National Grid pays Microsoft $1.75 billion to build a private power grid.
It is true that the cheque is being written by National Grid’s commercial arm, and not by the U.K. regulated distribution business whose wires carry power into eight million British homes. The trouble is that the distinction between a regulated utility’s commercial arm and its regulated rate base is precisely the distinction that makes the rest of the deal possible. The corporate parent is on the hook. The regulated entity is the source of the brand, the credit rating, the implicit guarantee. The public, which is the ultimate backstop of every regulated utility’s balance sheet, gets no vote on the deployment. The assurance that the investment “isn’t expected to affect [its] current five-year financial framework” is the standard corporate phrasing for: do not look too closely at the capital allocation.
The deal, announced in the first week of July, is for a 35 percent stake in Joulent, a U.S. company that The Wall Street Journal describes as developing “contracted power generation and high-voltage infrastructure for U.S. customers that require substantial amounts of power.” The customer, in the foundational case, is Microsoft. The foundational project — Project Kilby — will power a Microsoft-operated data centre campus in West Texas under a twenty-year agreement, is co-owned with Chevron, and is expected to begin operations in 2028. Final investment decision is to be made by the end of this year. Joulent, its investors say, is forecast to be free cash flow positive from the early 2030s. The $1.75 billion, the company says, will not affect its current five-year financial framework.
To read the press release at face value is to see a routine commercial investment by a U.K. grid operator looking to grow its U.S. business. To read the engineering substance is to see something else. The twenty-year agreement is the load-bearing financial structure of the whole arrangement, and the load it bears is the capital recovery horizon of the AI capex cycle. Microsoft is not just building a data centre; it is locking in a twenty-year power supply for one. Joulent, for its part, is not just building a power plant; it is amortising a twenty-year contracted-revenue stream against a generation-and-transmission capital cost that, in the absence of the contract, no power markets would clear at this scale on any commercially recognisable timeline. Both parties are locked in. The lock-in is the device that makes the deal financeable, and it is also the device that makes it unaccountable. The data centre will be built; the power will flow; the depreciation will be on someone’s books; and the externalities — water consumption, emissions, the stranded-asset risk if AI demand peaks before the twenty years are out — will be on someone else’s.
It is true that data centres require immense, continuous baseload power, and in the narrow sense in which engineers use the word, Joulent is solving a real constraint. The trouble is that “integrated power solutions” is the corporate euphemism for severing a critical piece of the physical layer from the public grid. What Joulent is actually building is a bespoke, tightly coupled generation-and-transmission node that bypasses the ordinary regulatory, ratepayer-sharing, and universal-service model of the public utility. The public grid operates on a pool-and-billing model: generators bid into a wholesale market, transmission operators move the electrons, and the cost is socialised across the rate base. Joulent’s architecture replaces that with a point-to-point physical and financial enclosure. The hyperscaler secures guaranteed capacity without absorbing the systemic cost of grid maintenance, resilience, or universal service obligations. It is the digital equivalent of a private toll road built across the public grid’s territory. It is the physical enclosure of the commons, wrapped in a twenty-year contract.
This is not an isolated transaction. As we noted when Chevron and Microsoft first announced the 2.7-gigawatt West Texas arrangement, the hyperscalers are no longer merely buying power; they are capturing the physical infrastructure that delivers it. The pattern accelerates. The same financialisation of essential infrastructure that just saw KKR pry a $4.2 billion renewable energy arm out of EDF now has a legacy regulated utility using its commercial arm to underwrite a private fiefdom for a technology monopolist. The capital goes where the rent is.
Kate Crawford’s Atlas of AI — the 2021 account of artificial intelligence as a logistical and extractive system rather than a cognitive technology — is the right place to read this. Crawford’s argument, in case it has not been encountered, is that AI is made from “vast amounts of natural resources, fuel, and human labour,” and that the abstraction the field performs in its public-facing register is the precise instrument by which the material infrastructure becomes invisible. National Grid’s $1.75 billion cheque is, among other things, a particularly legible entry in the ledger Crawford was keeping. The cheque, like the data centres it finances, like the power plants that will feed them, like the transmission lines that will connect them, is part of the physical plant of the AI economy. The fact that the cheque is denominated in the currency of a U.K. regulated utility is, in the end, a small detail. The unit of account is the kilowatt-hour. It always was.
Cory Doctorow’s chokepoint-capitalism frame, developed with Rebecca Giblin in the book of the same name, is the right lens for the architecture of the arrangement. A chokepoint, in Doctorow’s usage, is an intermediary that positions itself between two parties that cannot easily reach each other and extracts rent from the position. The AI data centre is itself a chokepoint, in the sense that Microsoft sits between the public and the AI compute, and the public is increasingly unable to opt out of the routing. The power provider is a chokepoint in the older and more traditional sense, in that Joulent-Chevron-National Grid sit between generation and the hyperscaler, and any power the hyperscaler uses has to pass through them. The grid, finally, is a chokepoint in the sense that the AI data centre cannot connect to the bulk power system without it, and the grid is the only public piece of the entire arrangement. The public, in other words, is the only party in the deal whose costs are externalised.
