The “Sandwich” Economy: Why AI’s Astronomical Utility Bills Are Saving Human Jobs

For three years, the dominant boardroom narrative in corporate America held that large language models would soon reach human-level competency across cognitive work, letting firms slash headcounts and replace salaried labor with low-cost digital minds. That narrative assumed intelligence would keep getting cheaper at software speed. On U.S. soil, it hasn't worked out that way. Electricity, liquid cooling, grid interconnection queues, transformers, and Blackwell-class silicon all impose hard physical costs that don't fall the way token prices do — and by 2026, those costs are visible everywhere from Dominion Energy's rate filings in Richmond to Uber's internal budget memos.

Nvidia's own VP of applied deep learning has said publicly that for his team, compute now costs more than the employees who use it. Uber's engineering organization burned through its full annual AI budget in four months. MIT researchers found that 95% of enterprise generative-AI pilots show no measurable return, and a separate MIT study found AI automation cheaper than a human in only 23% of vision-based workplace tasks. This piece goes inside the physical cost stack behind every American AI query — from a Loudoun County substation to a liquid-cooled rack in Abilene, Texas — and argues that the resulting "Sandwich Economy," in which a worker's caloric efficiency turns out to be an underrated competitive asset, is quietly capping how far full automation can go in the U.S. labor market.

For the past few years, a singular story has dominated American boardrooms and headlines: artificial intelligence will soon achieve human-level competency across cognitive tasks, letting corporations slash headcounts and replace expensive human labor with efficient, low-cost digital minds. It's a clean story. It's also the story that drove trillions of dollars of U.S. market capitalization into a handful of chipmakers, cloud platforms, and model labs, and it's the story every CFO has heard in some form from a vendor over the last eighteen months.

It rests on an assumption that turns out to be quietly false: that intelligence, once digitized, scales at close to zero marginal cost the way software traditionally has. Software doesn't need a dedicated power substation. A frontier LLM cluster does — and increasingly, it needs its own natural-gas turbines, its own water rights, and its own seat at the negotiating table with a regional grid operator. As AI deployment in the U.S. has scaled from chatbot demos to always-on enterprise agents, the physical resource stack behind every query — electricity, heat, water, and scarce silicon — has turned out to be the load-bearing constraint that the "replacement" narrative never priced in. Nowhere is that clearer than in the country that built most of this infrastructure first.

Every query processed by a large language model triggers a physical chain of resource consumption that has no equivalent in ordinary software: prompt in, a GPU cluster somewhere in Texas or Virginia spins up, power draw spikes, heat is generated, and liquid or air cooling has to remove that heat before the next query lands. Multiply that chain by billions of daily queries and millions of always-on agentic loops, and "software" starts behaving like heavy industry — closer to an aluminum smelter than to a SaaS product.

None of the five stages in that chain is optional, and none of them is getting meaningfully cheaper at the rate token prices are. That mismatch — falling sticker prices for intelligence, rising physical costs to actually deliver it on American grids — is the tension this article is about.

AI's appetite for electricity is stretching American grids harder than almost anyone forecasting five years ago expected. The Department of Energy's Lawrence Berkeley National Laboratory puts U.S. data-center electricity use at 4.4% of total national consumption in 2023, up from a little under 2% for most of the prior decade — and projects it will climb to somewhere between 6.7% and 12% by 2028. In absolute terms, that's a jump from 176 terawatt-hours in 2023 to a projected 325–580 terawatt-hours in 2028, a swing large enough that LBNL researchers themselves describe the underlying growth curve as one they've never had to model before.

Nowhere is that surge more concentrated than in Data Center Alley — the cluster of counties in Northern Virginia that PJM Interconnection, the grid operator for 13 mid-Atlantic and Midwestern states plus Washington, D.C., now calls "the largest concentration of data centers in the world." PJM's Dominion transmission zone, which covers most of Virginia, posted a summer peak load of 23,905 megawatts in 2025 — up 23% since 2019 — and a winter peak of 25,413 megawatts, up 45%. PJM expects that zone to grow faster than any other in its footprint through 2030, almost entirely because of data-center interconnection requests. Dominion's own resource plan calls for 27 gigawatts of new generation by 2039 just to keep pace, and in November 2025 Virginia's utility regulator approved a new rate class specifically for AI data centers, requiring them to pre-pay for the bulk of the transmission and generation capacity they request before it gets built.

