19 September 2026
This blog post is published in my personal capacity. The information and views set out in this post are mine and do not necessarily reflect the official opinion of the European Commission.
In short
Two pieces of news drove the EU sovereign AI discourse in the summer of 2026: the first was the essay Europe2031, which laid out a scenario where Europe falls behind on AI and becomes irrevocably dependent on the US. This was unbelievably well-timed, because later that same week the US government ordered Anthropic not to make their new Fable model available to non-US citizens, leading to it being unavailable for a period of time.
In Europe2031, one variable that is used to convey just how far Europe is behind the US in the AI race is compute. The piece discusses this in terms of gigawatts of installed data center capacity, which is a useful proxy, but not a metric that is easily understandable in terms of actual models. The goal of this post is to translate the numbers quoted into Europe’s current ability to both produce AI models and serve them to its population. Europe2031 can be interpreted as assuming that AI is a winner-take-all race where merely having a good model is no use if another actor has the best model. I find it important to note that while this is not guaranteed to be the case, it is still useful to think about where Europe’s ability to produce leading AI models stands.
The headline number we will use (which we will take directly from Europe2031, who have an analysis here) is that Europe (the Europe2031 data lists large data centers in the EEA, UK, and Switzerland) will have 2GW of data center capacity installed by the end of 2026. For comparison, the data lists US and Chinese capacities at 35GW and 5GW, respectively. As these numbers are difficult to interpret, we will try to calculate what 2GW of compute translates to in terms of the ability to train and provide sovereign AI at the scale of leading AI labs. One important caveat upfront: 2GW is the total facility power draw. Epoch AI, an organization that researches AI trends, estimates that GPUs (the type of chip used to develop & run most AI systems) account for only about 40% of power consumption in AI data centers, with the remainder going to cooling, networking, and other infrastructure, implying roughly 0.8GW actually available to run GPUs.
First, why are we measuring compute in watts? This is because the limiting factor in how many GPUs can be installed in a data center is governed by the amount of electricity the facility can safely consume while supplying the GPUs with power and keeping them at a safe temperature. While we can swap out the GPUs running inside a data center as new, more efficient, models come out, the amount of power a data center can draw is much harder to change. That also means that the question of what we can achieve with our 2GW is dependent on which types of GPUs are installed. For the purposes of this post, we will run the numbers using two scenarios: the first is that all data centers in Europe use the currently most common H100 GPUs1, the second is that we immediately upgrade all European compute to the higher-end B200 GPUs (the newest NVIDIA GPUs are currently only being rolled out to a few companies).
Before we move on to the actual numbers, we need to explain one more technical point: GPU usage for AI breaks down into two types: training and inference. During training, we “show” the model examples of the types of data we would like it to produce. Inference is the process by which models then generate output (code, articles, homework assignments). These two types of compute jobs differ in their demand profiles. To understand what Europe's current compute figure means, we want to know what kind of model we can currently train and how much AI output we can generate.
Most current frontier training demands are kept secret by the labs. Because calculating compute demand is reliant on many unpublished factors, we will use this estimate of the largest run tracked by Epoch AI at the time of writing to put European compute into context. Epoch AI calculates that training Grok 4 (trained in 2025) required 246 million H100-hours2 and $490 million to perform 5×1026 floating-point operations. We assume that Europe can split its compute budget the same way OpenAI did in 2024 and devote about 70% to R&D and 30% to inference (figures from other companies indicate that this estimate sits in a reasonable range across the industry). We can now estimate how training a 2025 frontier model would fit into our budget after accounting for the fact that only around 10% of all R&D compute appears to be spent on training the final model, with the rest going to experiments3.
Figure 1Training demand as a share of Europe’s training compute budget
We model the share of European training compute needed to train a 2025 frontier model.4 For the training compute demand needed for a 2026 frontier model, we assume the demand follows the 5x year-on-year trend observed by Epoch AI. We consider two scenarios, a fleet consisting only of H100s or only of B200s.
