Hook
Watch the order book, not the headline.
The headline that moved through the wires this week was a number: €13 billion. Google's largest single capital deployment in Europe to date. Four sites — an expansion at Hamina, new builds at Muhos, Vaala, and Kajaani. A projected €3.6 billion GDP contribution during construction, roughly 7,000 permanent jobs once operational, a 94 MW battery storage system bolted onto the Kajaani campus.
All of that is packaging. The order book — the part that actually prices risk — is a single structural line buried three paragraphs deep in the release: a 22-year power purchase agreement with Fortum for 50% of the output of the Loviisa nuclear plant. Ramp begins in 2028 at reduced volume. Full contracted share holds across 2030–2049.
A 22-year offtake on an existing European nuclear asset is not a data center announcement. It is a sovereign-duration trade dressed as capex. And it tells you something the AI-infrastructure bulls are not pricing: the marginal cost of intelligence has stopped being a semiconductor problem and become a power-plant problem. The companies that win the next compute cycle will not be the ones with the best chips. They will be the ones with the longest, firmest, cheapest contracted electrons.
Ruth Porat called it "Google's first nuclear energy deal outside of the United States" and "a really important cornerstone to everything that we are doing here." Read that sentence twice. Alphabet's President and Chief Investment Officer is describing a Finnish reactor as a cornerstone of a US technology company's operating model. That is a structural admission, not a press line.
The interesting question is not whether €13 billion is a lot of money. For Alphabet, it is a rounding error against a balance sheet that generates that in free cash flow inside a quarter. The interesting question is what happens to the asset-liability profile of a hyperscaler when it starts signing contracts longer than the depreciation schedule of the hardware that contract exists to power.
Context: Loviisa, Fortum, and the Balance Sheet Behind the Reactor
To price this deal you have to understand who Fortum is, and more importantly, what Fortum has survived.
Loviisa sits on the southern coast of Finland, roughly 100 kilometers east of Helsinki. Two VVER-440 V-213 pressurized water reactors, Soviet-origin design, commissioned in 1977 and 1981. Combined net output is just over 1 GW. Together they supply approximately 10% of Finland's electricity. Finland's grid is already roughly 99% renewable or nuclear by generation mix, and post-Olkiluoto 3 — the 1,600 MW EPR that finally entered regular commercial operation in 2023 after a fourteen-year construction saga — nuclear covers somewhere around a third of national demand.
So this is not a greenfield story. Finland does not need a new reactor to decarbonize. It needs something subtler: firm, dispatchable, price-certain baseload that survives a Dunkelflaute and does not spike to €500/MWh when the wind stops in the Baltic.
Loviisa's existing operating permits roll off within the current decade — the two units run on staggered licenses, the earlier of which expires in 2027, the later in 2030. Fortum has been pursuing extension through 2050. Without a capital program, the rational corporate decision for a 1970s-vintage VVER is retirement after 2030. That is the counterfactual against which everything else must be measured.
The €1 billion life-extension program embedded in Google's commitment is the deal's actual load-bearing wall. Everything else — the battery, the grid upgrades, the four sites — is downstream of a single binary: does Loviisa run past 2030, or does it not.
Now the balance sheet. Fortum is ~51% owned by the Finnish state. Between 2020 and 2022 it took one of the most spectacular write-downs in Nordic corporate history via its Uniper stake — a position that ended with Germany nationalizing the asset, Fortum booking losses in the multi-billion-euro range, and an €8 billion state bridge facility standing behind the group. The strategic response was a hard pivot back to what Fortum actually knows: Nordic hydro and Finnish nuclear.
A company emerging from that kind of drawdown does not casually underwrite a decade-scale capital program on a 1970s reactor. It needs someone to take the price risk off its book. That someone is a hyperscaler with a AA-rated balance sheet and an existential need for firm low-carbon power.
This is where the framing gets sloppy in most coverage. Commentators describe Google as "buying clean energy." It is not. It is buying certainty of supply at a contracted price for twenty-two years, and in exchange it is providing Fortum with the revenue visibility that makes a €1 billion capital program bankable. Fortum's own disclosure tells you the trade: the company expects its return on net assets (RONA) to rise by roughly 1.4 percentage points once the 50% capacity contract is fully realized.
Read that number carefully. Fortum is not saying the deal makes it rich. It is saying the deal makes its nuclear segment financeable at a lower cost of capital. A 1.4 percentage point RONA uplift on a de-risked asset base is the arithmetic signature of a risk transfer, not a revenue event.