Those costs don't stay contained to the data centers that caused them. PJM's capacity auction — the mechanism that pays power plants to be available on the grid's worst days — cleared at $28.92 per megawatt-day for the 2024/25 delivery year. For 2025/26 it jumped to $269.92, an 833% increase; the 2026/27 auction hit the federally approved price cap of $329.17. Monitoring Analytics, PJM's independent market monitor, attributes 63% of that increase directly to data-center load and estimates it adds roughly $9.3 billion a year in costs that get passed through to ordinary ratepayers. The Natural Resources Defense Council projects the average household across PJM's 13-state, 67-million-person footprint could be paying about $70 more per month by 2028 versus a no-AI-boom baseline. Dominion Energy filed for its first Virginia base-rate increase since 1992 this year, worth roughly $8.51 a month to a typical household. This is the part of the "AI is cheap" story that shows up on a residential bill in Fairfax or Prince William County before it shows up in a shareholder letter.

It helps to walk onto the data center floor itself, because the abstraction of "the cloud" breaks down fast once you're standing in front of the hardware. A traditional enterprise server rack draws somewhere around 7–12 kilowatts — about what a large central air conditioner pulls. A rack of Nvidia's GB200 NVL72 system, the workhorse of the current U.S. AI buildout, draws roughly 120 to 132 kilowatts: 72 Blackwell GPUs and 36 Grace CPUs wired into a single shared-memory unit, weighing about 1.36 metric tons, cooled almost entirely by pumped liquid rather than fans because air simply can't move that much heat out of a footprint that small. One rack. Ten to fifteen times the power draw of the rack it replaced. Multiply that by the tens of thousands of racks going into a single American campus and the numbers stop looking like an IT budget line and start looking like a power-plant interconnection application — because increasingly, that's exactly what they are.

The clearest illustration is sitting outside Abilene, Texas. The flagship Stargate campus — built by Crusoe for OpenAI and Oracle on the ERCOT grid, chosen partly for its proximity to cheap West Texas wind — is contracted for roughly 1.2 gigawatts of capacity and, per Oracle co-founder Larry Ellison, is being built out to house more than 450,000 Nvidia GB200 GPUs, enough draw for close to a million American homes. It's one node in a much larger buildout: OpenAI, Oracle, and SoftBank have announced Stargate sites in Shackelford County and Milam County, Texas; Doña Ana County, New Mexico; Lordstown, Ohio; and a Wisconsin campus, adding up to nearly 7 gigawatts of announced U.S. capacity and more than $400 billion in committed investment. A few hundred miles east, in Memphis, Tennessee, xAI's Colossus supercomputer runs on roughly 200,000 Nvidia GPUs today with a stated ambition to reach a million — enough that xAI has leased mobile gas turbines to bridge the gap while it waits on utility-scale power. None of these are hypothetical facilities. They're under construction, drawing permitted power, and in several cases already running training workloads.

The economics scale with the hardware. Industry cost trackers put the fully built-out capital cost of a GB200 NVL72 facility at roughly $1.5–2.5 million per megawatt of IT load once power, cooling, and building shell are included — 50–100% more per megawatt than a traditional air-cooled hall, even after accounting for the fact that a liquid-cooled rack packs five times more compute into the same floor space. That premium is exactly the kind of capital that has to be depreciated into the price of every American AI query, and it's part of why the "AI is simply cheaper than a person" argument has to be tested campus by campus rather than assumed.

Compute-heavy hardware generates enormous thermal loads, and historically, cooling systems have accounted for up to 40% of a data center's total electricity draw. A traditional server rack dissipates 5–15 kilowatts of heat; a GB200 NVL72 rack dissipates 120-plus. To avoid thermal throttling, U.S. hyperscale operators are racing to retrofit facilities with direct-to-chip liquid cooling and to lock in water-access agreements that amount to a second, less-discussed utility bill sitting directly on top of the electricity one.

Northern Virginia again supplies the clearest numbers, because it's the region that's been tracking them longest. The Environmental and Energy Study Institute estimates that all data centers across Fairfax, Loudoun, Prince William, and Fauquier counties consumed close to 2 billion gallons of water in 2023 — a 63% jump from 2019. Loudoun County alone, home to roughly 200 operational facilities, drew about 900 million gallons that year, enough that Loudoun Water now leans on potable municipal supply rather than reclaimed water to keep up. Nationally, EPA and LBNL data put direct U.S. data-center water consumption at 17.4 billion gallons in 2023, with projections of 38 to 73 billion gallons by 2028. A single 100-megawatt hyperscale campus using evaporative cooling — a fairly ordinary size for a modern AI training facility — can consume 1 to 5 million gallons of water a day at peak summer load, roughly the daily draw of a town of 10,000 to 50,000 people, with usage spiking two to three times higher during heat waves, which is precisely when the surrounding grid and water utility are also under the most local stress.

