Sovereign Compute: Why Governments Are Now Building the AI Factories

For most of the current AI cycle, the interesting money moved between private companies: a lab raised a round, a hyperscaler leased it compute, a chipmaker booked the revenue. Over the past few weeks that pattern has started to widen. Governments and state-backed institutions are now writing the checks directly, and they are writing them for the same thing everyone else wants — accelerators, power, and floor space. The AI race is quietly acquiring a second front, and it is being fought over infrastructure rather than intelligence.

The clearest signal came out of Brussels. The European Commission and member states have laid out a plan for up to €20 billion of public-private investment into five AI compute "gigafactories", with Spain fielding a Tarragona site backed by Santander, ACS and Telefónica. The framing matters as much as the number. Europe has spent this cycle watching frontier compute concentrate inside a handful of American hyperscalers, and it has concluded that renting capacity from someone else's data center is not the same as having it. This is what we have been tracking as Europe's AI infrastructure awakening: a continent that led on AI regulation now trying to lead, or at least not trail, on the hardware underneath it.

Canada is making a smaller but structurally identical bet. A three-year sovereign AI GPU cloud contract will see BUZZ HPC deploy 2,304 Nvidia Grace Blackwell GPUs for Bell AI Fabric, running Cohere's foundation models on domestic soil. The scale is modest next to a hyperscaler build-out, but the word doing the work is "sovereign." The point is not to out-compute the United States. It is to guarantee that a Canadian bank, hospital or ministry can run a capable model without its data or its dependency sitting in another jurisdiction.

Japan is approaching the same problem from the research end. Through NEDO and METI, the government has committed to a five-year AI robot foundation model program with a ¥387.3 billion FY2026 budget, selecting a SoftBank-led Noetra effort and an AIST consortium to build it. Where Europe is buying data centers and Canada is buying a cloud, Japan is buying a capability — foundation models aimed squarely at robotics and physical automation, the domain where its industrial base still has an edge. Three governments, three entry points, one instinct: do not let the entire stack be owned abroad.

It would be easy to read all of this as pure statecraft, but the private capital is moving in the same direction and often to the same places. Consider Firmus Technologies, which appears on our tracker twice in a matter of weeks. Nvidia agreed to supply 170,000 GPUs under a revenue-sharing partnership as Firmus builds a 360 MW data center in Batam, Indonesia, targeting up to $30 billion in cloud revenue over six years. Separately, a Blackstone-led consortium extended a $10 billion debt facility to build up to 1.6 GW of Nvidia-backed AI data centers across Australia. When a single infrastructure operator can attract a chipmaker's supply commitment, a mega-fund's debt, and — in the sovereign cases — a government's balance sheet, you are no longer looking at a software market. You are looking at something closer to energy or telecoms.

That comparison is the useful one, and it comes with an uncomfortable implication. Utilities are slow. They are capital-intensive, permit-bound, and measured in gigawatts and years rather than model releases and weeks. The AI industry has trained itself to expect the opposite — that everything compounds monthly. Sovereign compute programs will collide with that expectation. A €20 billion European gigafactory plan has to survive procurement cycles, national politics and grid connections; by the time the concrete is poured, the frontier model it was meant to train may be two generations old. The bet these governments are making is that the underlying demand for compute is durable enough that late capacity is still valuable capacity. On the evidence of the past year, that looks like a reasonable bet — but it is a bet, not a certainty, and public money makes the downside political as well as financial.

There is also a reason the timing is not a coincidence. The same weeks that produced these buildout announcements produced a fresh round of anxiety over chip access, as Nvidia prepared limited H200 shipments to China under strict US conditions and export policy became, once again, front-page news. When the supply of frontier accelerators is a lever that one government can pull on another, "build your own capacity" stops being an industrial-policy talking point and starts being a hedge. Sovereign compute is, in part, insurance against a world where the chips you need can be switched off by a decision made somewhere else.

None of this means the state is about to out-innovate the labs. The most capable models will almost certainly keep coming from well-capitalized private companies, and a government-owned gigafactory is not a substitute for a research culture that can use it well. Europe's harder problem was never floor space; it was the talent and the product velocity to turn compute into frontier systems, and €20 billion does not automatically buy either. The realistic prize is more modest and still worth having: guaranteed access, domestic control of sensitive workloads, and a seat at the table when the rules for frontier compute get written.

What is genuinely new is the altitude at which the AI race is now being run. A year ago the scoreboard was benchmarks and funding rounds. Increasingly it is megawatts, GPU counts and who holds the deed to the building. You can watch that shift accumulate in real time on our deals tracker, where the compute and infrastructure category has quietly become one of the busiest. The models still get the headlines. But the machines are starting to get the sovereigns.