Google puts €13 billion into Finland — its biggest European bet yet
Two very different kinds of infrastructure news this morning: Google committing the largest single investment it has ever made in Europe, and OpenAI's newest model quietly running a physics lab overnight. One is about where the compute goes; the other is about what the compute now does unsupervised.
Google will spend €13 billion ($15.1 billion) on AI infrastructure in Finland over the next two years, its largest single investment in Europe. Announced in Helsinki alongside Prime Minister Petteri Orpo, the package funds three new data centers on top of the existing Hamina campus — a converted paper mill Google has run for 15 years — plus clean energy deals and €31 million for local communities in Hamina, Kajaani, Muhos, and Vaala. Google's own numbers claim the 2027–2028 construction phase adds an average €3.6 billion a year to Finnish GDP and supports more than 37,000 jobs, with about 7,000 jobs once the sites are running at wages it says sit 24% above Finland's median. Treat those as company estimates — they are the kind of figures that get repeated as facts and never audited.
The more interesting part is what Google bought alongside the land. It signed a 22-year power purchase agreement with Fortum covering up to half the output of the Loviisa nuclear plant, keeping it running to 2050, plus a letter of intent to study new nuclear capacity, new onshore wind, and a 94-megawatt battery system near Kajaani that comes online in late 2027. That is the real signal: in Europe, the binding constraint on AI buildout is no longer chips or even sites — it is firm, clean, politically survivable electricity. Google is not just building data centers; it is underwriting a grid, and attaching public saunas, fishing piers, and AI upskilling for 4,400 workers to make that politically durable. Every hyperscaler expanding in Europe will now be measured against this template.
OpenAI says GPT‑5.6 Sol, driven through Codex, can calibrate a superconducting qubit chip on its own. Beatriz Yankelevich, a graduate student in MIT's Engineering Quantum Systems Group, connected Codex to the lab software that coordinates experiments and gave it written instructions for each measurement. On an uncalibrated six-qubit chip, the agent chose measurement parameters, ran the hardware, read the results, and then either refined the run or passed the result to the next step — finding transition frequencies, calibrating control and readout pulses, and measuring coherence time without her watching each step. The group now routinely runs agents overnight while researchers work in the cleanroom; Yankelevich says she checks in from her phone and steers when something needs fixing.
The honest caveat is the most useful line in OpenAI's writeup: the agent struggled when signals were weak or noisy, taking longer to find workable parameters and sometimes needing an experienced researcher to intervene. That is the same boundary we keep hitting — agents handle well-defined workflows with clean feedback, and stall on ambiguity. What makes this different from another benchmark post is that the feedback loop is physical, not synthetic: the agent is not optimizing a score, it is turning microwave pulses into measurements on a chip cooled to near absolute zero. Routine characterization that used to take a researcher several days per chip now runs unsupervised, which is the actual productivity claim worth tracking.
Sam Altman broke his silence on the Navier-Stokes credit fight, and his account has a confession buried in it. Writing on X, Altman said OpenAI initially believed a rival team had also solved the problem and proposed a joint release, offered to let them publish first and take the prize, and even floated making their researcher lead author on a rewrite of OpenAI's proof — an offer he says was harder to extend to an Anthropic employee who would not coordinate. He also says the team "threatened us with unfounded accusations of plagiarism," that the two approaches now appear different, and — most tellingly — that OpenAI tried this at all because of internet rumors last week that Anthropic's models had solved a millennium problem.
That last admission reframes the whole episode. This was not a research program that landed on a famous open problem; it was a competitive response to a rumor about a rival, run in days, and the credit arrangement was improvised afterward. We covered the original claim and the mathematician's objection in OpenAI's AI solved Navier-Stokes — and a mathematician cries foul. Altman's version is one side of a dispute involving people who are not all talking publicly, and the underlying math still has to be checked by people with no stake in the outcome — which takes far longer than a weekend.
What to watch: whether Finland's €13 billion converts into concrete megawatts on schedule, and whether anyone outside OpenAI and the rival team adjudicates the Navier-Stokes proof on the math rather than the messaging.
Which story here actually changes something — the grid deal, the lab agent, or the credit fight? Tell us in the comments.