Contrast this with the US nuclear deals that preceded it. Microsoft's arrangement with Constellation around Three Mile Island Unit 1 — rebranded Crane Clean Energy Center — is an 835 MW restart with a 20-year offtake, targeting 2028. Amazon's 1.92 GW arrangement with Talen Energy at Susquehanna is a co-location structure adjacent to an operating plant. Meta's 2025 RFP for 1–4 GW of nuclear capacity and its Clinton/Constellation offtake sit in the same family.
Every one of those is, structurally, a capacity-creation or co-location play. Google's Finland deal is a life-extension play. That is the analytical distinction that most of the tape missed.
US hyperscalers have been buying the marginal electron at an existing plant and then fighting about whether the plant's interconnection agreement permits it. FERC rejected the amended Susquehanna interconnection service agreement in late 2024 on a 2–1 vote, precisely because the co-location model threatens to shift transmission and reliability costs onto the broader rate base. That fight is unresolved.
Google sidestepped it. In Finland, the Loviisa electrons flow into the national grid. The contract is for capacity and attributes, not for physically isolated electrons behind a fence. There is no co-location controversy because there is no co-location. There is a PPA, a grid, and 22 years.
That is the real innovation in this deal: Google bought the outcome of the FERC fight without having to fight it. It paid for policy arbitrage.
Core: The Economics of Buying Twenty-Two Years of Certainty
Let me get into the mechanics, because the mechanics are where the alpha sits.
Additivity, not additionality
The first thing to test in any corporate PPA is whether the electrons are additional. Finnish commentators have already made the obvious objection: Loviisa already exists. Contracting 50% of its output does not create a single new megawatt-hour. It reallocates existing megawatt-hours from the merchant market to a single counterparty.
On a strict annual basis, that objection is correct. On a 22-year horizon, it collapses.
The counterfactual is not "Loviisa keeps running regardless." The counterfactual is "Loviisa's owners lack the contracted cash-flow certainty to commit €1 billion to a life-extension program on two VVER-440s, and the plant retires in the early 2030s." Under that counterfactual, Finland loses roughly 8 TWh per year of carbon-free baseload and the Nordic system loses its only non-hydro firm source outside Olkiluoto.
So the deal is not additional in the narrow sense. It is counterfactual-dependent — which is a stronger claim. I spent a chunk of my early career building liquidity-sustainability models, and the lesson that stuck was this: the number that matters is never the headline yield, it is the yield net of the mechanism that keeps the yield alive. A PPA that funds a life extension is worth more than a PPA that buys output from a plant that would run anyway, because it changes the terminal state of the system.
The uncomfortable corollary is that Google is now, functionally, a financier of Finnish national infrastructure. Alphabet shareholders are underwriting a portion of the durability of the Finnish power system. The externality runs the other way too — Finnish ratepayers capture the reliability benefit of a plant that a foreign technology company helped keep open.
That is not charity. It is a hedge purchased in the form of a positive externality. But let us be honest about the political economy: a partially state-owned utility signed a 22-year contract with a foreign hyperscaler, and the deal's bankability rests on the fact that the other side of the trade has a balance sheet capable of absorbing a decade of price volatility without blinking. That is industrial policy executed through a corporate PPA, with the state's balance sheet on one side and Big Tech's on the other.
What the contract actually is
The Fortum release describes a "50% capacity contract" alongside a memorandum of understanding covering new nuclear reactors and renewable energy exploration. The MoU is a press-release artifact. The capacity contract is the deal.
Structures in this family come in two flavors. A physical PPA transfers the actual electricity and the associated environmental attributes. A virtual PPA — essentially a contract-for-difference — settles financially against a reference price, with the buyer receiving a certificate stream and absorbing or capturing the difference between the strike and the spot.
The 1.4 percentage point RONA uplift points toward a structure that materially de-risks Fortum's cash flows rather than one that merely swaps price exposure. A pure CfD at a strike near forward curve would be RONA-neutral for the seller. A RONA-accretive structure implies Fortum is transferring a meaningful portion of its long-dated price risk to a counterparty with a lower cost of capital, and capturing the spread as a capital charge reduction.
Run the rough numbers. Loviisa produces on the order of 8 TWh annually at full output. Half of that is 4 TWh. Over 2028–2029 at reduced ramp, call it 1–2 TWh per year; over 2030–2049 at full contracted share, 4 TWh per year. Aggregate contracted volume lands somewhere in the 80–85 TWh range across the term.
Now amortize the €1 billion life-extension program across that volume. You get roughly €12 per megawatt-hour of embedded capital cost. Nordic day-ahead prices in recent years have cleared in a band that has touched negative territory during wind-surplus hours and spiked well above €100/MWh during winter scarcity events. The life-extension capex is not the deal's economic center of gravity. It is a modest adder on top of a marginal operating cost that, for a VVER-440 with sunk construction cost, sits in the €15–25/MWh range.
