Google's AI infrastructure bet: €13bn for Finland, $4.3bn for nuclear, and a satellite test for TPUs

Google is spending at a scale that makes the "AI capex bubble" conversation feel almost quaint. This week alone brought three concrete moves: a record €13bn ($15bn) infrastructure package in Finland, a $4.3bn nuclear-power agreement with Constellation Energy in the US, and the confirmed launch date for its first orbital data center test satellite. The through-line is simple: AI compute has become a power and physics problem as much as a software one.

Finland: nuclear power and three new data centres

The Finland package is Google's largest single investment in Europe, according to BBC News. It funds three new data centres — in Kajaani, Muhos and Vaala — plus an expansion of the existing Hamina campus, which Google converted from a former paper mill. Construction is scheduled for 2027–2028, and Google claims the buildout will support more than 37,000 jobs and add €3.6bn a year to Finland's GDP.

The energy side is the more interesting part. Google signed a 22-year contract with Finnish utility Fortum to buy up to 50% of the output from the Loviisa nuclear power plant. That's a long-term PPA structured to give Google predictable clean power while giving Fortum revenue certainty to keep the plant running. Finland's pitch — cool climate, low-carbon electricity, uncongested grid — also attracted TikTok, which announced a $1bn data centre in Kouvola earlier this week. The country is becoming the default European destination for AI infrastructure, and Google is paying a premium to lock in that position.

US: $4.3bn to squeeze more from existing reactors

Across the Atlantic, Google signed a 20-year deal with Constellation Energy to modernise 11 reactors at six sites in Illinois, Pennsylvania and New Jersey. The Verge reports the agreement represents more than $4.3bn of new Constellation investment and is expected to add 890 MW of capacity over five years — roughly the output of one large new reactor.

All six sites connect to the PJM Interconnection grid, which spans 13 states and includes Loudoun County, Virginia — the largest data centre hub in the US. PJM has been struggling to keep up with AI-driven demand; an independent market monitor filed a complaint with FERC last year arguing that PJM was allowing data centres to connect faster than the grid could reliably serve them. FERC recently sent PJM back to revise its auction rules to protect other ratepayers. Google's deal doesn't solve that bottleneck — it just secures a slice of the existing nuclear fleet before someone else does.

Space: testing Tensor TPUs in orbit

On October 1, Google will launch its first Suncatcher orbital data center test satellite. As Ars Technica notes, the experiment uses the same Tensor TPUs Google puts in ground servers, not radiation-hardened space hardware. The key questions are whether the chips survive launch vibration and radiation — which can flip bits and corrupt AI inference — and whether the cooling system can handle the heat.

Google's design uses thermal interface material to connect the TPUs to aluminium and copper heat pipes that radiate heat into space. But the system can only run for about 15 minutes at a time before the chips need to shut down and let the radiators catch up. Google will run Gemini models on the test, but the point isn't sustained compute; it's gathering data for future Suncatcher missions. The company still expects years of development before this becomes a product.

What it means for builders

For engineers and founders, the takeaway is that AI infrastructure is now a supply-chain story. Google is willing to sign 20-year nuclear contracts and build data centres in small Finnish towns because the marginal cost of being locked out of clean power is higher than the cost of committing early. The satellite test is a hedge — a bet that space-based inference might eventually matter for latency or energy reasons, but not something to build on yet. The Finland and Constellation deals, by contrast, are real capacity that will come online in the next few years. Watch the PJM regulatory fights and the Finnish grid connection queue: those are the real gating factors for Google's AI roadmap.