The engineering-substance aside, which I have been holding back, is what “integrated power solutions” actually means in the Joulent vehicle. The phrase is utility vocabulary for a particular corporate form: a single entity owns the generation, owns the high-voltage transmission, and signs a long-term contract with a single off-taker. The architecture is the vertically integrated utility of the early twentieth century, minus the public ownership of the wires and minus the regulatory standing of the public to set the rate. The customer is single; the off-taker is concentrated; the integration is bundling in service of a single project’s economics. This is a familiar corporate form — last seen in this corner of the energy economy in the build-own-operate-transfer deals of the 1990s, and before that in the rural electrification cooperatives of the 1930s, and before that in the holding companies whose abuses gave the country the Public Utility Holding Company Act of 1935. The Act, in case the historical detail is doing some work here, was passed in part because the holding-company form had been used to extract surplus from operating utilities and from ratepayers, and to leave the operating utilities holding the depreciation when the holding company moved on. The structural problem the Act addressed is the structural problem the Joulent vehicle is reconstituting. The Act was repealed in 1992. The structural problem did not go with it.
Chevron’s co-ownership is the part of the press release that has not received the attention it warrants. The oil major is repositioning itself as a power provider to data centres, which is to say that the same firms that, for the better part of a decade, told their shareholders that the energy transition was overhyped are now bringing their balance sheets to bear on the AI buildout. This is not, in the language of capital markets, a hedge. It is a re-platforming. The balance sheet that was built on finding and lifting hydrocarbons is now being redeployed to find and lift electrons for a different class of customer. The hydrocarbons are still being found and lifted, of course; they are now being sold to the same project that is electrifying them.
There is a Canadian-and-U.S. parallel worth drawing, because the geography inverts the politics and the inversion is instructive. Imagine a foreign utility — a French, or German, or U.K. one — investing in a Canadian data centre power project, co-owned with a Canadian oil major, with the electricity flowing to a U.S. hyperscaler. The transmission upgrades would be in Alberta or Saskatchewan. The rate-base impact would be in the foreign utility’s home jurisdiction. The Canadian public would bear the local environmental costs and would have no standing on the pricing. The political optics alone would be a scandal — front-page copy in The Globe and Mail within a week, a question in the House, an Order in Council response. The structure of the National Grid deal is the same. The geography is rearranged. The optics, in Texas, are different. The structure is the same.
The tradesman eye recognises the architecture. Vertical integration built to capture a single project’s economics, with the public system picking up what the project does not pay for. The mill model. The Selkirk story — Manitoba Rolling Mills acquired by Gerdau of Porto Alegre, several uncles laid off, father kept his job and watched a third of the line go — retold in kilowatt-hours. Extractive capital arrives, builds the infrastructure it needs, and leaves the public holding the depreciation. The mechanism is the same. The substance is different. The political economy is the same.
The deal is, finally, a bezzle. J.K. Galbraith coined the term in The Great Crash, 1929 for the gravity-defying interval when the money has been appropriated but the loss has not been discovered. Joulent’s implied equity valuation, derived from National Grid’s $1.75 billion for 35 percent, sits around the $5-billion mark at entry. It is a bezzle valuation: it depends on AI demand sustaining for twenty years, on the project operating as planned, on Microsoft’s AI business justifying the data centre capex, on the data centre remaining operational at planned scale. If any of these conditions fails — and the available indicators on AI demand sustainability are, to put it gently, mixed — the deal becomes a stranded asset. The corporate form is structured to shed the loss to whichever party is holding the bag when the bezzle collapses. The history of these structures, including the holding companies of the 1920s, suggests that the bag is generally held by the public. We are watching the bezzle pour into copper, steel, and switchgear, financed by a utility that still uses the word “Grid” in its name while systematically dismantling the premise of one.
The EDF-KKR deal we covered two days ago is the structural mirror image. There, a foreign utility was divesting its U.S. and Canadian renewable-energy arm to a private-equity buyer, on the grounds that the renewable portfolio did not justify the capital tied up in it. Here, a foreign utility is investing in a U.S. AI data centre power project, on the grounds that the AI buildout does. Both decisions are described, by the same kinds of executives at the same kinds of investor days, as routine capital allocation. Both decisions are recognisable, to the tradesman eye, as the same playbook at different stages. The capital goes where the rent is. The rent, in the AI economy, is in the data centre power supply. The capital will be there until the rent is not.
The public utility commissions on both sides of the Atlantic — FERC, the Texas Public Utility Commission, Ofgem — are evaluating this capital deployment against standard rate-of-return frameworks, treating National Grid’s stated aim of “capitalising on the country’s rapidly growing energy needs” as a routine commercial transaction. The regulated entities will offer the usual “stakeholder” submissions, arguing that private infrastructure investment relieves the rate base of the capital expenditure required to meet new demand. This is the manufactured-doubt playbook applied to grid governance: reframe a structural enclosure as a public relief. The data centre does not relieve the grid. It anchors itself to the most profitable node and leaves the rest of the network to absorb the systemic costs of resilience and transition.
There is a public consultation window — not on the National Grid investment, which is private, but on the open-access transmission requirements that should attach to the new infrastructure. The final investment decision is due by the end of the year. Deadlines are the only part of regulatory processes that the regulated actually respect, and the window to mandate open-access transmission requirements on this new infrastructure is closing with it. The data centres are private. The power is private. The capital is private. The grid, the water, the air, and the rate-base backstop are public. The work, in other words, is the cost accounting — to require that the AI infrastructure pay the full social cost of the social infrastructure it draws on, and to make that cost accounting visible to the regulators and the ratepayers whose balance sheet is the ultimate guarantee.