The €1 billion is not financing electricity generation. It is financing the option to keep selling electricity at all. That is a fundamentally different asset than a new-build, and it should be priced as such.
The duration mismatch nobody is underwriting
Here is where I diverge hard from the consensus read.
The consensus read is: Google bought cheap firm power, great for AI, bullish for the buildout. Fine. But look at the balance sheet mechanics.
Alphabet depreciates servers and network equipment over roughly six years, following the useful-life extension the company applied in the mid-2020s. The economic life of an AI training cluster, in revenue terms, is shorter than that — compute generations turn over fast enough that a large fraction of the value is captured in the first three years of deployment.
The power contract runs 22 years.
That is a maturity mismatch in the classic asset-liability sense, and it is the single most under-discussed feature of the entire hyperscaler nuclear wave. Google has locked in a long-dated fixed obligation whose value depends on the existence of a compute business that has not yet been built, running on hardware whose economic half-life is a fraction of the contract term.
Banks get regulated for less. A 22-year fixed-price offtake is, functionally, a long-dated bond issued by the buyer to the seller. Google has written a 22-year liability and booked it as an energy strategy.
Now, the counterargument is fair: Google is not obligated to run AI workloads to honor the PPA, and the contract is sized — 50% of one plant's output, roughly 4 TWh — against a company whose global electricity consumption runs in the tens of terawatt-hours annually. For Alphabet, this is not a concentration risk. It is a portfolio position.
But here is the thing about portfolio positions: they get marked to market in the terminal value. If the AI capex cycle disappoints in the early 2030s, Google does not face a liquidity event on this contract. It faces something more insidious — a stranded duration problem, where a fixed-price obligation outlives the revenue stream it was designed to protect, and shows up as an above-market cost line for a decade.
That is not a solvency risk. It is a multiple risk. And multiples are what justify the capex.
Firm CFE versus total CFE
There is a second-order point that gets lost. Google's stated goal is 24/7 carbon-free energy by 2030 — matching every hour of consumption with carbon-free generation on the same grid in the same hour. Most corporate buyers satisfy this with annual matching, which is a much weaker standard. Loviisa's output is already carbon-free. Contracting it does not increase Google's carbon-free energy percentage in Finland. Finland's grid is already ~99% clean by generation.
So what does it buy?
It buys firmness. It converts a variable, price-exposed clean-energy position into a fixed, contracted, weather-independent one. In a system where wind and solar dominate the marginal supply stack, the scarce commodity is not the megawatt-hour. It is the megawatt-hour that is guaranteed to exist when the wind is not blowing and the sun is not up.
That is the repricing happening under the surface of every one of these deals. Two decades of energy transition have made carbon-free electricity abundant and firm carbon-free electricity scarce. Fortum's RONA uplift is a direct measurement of that scarcity premium. And the 94 MW battery at Kajaani is Google's acknowledgment that even a nuclear-backed contract needs a buffer at the point of consumption.
(A 94 MW/4-hour buffer solves a different problem than baseload scarcity — it solves ramp and grid-service constraints at the interconnection. Note what it implies: at the site level, Google is building the demand-side flexibility that in a previous decade would have been provided by someone else's peaking plant, or, in the crypto markets I spend most of my time on, by a bitcoin miner's curtailment strategy.)
The constraint nobody priced: fuel and waste
Two non-obvious operating constraints sit under this deal, and neither appears in the press coverage.
Fuel. Loviisa's VVER-440 reactors were built around Russian-origin fuel assemblies. Fortum has spent the post-2022 period diversifying supply, qualifying Western-supplied fuel from Sweden and France. A life extension through 2050 converts a temporary fuel-diversification program into a permanent supply-chain requirement. That is not a technical footnote — it is a sovereignty condition. The entire premise of "sovereign AI" in Europe depends on a fuel cycle that does not route through a hostile jurisdiction for the next quarter century.
Waste. Finland is the only country in the world with a licensed deep geological repository — Posiva's Onkalo facility at Olkiluoto, designed to encapsulate and bury spent fuel from both Olkiluoto and Loviisa. A license extension to 2050 extends the spent-fuel inventory curve by twenty years and pushes encapsulation and disposal scheduling out accordingly. There is a real question, under-documented in the trade press, whether the disposal timeline is dimensioned against a 2050 reactor life or an earlier shutdown date.
