When I say artificial intelligence doesn't kill people, LLMs do, the objection I hear most often is that this is a semantic quarrel: that "AI" and "large language model" are near enough as to be interchangeable, and insisting on the difference is pedantry.
Consider Boxtown, a neighborhood in Memphis, Tennessee, founded after emancipation by formerly enslaved people who built their first houses from discarded railroad boxcars. Its ZIP code is 95% black, with a median household income under $37,000 and a poverty rate above 31%. Residents live alongside an oil refinery, a steel mill, and chemical plants, and report asthma and cancer rates far above national averages, all of it predating what came next.
What came to Boxtown next was a supercomputer. In 2024, xAI built Colossus there, over 230,000 graphics processing units, in 122 days, and powered it with dozens of methane gas turbines for which it did not have permits. It was only caught breaking the law when the Southern Environmental Law Center flew thermal-imaging drones over the site and counted 35 of them, roughly enough capacity to power 280,000 homes. The Shelby County Health Department later issued Clean Air Act permits covering 15 of them, a decision environmental groups have appealed. In April, the NAACP sued xAI and a subsidiary over a second facility across the state line in Southaven, Mississippi, alleging 27 more unpermitted turbines emitting more than 1,700 tons of nitrogen oxides, 180 tons of fine particulate matter, and 19 tons of formaldehyde annually, which the complaint contends would make it the largest industrial source of nitrogen oxides in the greater Memphis area. The allegations have not been adjudicated.
Nothing about artificial intelligence required that. The difference between "AI" and "LLMs" is measured in gigawatts, aquifers, and the air over a neighborhood that has been absorbing other people's industry for a century. The evidence that settles it does not come from nostalgia about simpler computing. It comes from what the best AI laboratories in the world have published in the last few years.
Start there, because it reframes everything that follows.
This year a team including Yann LeCun, a Turing Award laureate and one of the architects of modern deep learning, published a world model trained end to end on a single GPU in a few hours. Fifteen million parameters. It plans roughly 48 times faster than systems built on foundation models, using around 200 times fewer tokens. LeCun's research program is, explicitly, an argument that scaling language models is the wrong road. And the man himself left Meta for good in November 2025, calling LLMs a "dead end."
DeepMind's GraphCast, published in Science, generates a 10-day global weather forecast on one tensor processing unit in under a minute and beats the European gold-standard supercomputer forecast in more than 90% of tested variables. FourCastNet, from the group that pioneered neural operators for physics, is estimated to use 12,000 times less energy than the numerical weather model it substitutes for. That is artificial intelligence that reduces the world's computational load.
A neural network built by DeepMind and the Swiss Plasma Center controls all 19 magnetic coils of a tokamak fusion reactor, in real time, running on the reactor's own control hardware. It was published in Nature. Confining fusion plasma is a harder real-time problem than holding a conversation, and it does not require a campus in Loudoun County, Virginia.
At the other end of the scale, TinyML models run on microcontrollers drawing under a milliwatt, in less than 512 kilobytes of memory. ABI Research projects 2.5 billion devices shipping with that capability by 2030: billions of AI deployments that will never contact a data center at all, running for years on a coin cell.
A commercial counterfactual is also emerging that receives almost no attention in this debate: firms building models the opposite way: a purpose-built architecture for a single task, trained on a customer's own data, deployed on the customer's own hardware, on-premise or air-gapped. Not a distillation of a frontier model, not open weights with a fine-tune applied, not a wrapper around a rented application programming interface. The model runs where the data already are. Every enterprise task served this way is an inference that never enters a hyperscale facility.
Now set that against what is actually being constructed.
U.S. data centers consumed roughly 4.4% of the nation's electricity in 2023 and are projected to reach 6.7%-12% by 2028, according to Lawrence Berkeley National Laboratory, which attributes the doubling of data center demand between 2017 and 2023 largely to AI servers. Put another way, one-ninth of all electricity consumed in the United States will go to powering these facilities by the end of the decade. The International Energy Agency has even more damning figures: Electricity use in accelerated servers, the GPU-dense racks that train and serve large models, is growing about 30% a year against 9% for conventional servers, and accounts for almost half the net increase in global data center consumption. The compute used to train frontier models has grown four to five times per year for over a decade.
But it's not the data center hyperscalers picking up the electricity tab, by and large. Rather, the bill arrives at U.S. households. In the PJM Interconnection, serving 65 million people, capacity prices rose from $28.92 per megawatt-day to $329.17 in two years. Data centers were responsible for 63% of one auction's increase, about $9.3 billion recovered from ratepayers, with the Natural Resources Defense Council projecting roughly $70 a month in additional household costs by 2028. In that same record auction, PJM fell 6,625 megawatts short of its own reliability target for the first time in the capacity market's history. A Bloomberg analysis of some 8,000 facilities found that about two-thirds of new U.S. data centers built or in development since 2022 sit in areas of high water stress. Coal plant retirements have been delayed or canceled across seven states. A decommissioned coal station in Homer City, Pennsylvania, is being rebuilt with seven gas turbines to serve data centers onsite. A study of 244 large facilities found an average of 1,688 workers during construction and 157 permanent jobs afterward. Quite a thin return for the 42 states offering the industry many billions in sales tax exemptions.
Put the two halves together, and the conclusion is unavoidable. A neural network can run a fusion reactor from a control rack. A world model can be trained on one GPU in an afternoon. A weather forecast that once required a supercomputer now takes a minute on a single chip. Meanwhile, an industry is reopening fossil plants and draining stressed aquifers, all to deliver a chatbot.
Simply put, that's not a technical necessity — it's a piratical business model. The largest feasible model, in the largest feasible building, rented to you by the token. It may well be a defensible commercial strategy — and if it is merely a speculative investment scheme, should it not be regulated as such? What it is not is a law of nature — that's not what "artificial intelligence" means.
Which is exactly why the vocabulary matters. "AI" is a research field 70 years old, encompassing the spam filter on your phone, the classifier helping a radiologist, the software driving a rover across Mars on a 133-megahertz processor, and the protein-structure model that won the 2024 Nobel Prize in Chemistry. A large language model is a 4-year-old consumer product category. When a press release announces an "AI data center," it borrows the credibility of the former to describe infrastructure built for the latter and converts a corporate capital decision into something that sounds like weather.
You cannot hold weather accountable. That is the point. "AI" has no address, signs no interconnection agreement, and files no water permit. Companies do those things, deliberately, for products, on schedules they control. What's even more galling is that no regulator requires them to say so under their own name. A company that is raising money from public investors must file an honest account of what it's actually building, while a company breaking ground on a data center does not.
None of this requires believing language models are worthless or that computation should be free of physical cost. It merely requires an honest ledger. Honest ledgers need correct nouns, and someone with the authority and wherewithal to check that what they're saying adds up.
The question of an honest ledger runs entirely unanswered all the way to the halls of Capitol Hill. No federal agency can require a frontier AI developer to disclose what it built, or what data it used to build it, or how much compute it used, or what running it might cost our energy grid. The data center hyperscalers and LLM model builders of Silicon Valley are more or less entirely unregulated in any meaningful way that is accountable to the public.
The residents standing up at county meetings from Virginia to Arizona are not confused about technology. They are no fools and understand the terminology bait-and-switch being played on them perfectly well. They use a dozen AI systems before breakfast without complaint. They are objecting to a particular building, drawing particular power, consuming particular water, for a particular product, built by a company whose name is on the rezoning application. They want meaningful public oversight of the purported magic bullets that they're being asked to pay for in their own monthly bills.
Every word of that objection is specific. Only the industry's words are vague, and it is vague in precisely one direction. Ask who that serves.