Silicon Valley is strip-mining California’s water and power, and the governor is its partner.
It is true that the current generation of large language models produces genuinely useful outputs — protein-structure prediction, code generation, pattern recognition in medical imaging — and that AlphaFold, from Google’s DeepMind, has done real, Nobel-worthy science. The trouble is that every one of those outputs runs on physical infrastructure that consumes electricity and water at industrial scale, and the companies building that infrastructure have, with the governor’s active assistance, blocked the most basic requirement to disclose how much. The public case for the AI buildout is that it will cure disease, solve climate change, and make every worker more productive. The engineering reality is that it is, right now, the largest private seizure of public water and electrical infrastructure California has seen since the Gold Rush.
Mark Arax, writing in the Guardian this month, traces the through-line from the gold seekers of 1849 to the almond barons of the San Joaquin Valley to the datacenter farms rising in Santa Clara and San Jose. The architecture is the same at every stage. A scarce shared resource — gold-bearing gravel, groundwater, electrical capacity — is appropriated at scale by a small number of operators. The costs are socialized to the surrounding population. The profits are privatized. And the disclosure regime that would let the public see the cost-benefit ledger is defeated, every time, at the critical moment.
The central case study is not a tech company but a water district in Fresno County. In Pleasant Valley, a trust-fund heir named Jimmy Anderson — who flies his private jet between a country-club house in Fresno and a Pebble Beach estate — used his position as the district’s largest landowner to rewrite the local groundwater allocation formula. The old plan defined “irrigated acres” as land that was planted and watered year after year. Anderson’s new definition expanded the category to include 8,000 acres of mostly barren ground, 5,000 of which belong to him and his family. The result was that Anderson seized 1.5 billion gallons of annual groundwater rights from his neighbors. He is now selling those rights back to the very farmers he took them from, at $200 an acre-foot, and expects to clear close to a million dollars this year.
This is not a scandal in the sense that anyone will be prosecuted. It is a business plan, and it is entirely legal. California’s Sustainable Groundwater Management Act, passed in 2014, did not define “irrigated.” Anderson’s lawyers found the gap. The law, in the way of all such compromises, was designed to phase in restrictions slowly enough that the agricultural industry wouldn’t complain — which meant, in practice, a decade-long race to the bottom of the aquifer. By the time the phase-in ended, the biggest landowners had locked in their claims. The small farmers and the residents who drink from wells were left with the costs.
“Farming water,” Anderson’s rivals call it. The phrase names precisely what the datacenter operators are doing with electricity. They secure below-market power-purchase agreements, lock in long-term grid access, and extract value from a shared utility whose costs are borne by ratepayers who never asked for the load. They farm municipal water supplies under agreements whose terms they have, with the governor’s help, made it illegal to disclose. They farm tax incentives and streamlined permitting from eager municipalities — the datacenter equivalent of Anderson’s redefined irrigated acres, a reshaping of the rules to concentrate a shared resource in private hands.
The numbers are stark. In Santa Clara, home to 57 datacenters, the city’s own municipal utility — Silicon Valley Power — now supplies more than 60 per cent of its electricity to these facilities. Residents pay higher rates to subsidize the industrial load. San Jose, California’s third-most-populous city, is partnering with Pacific Gas and Electric in a $1.5 billion deal to triple electrical capacity over the next decade. PG&E is covering $670,000 in salary and benefits for two city-hall employees whose explicit job is to recruit more datacenters. The utility is, in effect, paying municipal staff to grow its own customer base — a circular arrangement that only looks like public policy if you don’t read the wiring diagram.
Here is the engineering-substance question the state does not want answered: how much water do these facilities consume? A single hyperscale datacenter can draw between one and five million gallons per day. California’s legislature passed a bill in September 2025 requiring operators to disclose and certify their water consumption as part of their local business licenses. Governor Newsom vetoed it, calling the measure too onerous. Newsom, whose political career has been funded by the tech industry and who recounts the unveiling of the first iPhone as a quasi-religious experience, has made a pattern of vetoing guardrails on AI while delivering speeches about “abundance.” The word is doing a lot of work: abundance for whom?
The regulatory architecture is gift-wrapping. The California Energy Commission exempts any datacenter requiring fewer than 100 megawatts of backup diesel power from its more demanding review. That threshold — enough electricity to light 75,000 houses — is not a serious environmental standard. Since 2011, the commission has fully vetted 15 datacenters under this process. Not one was denied.
