Europe is lagging behind in AI. Powerful models mostly come from the US and China, whose AI companies have superior access to the capital, talent, data and computing power required to compete at the frontier. The AI Gigafactories initiative, as part of the AI Continent Action Plan, seeks to address a key structural weakness in the EU: the lack of computing power, or “compute”, provided by specialized AI chips such as GPUs.1 AI developers need an increasing amount of those chips to train competitive models and deploy them at scale. A Gigafactory would be a large data center housing ~100.000 GPUs.
Counting Gigafactories
The official goal of the European Commission is to build up to 5 Gigafactories. But the math behind that number isn’t entirely clear: According to the Commission, one Gigafactory will cost €3-5bn in capital expenditure. Some experts think that's too optimistic, so let's assume – a bit more conservatively – that the price tag is closer to €6bn. Public funds from the EU/member states will cover about 1/3 of the cost, i.e. €2bn per Gigafactory. The private sector is supposed to cover the rest.
Now, the EU’s InvestAI initiative has officially earmarked €20bn in total for the Gigafactories. This would pay for ~10 of them, rather than just the ≤5 mentioned in EU documents and media reports.
Is that realistic? Enough interest seems to be there: In response to the EU’s open call, 76 parties expressed their interest to build a total of 60 Gigafactories throughout the EU. The initial plan to build 5 Gigafactories could thus turn into something much bigger.
But would it make sense?
Assessing Gigafactories
On the one hand, the EU’s announcement to invest in computing infrastructure is clearly a step in the right direction. As different sectors of society will increasingly rely on AI, compute becomes a strategic resource. Europe doesn't have very much of it, compared to the US and China (and, perhaps in the future, Middle East countries like the UAE or Saudi-Arabia).
On the other hand, the strategic vision is rather opaque: How much compute does Europe need, for what purposes, where, and used by whom? Let’s look at some possible scenarios.
Scenario 1: Europe focuses on AI adoption and heavily integrates US-hosted models throughout the economy and public administration. Only the most security-relevant applications, involving sensitive citizen data or related to national security, run on European data centers.
Rationale: Europe might conclude that it simply cannot compete at training frontier models, or that due to more pressing priorities, it doesn’t want to invest the enormous amount of resources required for that. And it might shy away from training domain-specific models out of fear that leading US companies could easily replicate any such efforts as soon as they get to truly powerful AI systems. In cases where the EU lacks competitive homegrown models and the US is very restrictive about sharing frontier model weights with foreign data center companies,2 the need for sovereign inference capacity also becomes less pressing.3
Implications: The value of Gigafactories in this scenario is not clear. If AI models are trained elsewhere and (mostly) inferenced over the cloud or by US data center companies, the need for sovereign European compute is generally limited. For the few applications that Europe wants to inference at home, smaller, decentralized clusters seem preferable over Gigafactories. Inference-focused clusters can be smaller, because they process smaller batches of data. And ideally, they are dispersed over a country, for two reasons: 1) Inference favors low latency, which means that data centers have to be physically proximate to the users they serve, which are themselves dispersed over the country. Second, for sensitive applications, concentrating inference in one big cluster would create vulnerabilities to sabotage and data theft. Some amount of redundancy and dispersion protect against that, on top of high information security standards. In short, an adoption-focused AI strategy favors small, decentralized, highly secure inference clusters over centralized Gigafactories.
Scenario 2: Europe trains and inferences its own models, but with a focus on specific areas or use cases – e.g. industrial foundation models or public-sector AI tailored to European administrative processes. Other than that, it draws heavily on US-hosted models.
Rationale: With this intermediate strategy, Europe would try to leverage existing strengths to lead in some crucial areas of AI development and deployment, without committing to the massive infrastructure buildout necessary to train foundation models for general use.
Implications: In this case, Europe does need more compute to train and inference its own specialized models – though how much exactly depends on several factors (e.g. which sectors the EU wants to serve with sovereign models, how compute-intensive it will be in the future to train industrial foundation models, how many European companies have the skills and willingness to do that). Right now, there doesn’t seem to be much clarity either among European industry or the EU about these questions. Therefore, it’s hard to assess how many Gigafactories Europe would need to train competitive specialized models. Either way, Gigafactories would have to be complemented with smaller, decentralized clusters for inference.
Scenario 3: Europe aims for sovereign frontier models, trained and hosted in the EU. European countries join forces to amass the enormous capital and computing infrastructure needed for this once-in-a-century effort.
Rationale: If EU leadership recognizes AI as a century-shaping, industrial-revolution-scale technology – and acts accordingly –, it would have to adopt a ‘whatever it takes’ approach. On this view, the safest bet to secure economic competitiveness and geopolitical sovereignty in the AI age might be to compete at the frontier, even if it requires a staggering amount of political ambition and resources.
Implications: In this scenario, Europe needs a lot (!) more compute (and other things, including capital, talent and access to data). However, even 10 Gigafactories would likely not be enough: By the time of their completion, they will be much too small to train frontier models. Even if the EU were to buy 1 million GPUs and bundle them together in one large data center, this would still pale in comparison to the 20-million-GPU superclusters planned in the US and the UAE. If Europe really wants to train (and inference) sovereign frontier models, it needs something much more ambitious than the current Gigafactories initiative.
Conclusion
I'm not sure which of these three scenarios the EU and its member states are currently aiming at. Official statements by political leadership often sound like Scenario 3, the numbers behind the Gigafactory initiative sound more like Scenario 2, and the actual level of political ambition is closer to Scenario 1.
‘Wait and see’ is not the right way to deal with this uncertainty. Given how long it takes to build new data centers and the corresponding energy infrastructure, it's good to see the EU take initiative. But it urgently needs strategic clarity to ensure that infrastructure expansion aligns with its long-term policy goals.
GPUs = Graphics Processing Units. In this blogpost, GPU counts generally refer to Nvidia H100 equivalents – a common metric to compare the computing power of different chip types.
Weights are the mathematical parameters of an AI model: a huge set of numbers that determine a model’s output to a given input. They are the core of an AI model’s “intelligence” and thus among an AI company’s most sensitive intellectual property. Serving a US model from European soil requires the weights to be stored in a European data center. The US government might be reluctant to allow that, e.g. if it doesn’t trust European cybersecurity standards to prevent model theft, or it might require that US models may only be served by US data center companies.
Inference is the process of deploying a trained AI model, in which the model generates an output for a given input.