Google's €13bn Finnish AI buildout needs power in 2027, but the nuclear PPA only ramps to 50% from 2030 - is there a reliability gap?

On September 9, Google announced its single largest investment in Europe: at least €13 billion across 2027–2028 to build and expand AI data centers in Finland, wrapped in what reads like a tidy energy package. A 22-year power purchase agreement with Fortum covering up to 50% of the output of the 1,014-megawatt Loviisa nuclear plant. Another 629 MW of new onshore wind. A 94 MW battery at the Kajaani site. Headlines rounded the nuclear piece to "about 500 MW," which is the kind of number that makes a deal sound secured.

Read the delivery schedule and the impression gets weaker. That ~500 MW is the back end of a ramp, not what shows up in 2027. The PPA begins in 2028 at a reduced capacity and only climbs to contract for 50% of Loviisa—the ~500 MW figure—from 2030 through 2049. The data centers, meanwhile, are being built and switched on in 2027–2028. The two-year construction program is the front of the story; the firm nuclear block arrives up to four years later. That mismatch is the false narrative worth testing: the announcement's headline offtake does not match the timing, and what actually carries the 2027–28 buildout is intermittent wind, a small battery, and the existing Finnish grid.

What carries the load before 2030

Google's clean-energy claims rest on a real but limited toolkit for the gap years. The 629 MW of contracted onshore wind is new-to-the-grid capacity from Valorem and Suomen Hyötytuuli, but wind is variable: at a Finnish onshore capacity factor in the low 30s, 629 MW produces on the order of a couple of terawatt-hours a year, roughly the sustained output of only about 200 MW of baseload. Independent analysis of the announcement came to the same shortfall conclusion, noting the wind "would barely support 200 MW of load" and that the Hamina campus alone draws an estimated 60–70 MW minimum. The 94 MW battery, slated for late 2027 at Kajaani, helps balance the grid over minutes to hours, not over a cold, windless week—exactly the reliability problem Finnish winters pose.

Google, to its credit, doesn't pretend otherwise. Its own modeling via AFRY is blunt: keeping price and volatility impact fully mitigated needs on the order of 5–7 GW of new generation and flexibility for every 1 GW of new baseload demand. That is the honest engineering version of the balance sheet, and it says the 629 MW of wind and the 94 MW of battery are a start, not a match.

Two caveats stop this from being an outright "they can't power it" story. First, data centers do not hit full load on day one; demand ramps as servers are installed, so peak draw trails the construction start. Second, Google matches clean energy on a regional, hour-by-hour basis averaged over the year, not by wiring each rack to a specific plant—so an annual energy gap is not the same as a firmness gap. The residual risk is concentrated in those windless winter hours and in whether the grid north of Helsinki actually has the headroom to absorb a new downtown of baseload while the new interconnects and wind get built.

Why the schedule is vulnerable

The watchable risk is not nuclear—Loviisa already operates—but everything that must newly connect to reach the 2027–28 and 2030 milestones. On the nuclear side, the deal is a life-extension contract that pays Fortum to spend; Fortum itself says about 80% of the projects needed for the extension, representing hundreds of millions of euros of investment, still lack a final investment decision. That money determines whether the uprate (an added 38 MW targeted for 2028) and the ramp to 50% actually happen on time. The PPA is necessary for that revenue certainty but not sufficient for the execution.

On the wind and grid side, the exposure is permitting and interconnection. New Finnish capacity is being added against a backdrop where the ready pipeline is thin: Fortum cites only about 1.2 GW of renewables ready to build against a roughly 8 GW pipeline still sitting in the permitting phase. Grid connections for the northern sites require Fingrid connection agreements, and downstream of any approval sit the local opposition and court appeals that routinely stretch wind projects in Finland by years. Any of these slipping is what pushes the 2027–28 balance back onto the battery and the grid.

The evidence that confirms—or breaks—the timeline

Because the claim is a schedule claim, it has clean falsifiers. On the confirm side, the gap closes on time if each of these lands: Loviisa's PPA actually delivers partial nuclear power from 2028 and ramps toward 50% by 2030 as Fortum's delayed investment decisions are made; the 629 MW of wind reaches commercial operation on its slated dates rather than stalling in appeals; the 94 MW battery connects in late 2027 as announced; and the Fingrid connection agreements for Hamina, Kajaani, Muhos, and Vaala hold their queue dates.

On the falsify side, the reliability claim breaks the moment any of these publicly slips: a wind project pushed past 2027–28 by a permitting appeal or a re-permit; a Fingrid connection date for a northern site deferred, exposing grid constraints; Fortum's life-extension investment decisions for the remaining projects not materializing, freezing the Loviisa ramp at a small early capacity; or a sustained winter price spike or constraint episode showing the region cannot absorb the buildout in the cold, still-air hours the battery cannot cover.

None of this makes Google's announcement a sham. The nuclear block is structural and firm—it is the real anchor from 2030 on, and extending a plant that supplies 10% of Finland's electricity is a genuine act of supply security. But the "~500 MW" headline compresses a back-loaded schedule into a present-tense promise. What is actually secured in 2027–28 is a smaller, intermittent, and grid-dependent position, and the honest question for anyone sizing the story is not whether Google will power the buildout—it is whether, on this schedule, it can do so reliably in the Finnish winter before 2030. That is a question with observable answers, if you watch the right milestones.