Life-extension economics look clean until you price the back end. Anyone underwriting a reactor to 2050 is underwriting a waste-management schedule to 2050 and beyond. That cost is not in the €1 billion. That cost is a social contract.
The regulatory artifact angle
I have spent the last year navigating MiCA compliance for cross-border fund operations, and one observation transfers cleanly here: Europe writes the rulebook before the market arrives; the United States writes enforcement actions after the market has already positioned.
The EU's Electricity Market Design reform actively promoted long-term PPAs and two-way contracts-for-difference as instruments of price stability and decarbonization. The taxonomy classified nuclear as a transitional sustainable activity. The permitting reform packages shortened timelines for grid and generation projects. Google's Finland structure is not a loophole. It is a compliance artifact — the deal exists in its current form because Brussels and Helsinki built the rails first.
Now contrast with the FERC co-location fight in the US, where the regulatory question (who pays for transmission when a hyperscaler co-locates behind a nuclear plant's meter) was left ambiguous and is now being litigated after the capital has been committed.
The difference in execution risk between those two regulatory environments is not small. It is the difference between a signed contract and a contested docket.
Contrarian: What Happens When a Technology Company Starts Behaving Like a Utility
The consensus interpretation of Google's Finland play is straightforwardly bullish: hyperscalers are securing the energy bottleneck, AI buildout continues, nuclear is back, Finland wins.
I want to argue the opposite signal is embedded in the same fact pattern.
When the marginal cost of your core product is set by the marginal cost of firm power, you are no longer a technology company. You are a utility with a software wrapper — and utilities get utility multiples.
Consider what Alphabet has actually done here. It has signed a 22-year fixed-price offtake. It has committed to fund life extension on an existing generating asset. It has built battery storage at a campus to manage grid constraints. It has, per the release, folded grid improvements into its capital program. It is, in substance, deploying capital into regulated-asset-like infrastructure because that is where its marginal production constraint now lives.
Every one of those behaviors is a characteristic of an infrastructure investor, not a growth technology company. The capital intensity is moving in the direction of a utility. The revenue duration is moving in the direction of a utility. The regulatory exposure — nuclear licensing, grid interconnection, environmental permitting — is moving in the direction of a utility.
If the cash flows begin to look like a utility's, the terminal multiple will eventually converge toward a utility's. Markets are slow about this, but they are not wrong about it forever.
This is the same structural trap that swallowed the crypto mining sector. In 2021, miners looked like high-beta technology companies with extraordinary optionality on the hash price. By 2025, the survivors had quietly become energy traders with GPUs — signing long-dated power contracts, negotiating demand-response agreements, and selling capacity into grid-balancing markets. The multiple compressed from "growth technology" to "power arbitrage with a hardware cycle." Those miners who pivoted to HPC hosting for AI customers are now recognizable as data center REITs with a legacy business attached.
Google is not going down that path wholesale. But the direction of travel of its capital allocation is legible, and the market is not yet discounting the multiple consequences.
The additionality illusion, revisited
A second contrarian beat. Read the deal as a supply-preserving hedge rather than supply expansion.
Loviisa's 1 GW exists. The deal prevents it from disappearing. That is valuable, and it is also bounded. It does not add a single megawatt of new firm capacity to a European system that will need something on the order of several hundred gigawatts of new firm capacity over the next two decades if electrification of heat, transport, and industry proceeds at policy pace.
Life extension buys time. It does not buy growth. Every hyperscaler that signs a life-extension deal instead of a new-build deal is deferring the capacity problem, not solving it. Google's MoU with Fortum on new reactors is the honest acknowledgment of that gap — an MoU is not a reactor, and a first-of-a-kind European SMR or large reactor ordered today will not deliver electrons until the mid-2030s at the earliest.
The gap between now and then is where the interesting trades live. Firm power in Northern Europe is going to get structurally more expensive before it gets cheaper, because the marginal buyer — AI datacenters — has price-insensitive demand and a balance sheet that can outbid any industrial competitor for a scarce input.
That has consequences. Finnish pulp and paper, Finnish metals, Finnish district heating — all of them compete for the same electrons. When a hyperscaler signs a 22-year offtake at a contracted price, it removes that volume from the merchant market. Merchant-market participants, including Finnish industrial consumers, now face a thinner spot market with more volatility, not less. The counterfactual is that Loviisa might have closed entirely, which would be worse. But the honest framing is a trade-off between system-level scarcity and subscriber-level certainty, and Google is the subscriber.
Announcement versus execution
Third beat. The gap between the press release and the electricity.
