Let me begin with a concession. It is true, in the narrow sense in which clean-energy consultants usually mean it, that the AI data-center buildout is doing what the Biden administration’s billions in subsidies could not: it is making wind, solar, and battery storage pencil out for utilities that had quietly shelved their decarbonization timelines. The Guardian piece that landed Friday catalogs the receipts — Nextpower posting 20 percent year-over-year growth, Google dropping the world’s largest grid-scale battery onto a Minnesota data center, DTE Energy in Michigan building a 330-megawatt battery system because it was the only way to meet Oracle’s timeline. The iShares Global Clean Energy ETF, which fell roughly 80 percent between late 2021 and early 2025, is up about 52 percent over the last year. The paradox the piece names — that the same AI buildout threatening to derail grid decarbonization is also the single largest private-sector driver of renewables deployment — is real, and worth taking seriously on its own terms.
The trouble — and here it is worth being precise about what “clean energy” means in the context of a load the size of a mid-sized city arriving on an eighteen-month timeline — is that the paradox dissolves the moment you stop treating “the grid” as a fixed object and start reading what the tech companies are actually doing rather than what their policy proxies say they want.
DTE Energy, the large Michigan electric utility, is fast-tracking grid connections for a $7 billion Stargate data-center project on southeast Michigan farmland. Rural residents rallied in December. They said the project could raise residential electricity rates and endanger the water supply. They were not wrong about the mechanism, and they are not the only ones. In Wisconsin, regulators with no renewable energy standard are building about fifteen wind or solar facilities to accommodate Microsoft and Oracle data centers, though those also include some natural gas. Nevada’s utility has already warned that AI data centers may force continued reliance on fossil fuels, and the pattern is repeating in state after state. This is not a clean energy transition. It is a fossil fuel expansion powered by a different name, and the name is the AI buildout’s electricity appetite.
What the Guardian piece documents, without quite naming the mechanism, is the vertical-integration play the hyperscalers have been telegraphing in plain sight since at least 2024. Douglas Jester, the clean-energy consultant the piece quotes, is exact: “It looks to me like they’re setting up to be vertically integrated to supply their own electricity, and they’ll drive a lot of development.” That is not a paradox. That is a bypass.
The numbers tell the story. Bloom Energy produces solid-oxide fuel-cell systems that can deploy in as little as ninety days — roughly one-forty-eighth of the twelve-year grid-interconnection queues the Guardian references — and its stock is up 1,338 percent over the last year. Oracle just signed with them. Bloom’s fuel cells emit carbon dioxide, however efficiently, because an electrochemical process that burns natural gas still produces carbon dioxide. Call that clean if you want. The carbon dioxide does not care what you call it. When gas is faster than solar, tech buys gas. Lucas Davis, the UC Berkeley energy economist, put it plainly: “I would say tech is desperate for electricity and oftentimes it’s going to whatever is the quickest — it could be the fuel cell, it could be natural gas turbines, or it could be solar and batteries, but the underlying demand is electricity.”
Desperation, not benevolence. The clean-energy boom the hyperscalers are driving is a byproduct of the speed constraint, and the speed constraint is producing two separate electrical systems. The first is the regulated grid, which will continue to carry the political and financial costs of decarbonization for everyone else — the coal and gas plants kept online to meet the hyperscalers’ load while the transmission upgrades are built, the rate increases spread across residential and small-business customers. The second is the private energy archipelago the hyperscalers are constructing behind their own meters: Google’s Minnesota battery, DTE’s Michigan battery, the Texas “off the grid” campus combining wind, solar, batteries, and gas. The archipelago’s assets are real, and they will displace some fossil generation that would otherwise have been built. But they are being built for the hyperscalers, on the hyperscalers’ timeline. They do not decarbonize the grid that serves the rest of us. They decarbonize the part of the electrical system the tech companies are quietly seceding from.
This is the structural move that the Guardian piece’s framing — “paradox” — obscures. A paradox is a contradiction that resists resolution. What is happening here is a deliberate unbundling, and the extraction mechanism is older than the cloud layer. Cory Doctorow’s enshittification framework names the digital instance of a more general phenomenon: a platform that is “good to its users” until it accumulates enough scale to route around the constraints that made it behave. The four forces that historically constrained extractive behavior — competition, regulation, self-help, and labor — apply to the electrical grid as surely as they apply to a social network. Competition among electricity buyers kept industrial loads tethered to the regulated rate base. Regulation — the obligation to serve, the integrated resource planning process — forced utilities to build generation and transmission for the common load. Self-help, in the form of on-site generation, was always technically available but economically marginal. And labor — the utility workforce, the state public-service commissions, the intervenor community — provided the political counterweight that kept the system from devolving into pure cost-allocation warfare.
