mix and upgrade path – Establish which generations are obtainable now, and what happens to your pricing and capacity when the provider’s own allocation shifts.
- Jurisdiction of the operator, separately from the location of the racks – Who can be legally compelled to act on a system is a different question from where it physically sits, and the comparison of US hyperscale and EU private cloud sets out why the two come apart.
- Exit and portability terms – A chain this concentrated produces sudden repricing, and the cost of moving a workload is the only real defence against it.
- Residency commitments that survive a capacity crunch – Get in writing what happens to your data and workloads if a provider has to shift capacity between sites, which is covered in the post on data residency in Europe.
None of this makes a European cluster independent of Taiwanese packaging or American design, and claiming otherwise is how sovereignty arguments lose credibility.
The silicon is a commodity you queue for, while the legal perimeter around the workload running on it is a choice, and the data sovereignty pillar sets out what that perimeter has to cover.
Neurotechnology Cloud engineers AI factory environments and operates private AI cloud services on EU-resident infrastructure in the Baltics through its AI Factory engineering service, with operational control held entirely within the European Union.
If you are planning cluster capacity under these constraints, reach out to us here.