In San Jose, the city approved a three-datacenter cluster on an 18-acre site in the Edenvale neighborhood — 547,000 square feet of computing equipment cooled with refrigerant and drinking water, backed by a power substation and three dozen diesel-fired generators — without conducting an environmental impact study. The site, on old hay ground where residents had wanted a park, sits surrounded by schools and $1.3 million houses, in the same part of town where Fairchild Semiconductor spent the late 1970s leaking hundreds of thousands of pounds of volatile organic compounds into the soil and groundwater. Fairchild landed on the EPA’s superfund list. Santa Clara County has 23 such toxic sites — more than any other county in the nation. The data centers’ diesel generators will add their own emissions to an air basin that already fails federal ozone standards.
This is what happens when the four forces that ordinarily constrain extractive industry are removed one by one. Competition does not apply — datacenter operators do not compete on water or energy efficiency because no regulatory framework forces them to. Regulation has been captured at the state level, where the governor vetoes disclosure and the commission rubber-stamps exemptions, and at the local level, where municipal governments are too busy recruiting more centers to ask what the existing ones are costing. Self-help is foreclosed — residents cannot build their own utility, cannot opt out of PG&E ratepayer obligations, and cannot compel disclosure without legislation the governor will kill. And labor, the fourth constraint, is irrelevant: once built, these facilities hardly employ anyone. There is no workforce inside the machine whose moral injury might slow the extraction down.
The pattern here is the one Cory Doctorow described as enshittification, applied not to a digital platform but to physical infrastructure. First the system is good to its users — the early internet was fast, cheap, and lightly regulated. Then it is good to its business customers — the datacenter operators get tax breaks, streamlined permitting, and below-market power. Then it extracts from both — ratepayer subsidies, water depletion, diesel emissions in residential neighborhoods — and delivers the surplus to shareholders. And if the AI bond covenants come due before the revenue materializes, the overbuilt infrastructure sits idle, and the ratepayer and taxpayer are left holding the bag — exactly as California’s overbuilt gas-plant fleet saddled ratepayers with billions in costs after the 2000–2001 energy crisis. The AI industry spent nearly $1 trillion on datacenters in 2025 alone, according to McKinsey, on the assumption that demand for AI inference will grow exponentially for years. Already, pending AI listings are binding U.S. retirement funds to the sector’s fortunes. If the bet collapses — if, in J.K. Galbraith’s term, the bezzle, the interval between a fraud’s commission and its discovery when the perpetrator has the gain and the victim does not yet feel the loss, closes — the costs will be socialized.
The refusal to disclose water use is itself the most important datum. A system whose operators will not let you measure its inputs is a system whose operators already know the measurement would change the political calculus. Kate Crawford made the point at book length in Atlas of AI: artificial intelligence is not a disembodied intelligence floating in the cloud but a logistical-extractive system built from natural resources, fuel, and human labor. The “cloud” is a building the size of an aircraft carrier with no windows and three dozen diesel generators out back.
And what does this industrial apparatus produce? San Jose’s deputy city manager, asked by Arax what crop comes off the fields of datacenters, paused and said they are good for business. It is worth being specific. They produce no taxable goods. They employ almost no one. The property taxes they pay are capped by Proposition 13. What they do produce — as Gemini, Google’s own AI, confirmed when Arax asked it — is a system that “is actively mining the human mind” and “atrophying our brains each time the software performs a thinking task for us.” They produce a compensation structure that drives San Francisco housing prices to levels only tech workers can afford, while the warehouse workers and delivery drivers who make the physical economy run commute from ever farther away. They produce, in short, the extraction of a shared resource — attention, memory, cognitive capacity — and nothing material in return.
Harold Innis argued that Canada’s political economy was shaped by the extraction and export of successive staples — fur, timber, wheat, minerals — each of which left the hinterland dependent on the metropole that processed and priced the staple. California’s story tracks the same arc across gold, oil, water, agricultural commodities, and now compute. Each staple enriched a small number of operators concentrated in San Francisco or its successor nodes. Each left the producing regions bearing the environmental and social costs. And each was enabled by a regulatory apparatus that, at the critical moment, chose not to measure the extraction or, having measured it, chose not to act. The AI datacenter is the latest staple, consuming land, water, and electricity, and yielding — nothing you can touch, smell, or taste.
There is a difference between a factory that makes something and a factory that extracts something. A datacenter is the latter.
At the Malakoff Diggins state historic park in the Sierra Nevada, you can stand at the lip of a crater blasted out by hydraulic mining in the 1870s. There is a mercury-tainted pond, a few ducks, some twisted pines, and the sheared-off walls of a mountain leaking mottled red iron. The state declared it a historic park. It is California’s habit to name its plunder sites as monuments after the plunder is done. That park is the bezzle, closed.
There is a disclosure bill due back in the legislature this session. The deadline for public comment is approaching. Deadlines are the only part of regulatory processes that the regulated actually respect.