Many countries invest in AI computing infrastructure by building GPU clusters. Computing resources on these clusters are typically made avaliable through an application-based process, and access is generally possible for academics and other researchers. The total amount of public GPU compute available in a given country varies quite a lot; we wouldn't necessarily expect a small country to invest as heavily in AI infrastructure as a much larger country would. Plotting each country's GDP against the total GPU FLOPS available across all of its public clusters can give a sense of each country's level of investment in AI research. I spent some time searching the web and massaging data (with some help from Claude and Gemini) to produce the plot shown below.

The data for this plot is available in this GitHub repo. Some of the exact numbers were hard to pin down, and while I ultimately tabulated ~80 clusters, it's entirely possible that I missed some—please feel free to open a PR or issue if anything looks off. FLOPS are measured here in terms of float16 or bfloat16 performance; while there is no perfect metric to measure compute capacity, I feel this is the most relevant and easily measurable across all of the diverse hardware in all of these clusters. The trend and patterns look pretty similar if you consider simply "# of GPUs” or “total VRAM”.

There is, unsurprisingly, some correlation between a country's GDP and the size of their public AI infrastructure. The plot also clearly demonstrates a set of “overachievers”, i.e. countries with an unusually large amount of GPU compute that seem to closely track their own fit line: Finland (thanks to the LUMI cluster), Switzerland (ALPS), Italy (Leonardo), Germany (JUPITER), and the USA (Aurora, El Capitan, Frontier, etc.). I won't comment on the countries that fall on the other side of the main fit line!

Interestingly, Canada currently falls right on the main fit line. I initially undertook this exercise to get a sense of how Canada would stack up after the $980M CAD SCIP funding is put to use. In the simple case where this funding is used to build out a single large cluster built from today's hardware, we might hope that we will get a new cluster with around 10,000 B200s (assuming total system cost of around $100,000 CAD per B200, which is not far off from other recent purchases that I'm aware of). This would bring online 22.5 EFLOPS of additional compute, bumping Canada up from the main trend line to sit comfortably among the overachievers. Canada is not alone in their efforts to expand compute capacity; a notable example is South Korea, which has claimed to be in the process of spinning up 50,000 GPUs for the public sector. Let's hope we can bring this compute online soon!

Update: After performing this analysis, I was alerted by Gül Sena that Epoch AI has already done the work of collecting a list of datacenters. Our lists mostly overlap, so I think the analysis provided here stands (perhaps more robustly).