Schneider Electric and Foxconn are building the power-hungry assembly line that will extract billions from electricity ratepayers.
It is true, in the narrow sense in which press releases usually are, that Schneider’s partnership with Foxconn will produce modular power and cooling assemblies that make AI data centers more efficient. Closed-loop cooling and power-usage effectiveness ratios of 1.2—roughly four-fifths of consumed energy reaching the servers—are real engineering. The trouble is that efficiency at the module level does not determine consumption at the system level, and the system is being built to consume as much power as the grid will sell it. A lower PUE doesn’t just waste less electricity as heat; it lets the operator cram more servers behind the same grid connection. Total load goes up, not down. That is the Jevons paradox at data-center scale: making each unit cheaper to deploy doesn’t reduce the number deployed when four companies are racing one another and the capital markets are funding the race.
The Schneider-Foxconn deal accelerates this lockup. Modular power and cooling assemblies mean shorter construction timelines, and shorter timelines mean more megawatts consumed sooner. The efficiency gains don’t reduce total consumption; they reduce the time to first consumption. The partnership is the downstream hardware that makes the hyperscalers’ nuclear power purchase agreements profitable—and those agreements are the most legible single indicator that the AI buildout is a real-economy extraction event.
The last two years’ contracts make the appetite legible. Meta’s twenty-year PPA covers the entire 1.1-gigawatt output of Constellation Energy’s Clinton Clean Energy Center in Illinois—one company, one nuclear plant, two decades of dedicated baseload. Microsoft has locked up 835 megawatts through its Three Mile Island restart deal with Constellation, accelerated to 2027 with a billion-dollar federal loan. Amazon Web Services co-locates with Talen Energy’s Susquehanna nuclear plant, having purchased the adjacent Cumulus data center campus. Three contracts, three nuclear plants, the same four companies. These are not speculative ventures. They are the largest transfer of electrical generating capacity from the public grid to a handful of private companies since the deregulation wave of the 1990s.
This is the enshittification of the electrical grid—the four-stage sequence Doctorow mapped: a service made good for users, then turned against business customers, then value extracted from both for shareholders. The hyperscalers lock in cheap, reliable power with long-term contracts that often include favorable rate treatment and tax abatements. Utilities, eager for large industrial customers, approve the deals and pass the costs of grid upgrades to residential and small-business ratepayers. The regulatory constraints that once required public-interest review have been systematically weakened, leaving utility commissions to ratify what the companies have already negotiated. The people who pay the bills are not the people who negotiated the contracts.
When a hyperscaler signs a twenty-year PPA for a nuclear plant’s entire output, that generating capacity disappears from the public grid. The next user—the household, the hospital, the manufacturer—pays more for the remaining capacity or does without. No single regulatory proceeding systematically tots up the grid-capacity withdrawals of all hyperscaler PPAs across jurisdictions. The consequences are already material: constrained capacity, rising transmission costs, and planning cycles that cannot keep pace with demand.
Foxconn’s pivot to AI infrastructure warrants the skepticism its prior pivot earned. The company that negotiated a multibillion-dollar incentive package from Wisconsin for an LCD factory that became something considerably smaller and less transformative than what was announced is now repositioning as an AI data center module manufacturer. Its track record on delivering what was promised to the communities that subsidized it is a public document. Schneider Electric, for its part, is a sophisticated infrastructure firm; it knows precisely what its modules will be plugged into. The partnership scales a business line that has been a growth engine for years.
The supply-chain pattern extends beyond any single deal. Amazon and Corning signed a multibillion-dollar fiber-optic deal for AI data center interconnect, tilting Corning’s production toward the four companies purchasing cable at a scale no one else matches. STMicroelectronics raised its data-center revenue target last month and its shares rallied 8 percent on the same pivot: the semiconductor supply chain reorganizing around hyperscaler demand. The industrial base is committing productive capacity to the AI buildout because the money is there, and the money is there because the hyperscalers have decided to spend it.
The playbook is older than silicon. When my father’s rolling mill in Selkirk, Manitoba was bought by Gerdau of Brazil in 1995, the new owners modernized the plant, increased line speed, and cut a third of the workforce. The electricity that powered the mill came from Manitoba Hydro, a Crown corporation with a public-service mandate. The modernized mill used more power, and the surplus flowed to shareholders in Porto Alegre. The infrastructure was more efficient. The people who lost their jobs did not get the efficiency dividend. The AI data center buildout—modular, optimized, closed‑loop—is the same story with a newer vocabulary. The companies supplying the hardware, Schneider and Foxconn, will book the revenue. The hyperscalers will book the compute. The ratepayers will book the bill.
Production is scheduled to begin later this year. The public utility commission dockets where the rate cases will be filed are already being prepared. The deadlines for public comment will be announced on a Friday afternoon, in a format that requires Adobe Acrobat, with a comment period just long enough to satisfy the statute. The work, as always, is to notice.