The PPA begins in 2028 at reduced capacity. It scales to 50% of Loviisa's output across 2030–2049. Today is 2026. Between now and 2030, precisely nothing physically changes at Loviisa, at Hamina, at Kajaani, at Muhos, or at Vaala, beyond construction activity.
The regulatory path is not trivial. License extensions to 2050 require government approval, safety assessments, a qualified fuel supply chain, and a waste-management schedule that holds. Grid interconnection in northern Finland depends on Fingrid's transmission build-out, which is subject to the same permitting and capital constraints as every other European TSO.
I have spent enough time in the announcement-to-execution gap professionally to hold a default skepticism about headline numbers. In 2022, I watched funds price recovery probabilities on Celsius and BlockFi claims off press releases and bankruptcy dockets rather than off actual balance-sheet analysis. The ones who read the schedules — the actual asset schedules, not the press summary — were the ones who got paid. You underwrite the decade with the decade, not with the quarter.
€13 billion is a commitment. It is not a wire transfer. The four-site build is staged, the PPA is back-end loaded, and the €1 billion life-extension program is contingent on a licensing decision that has not yet been rendered. The correct way to read this is as a twenty-year option purchased with quarterly drawings.
The crypto read-across nobody is drawing
This is my home territory, so let me be direct.
Bitcoin mining and hyperscale AI compute are competing for exactly the same input: firm, cheap, low-carbon electricity at scale. Where they differ is optionality. A miner can curtail. A training cluster cannot. A miner's revenue is a floating commodity price; a hyperscaler's contract is a fixed 22-year obligation.
That asymmetry is about to reprice the entire energy-services market. As hyperscalers lock up the firm baseload, the residual demand for flexible, interruptible load becomes structurally more valuable — because the system needs someone to switch off when supply is tight, and the hyperscaler contractually will not.
Which means the surviving miners' curtailment capability, long dismissed as an operational nuisance, becomes the product. The miners who understood this — the ones that converted their sites into grid-service assets, signed demand-response agreements, and pivoted to HPC hosting — are already trading at infrastructure valuations while the pure hash-price miners trade at scrap value.
There is a second, quieter read-across. Google's 24/7 carbon-free energy standard requires hourly matching, which requires hourly granular certificates, which requires registry infrastructure that can track MWh-level provenance across a grid. That is a provenance problem. It is the same class of problem that blockchain infrastructure is genuinely, non-speculatively good at solving — and if the granular certificate market ever tokenizes at scale, the settlement layer will not be a hyperscaler's internal database.
I analyzed a comparable structural dynamic in 2021, when I modeled liquidity sustainability across DeFi pools and found that 85% of headline yield was coming from inflationary emissions rather than genuine fee generation. The tell was in the mechanism, not the number. Here, the tell is in the duration, not the euro figure. Liquidity is always a claim on someone else's balance sheet. In DeFi, it was a claim on a token emission schedule that had a terminus. In power, it is a claim on a plant that has a license that expires in 2030 unless someone pays to move it out to 2050.
Google is the someone. That is the whole trade.
Takeaway
The market will file this under "AI infrastructure" and move on. The correct filing is under duration.
What Google bought in Finland is not compute capacity. It is twenty-two years of price certainty on a scarce, firm, carbon-free input, purchased at the exact moment when that input's scarcity premium is being discovered by the entire Western industrial base simultaneously. Fortum got a 1.4 percentage point RONA uplift and a bankable capital program. Finland got €3.6 billion of construction GDP, 7,000 jobs, and a reactor that survives past 2030. Alphabet got a cornerstone.
Everyone won, which is precisely why the structure will be replicated — and why the second-order effects will be mispriced.
The thing to watch over the next four years is not the €13 billion. It is whether the life-extension license is granted, whether the fuel diversification program holds through qualification, and whether the next cohort of hyperscaler deals in Europe follows the same life-extension template or escalates to genuine new-build underwriting. Because the second category carries construction risk that no PPA can absorb, and the moment a hyperscaler starts underwriting greenfield nuclear, the balance sheet becomes a different animal entirely.
And watch the merchant market. If Finnish industrial consumers start paying more for power because a hyperscaler locked in half a reactor's output for twenty-two years, the political economy of "sovereign AI" in Europe changes shape. Sovereignty that transfers cost from shareholders to ratepayers has a shelf life measured in election cycles, not decades.
For positioning: firm power in constrained Northern European grids is the trade of this cycle, and the instruments that express it are not equity in hyperscalers. They are the contracted generators, the grid equipment suppliers, the fuel-cycle qualification specialists, and — yes — the flexible-load operators who can curtail when the firm load cannot.
The crowd is staring at the capex line. The order book is in the contract term.