H100 fleet 7.2B H100-equivalent hours a year of available R&D compute
B200 fleet 11.3B H100-equivalent hours a year of available R&D compute
Source: Epoch AI, Grok 4 training resources and R&D vs training compute; Europe 2031 compute forecast.
The figure above treats Europe’s compute as being housed in one large data center, meaning that we can actually use all GPUs optimally with little latency. In practice, large training runs usually require GPUs to be in close proximity. As Europe’s compute is highly fragmented, we need to acknowledge that the numbers above represent an absolute (currently non-existent) best case. Still, we see that under these assumptions, Europe could train a 2025 frontier model while having compute devoted to inference. However, Epoch AI estimates that frontier model training demand grows about 5x year-on-year and the 10% figure for the proportion of R&D compute going to final training runs seems stable across a few companies. If these trends hold, we see that training the estimated 2026 frontier would exhaust Europe’s training compute budget even after upgrading our GPU fleet to B200 GPUs. Holding the frontier seems very unlikely even in the best of worlds for a pan-European AI effort even before considering that inference compute demand is also rapidly growing.
By now, it should be pretty clear that Europe’s compute is not anywhere close to enabling us to compete in terms of model size with the American frontier labs (we will examine below why this is unlikely to change). But, if we waved a magic wand and received a frontier-sized model today, could we at least serve it to European citizens without relying on external compute?
AI consumption is measured (and billed) in tokens. One token corresponds roughly to ¾ of a word. Exponential View estimates a global token demand of around 13 quadrillion tokens a month. We will take this as our baseline estimate and extrapolate a 15% share of consumption in Europe, giving us 1.95 quadrillion tokens a month5 consumed in Europe. We will make the optimistic assumption here that these tokens are demanded at a constant rate of 2.71 trillion per hour6. Calculating the compute demand for inference is quite sensitive to the specific models used, and no sufficient details are published for current frontier models. Epoch AI tried to estimate global inference capacity in May 2026 by using the Kimi K2.6, most capable open-source model at the time. Before we move onto the numbers, it’s important to understand that different types of AI uses have different token usage profiles. For example, using a model as a chatbot or for document summarization requires very different input/output token ratios (input tokens are the tokens given by the user, while output tokens are produced by the model) than using an AI for coding or agentic tasks. Epoch AI provides numbers for different inference profiles, and the table below shows us how their estimates translate onto European compute capacity and token demand. Note that these numbers are calculated assuming that we could use the newer fleet consisting only of B200 GPUs, which provides an upper bound on Europe’s actual current inference compute capacity.
Figure 2Serving capacity by query mix, against Europe’s 2 GW
We extrapolate European token serving capacity (tokens processed per hour) from Epoch AI’s numbers7. The table shows the share of total European compute and the share of assumed European inference compute (30% of the total) needed to meet inference demand (assuming inference demand is either all general use, all agentic coding, or all long-context, but not a mix of the three).
| Query mix (input : output) |
Total European inference capacity (tokens processed/hour) |
Share of total compute needed to meet demand |
Share of inference compute needed to meet demand |
|
|---|---|---|---|---|
Source: Luke Emberson and Jaime Sevilla (2026), “Is a compute crunch coming?”, Epoch AI; Exponential View, total tokens processed across all providers.
We can see that all three scenarios fit into our inference compute budget. However, there are some important caveats here, including the fact that demand at peak might be many times higher than the assumed 2.71 trillion tokens per hour. Additionally, Kimi K2.6 is likely a lot smaller than GPT-5.5 (see Note 1 here) and thus less computationally expensive compared to “true” frontier models. To make matters worse, Epoch AI estimates that demand at current prices could be rising 10x year-on-year, while the compute forecast in Europe2031 only has European compute rising to 5.2GW by 2027. Therefore, a similar story as with training emerges: Under optimistic assumptions, Europe has enough inference compute for now, but is very likely to be underpowered going into the near future.