The hyperscalers have now disabled all four constraints in sequence. Competition among electricity buyers has collapsed: a single data-center campus can represent a load larger than the entire residential demand of a mid-sized city, turning the utility’s relationship with the hyperscaler into a bilateral negotiation, not a market. Regulation has been neutralized by the speed mismatch — a twelve-year interconnection queue means the hyperscaler can build its own generation-and-storage complex before the utility can complete the regulatory process to serve it. Self-help is now the primary strategy, not a marginal supplement. And labor — the ratepayer advocates, the public-service commissions, the intervenors — is fighting a rearguard action at a zoning hearing in rural Michigan while the infrastructure decisions are being made in C-suites.
The ratepayer question is the tell. When the largest new electrical loads exit the regulated rate base, the fixed costs of the grid — transmission, distribution, the aging gas and coal plants the utilities are keeping online — are spread across a smaller pool of remaining customers. The hyperscalers’ energy archipelago lowers their costs and speeds their timelines. The residential ratepayer, the small business, the school district — they pay for the grid the hyperscalers no longer need. The clean-energy buildout the data centers are financing is almost entirely utility-scale — big batteries, big solar farms, big fuel cells — and does little for the homeowner who wants panels on her roof. That segment of the industry remains sluggish, a casualty of the same high interest rates and policy whiplash the hyperscalers are routing around.
The second Trump administration’s hostility to clean energy accelerated this divergence. The Guardian notes that the administration “canceled the government programs that had helped wind, solar, and electric vehicles.” The policy vacuum — the removal of federal carrots and the signal that fossil-friendly regulation would persist — did not stop the hyperscalers; it pushed them harder into the bypass. With no reliable public-sector path for grid decarbonization, private on-site generation became not just a speed play but a hedge against policy uncertainty. The hyperscalers didn’t wait for Washington to come around; they built their own energy future, fast enough to leave the regulated grid and its ratepayers absorbing the policy fallout alone.
The net effect on total U.S. carbon emissions depends on whether the archipelago’s clean-energy buildout displaces more fossil generation than the grid’s delayed transition adds. The Guardian does not hazard a number, and I will not confabulate one. But the direction of the asymmetry is legible even without the exact tonnage: the archipelago decarbonizes the part of the system the tech companies control and pay for; the grid burns more gas and coal to serve the rest of us, and the rest of us pay for it. The two-thirds of planned U.S. AI data centers built on drought-hit land add water to the ratepayer’s ledger alongside the electricity bill.
A BlackRock portfolio manager overseeing the firm’s flagship sustainability funds told Bloomberg that the sector is “well prepared to withstand a potential AI bust,” and that sustainable-energy equities “could stand to even further benefit as U.S. rates come down.” Translate that from finance to English: the clean-energy assets the hyperscalers are financing — the batteries, the solar farms, the fuel cells — will exist regardless of whether the AI models they power are profitable. The climate benefit, such as it is, is a byproduct of a capital-allocation decision, not a climate decision. That does not make the benefit unreal. It makes it contingent — contingent on the AI buildout continuing long enough to finance the assets, contingent on the assets being operated rather than stranded, contingent on the regulatory environment not shifting in ways that re-tether the hyperscalers to the grid they are currently exiting.
The structural story does not end with the equity analyst’s reassurance. It ends with the ratepayer in rural Michigan, who is being told that a $7 billion data center will be built on farmland, that the local utility is fast-tracking it, that residential rates may rise, that the water supply may be at risk — and that, somewhere in Minnesota and Texas, the same companies building the data center are also building very impressive batteries that will not lower her bill by a single cent.
The public consultation in Michigan is presumably still open. The intervenor dockets at the state public-service commission are presumably still accepting filings. The hyperscalers’ energy archipelago is being built on specific permits, specific zoning decisions, specific utility-rate cases, and specific water allocations — all of them public records, all of them contestable. The death of the regulated-grid model is not an inevitability. It is a series of decisions made by named people in docketed proceedings. The decisions can be different decisions.
The work is to be done.