Comparing the compute demand trend alongside the Europe2031 forecast for European compute buildout shows us that even after the announcement of AI Gigafactories by the European Commission (which are included in Europe2031's data), the compute gap between Europe and the AI frontier is growing rapidly.
Still, there’s no guarantee the current global compute trend can keep up: even the US compute buildout might not keep pace with demand given that the Wall Street Journal reports 60% of data center capacity set for completion in 2027 had not started construction as of May 2026.
Additionally, even if European leaders wanted to deploy the capital necessary to compete at the frontier, it seems doubtful that European voters would accept the high costs of AI hardware and electricity demands of data centers. In the US, AI is extremely unpopular among voters, which is not an encouraging sign for the future speed of data center construction on either side of the Atlantic.
Let me reiterate that the necessity of continuously scaling the largest AI models is a load-bearing (and controversial) assumption of this piece that you should not uncritically accept. Still, the numbers today indicate that if controlling frontier AI is essential to national security (which is not a given, but has been claimed by the US government in its decision to block Fable for export), Europe currently would have trouble competing. While it has the compute to serve its current population using cheap open-source models, owning both creation and distribution of frontier AI is far outside the reach of European compute capacity.
Still, there are question marks around the compute forecast for the US laid out in Europe2031. Current construction capacity is not keeping pace with announcements and the current economics are far from guaranteeing a positive outcome. Whether the investment pays off in monetary terms matters less in national security contexts and other winner-take-all scenarios, but for many use cases, cheap open-source models with significantly lower compute requirements than the current frontier might suffice. At the moment, subsidized AI costs allow all of us to essentially use the equivalent of a Ferrari for our daily commute where a Toyota would do. If cheaper models end up being the dominant category, we might see a shift away from gigantic datacenters, where the gap between Europe and the US would be a lot less significant. All those questions necessitate a well-informed dialogue about how European leaders see AI developing, how much the European public is willing to spend on supporting compute buildout in Europe, and how European trade policy should adjust to the threat of the US cutting us off from access. While these questions are far above my own paygrade, I hope to at least have put the numbers around our current resources into better context.
I thank Luke Emberson for the constructive feedback on this post.
None of the data centers shown on this map by Epoch AI have any data center in Europe housing any GPU newer than the H100 model as the dominant GPU at scale, and H100s were the most sold NVIDIA GPU model over the past few years, so we will run the numbers assuming we can use up the entirety of our 0.8GW budget using H100s.
The 246 million H100-hours figure comes directly from Epoch AI’s estimate. The 2026 frontier estimate applies Epoch AI’s observed 5×/yr growth to give ~1.23 billion H100-hours. For B200s, the same job requires roughly H100-hours ÷ 2.25 (due to compute improvements), each drawing 1,000W vs 700W, giving a factor of ~1.57× to convert from H100-hours.
While the rest of the R&D budget might not be fully attributable to the final run that produces the frontier model, it is nevertheless necessary to sustain competition at the frontier level.
We assume the model has the same compute requirements as the final training run for Grok 4 using Epoch AI’s numbers, assuming a 10% share of R&D compute devoted to the final run and a 70/30 split between R&D/inference compute.
Applying a 15% European share derived from country usage in the Anthropic Economic Index (September 2025) gives about 1.95 quadrillion tokens a month or 2.71 trillion tokens an hour.
Demand assumes tokens arrive uniformly across all hours.
As above, we assume 40% of data center power being used on the GPUs themselves and 1000W per B200 (Epoch gives their numbers in GB200 NVL72 racks, which we convert into individual GPUs for simplicity). This gives us:
Epoch AI’s inference model reports output tokens per second for a GB200 NVL72 rack serving Kimi K2.6. So, for the 8k:1k usage, Epoch lists 400,000 output tokens per second per rack, giving a total European inference capacity of 11,111 × 9 × 3,600 × 400,000 = 144 trillion tokens an hour.