That's pushed the largest U.S. operators toward closed-loop and direct-to-chip liquid cooling, which cuts water withdrawal sharply in exchange for a higher electricity bill and a steeper up-front capital cost — a trade-off, not a free efficiency gain. Even after those retrofits, the underlying cost of a unit of American AI compute stays stubbornly high, no matter how cheap the model licensing looks on a rate card.

The capital required to secure frontier compute in the U.S. remains extraordinary, and almost all of it is domestic construction spending. Morgan Stanley's tracking put Big Tech's 2026 AI-related capital expenditure at roughly $740 billion — Amazon, Microsoft, Google, and Meta alone account for most of that — and Gartner projects worldwide IT spending will reach $6.31 trillion in 2026, up more than 13% year over year, with U.S. AI infrastructure as the single largest driver of the increase. That's capital that has to be depreciated somewhere, and increasingly it's depreciated straight into the per-query cost of running an American AI agent instead of a human employee's paycheck.

The clearest admission of this came not from a critic but from inside the supply chain. Nvidia's own vice president of applied deep learning, Bryan Catanzaro, told Axios in April 2026 that for his team, the cost of compute now runs far beyond the cost of the employees using it — GPUs, electricity, and cloud spend combined have overtaken payroll for the same group. When the company selling the picks says the picks now cost more than the miners, the "AI is simply cheaper" argument stops being a given and starts being a claim that has to be tested case by case, project by project, across every U.S. company weighing the same trade.

When U.S. companies model the ROI of replacing an employee with AI, they frequently fall into what researchers call the Turing Trap — assuming automation is cheaper simply because it's faster. The empirical record increasingly disagrees. A widely cited MIT CSAIL analysis of the technical and financial requirements for matching human performance found AI automation economically viable in only 23% of vision-based workplace tasks; in the remaining 77%, the combined cost of acquiring, adapting, and maintaining the AI system exceeded a competent employee's wage for the same job — the same conclusion, running through a different model, that a Loudoun County utility planner reaches when pricing out a new substation.

MIT's Project NANDA reached a parallel conclusion at the organizational level: despite an estimated $30–40 billion in enterprise generative-AI spending, roughly 95% of pilots have produced no measurable P&L impact, with just 5% of integrated deployments capturing meaningful value. The gap wasn't primarily about model quality — it was about the hidden operating cost of getting an agent from "mostly working" to "reliably unsupervised."

"Token anxiety" compounds the problem. Salaried employees are a fixed, predictable cost. AI agents are billed by the token, and in always-on enterprise deployments — agents that continuously query databases, summarize sprawling message histories, and re-validate their own code — token consumption can multiply unpredictably. Uber's engineering organization reportedly exhausted its entire 2026 AI budget in four months because of exactly this dynamic, turning what was budgeted as a fixed software cost into something closer to a variable utility bill that spikes with usage.

Set against 132 kilowatts of Blackwell silicon, the American worker at a desk in Richmond or Columbus looks almost comically efficient. A person can perform complex, unstructured cognitive work for an eight-hour shift powered by roughly 2,000–2,500 food calories — the energy equivalent of a couple of sandwiches, a bag of chips, and a coffee — which works out to a sustained power draw of well under 150 watts for the whole body, brain included. Training and running a neural network to approximate that same cognitive range at scale requires orders of magnitude more physical energy once GPUs, cooling, and networking are all counted in.

The gap shows up just as starkly in physical labor. Reporting on the coming wave of humanoid workplace robots estimates that a robot performing a shift of manual tasks comparable to a human's typically consumes on the order of 10–15 kilowatt-hours over sixteen hours of operation — every step, grasp, and glance requiring sensors feeding a power-hungry compute stack — versus roughly four liters of water and 2,000 calories for a human doing similar work. It is not a knock against the technology; it is simply a reminder that biological cognition and motor control evolved under a brutal energy budget that silicon has not yet had to meet.

Faced with mounting inference bills, capacity-auction sticker shock, and hardware capex, forward-thinking U.S. enterprises are pivoting away from full automation toward augmentation. Rather than building fully autonomous agents to replace employees outright, organizations are turning existing teams into "super-users" — pairing smaller, fine-tuned, often open-weight models (meaningfully cheaper to run than frontier proprietary systems, and easier to host on-premise away from the priciest grid regions) with human oversight for the highest-stakes 20% of a task.

That last 20% is where the real cost curve lives. Getting an AI system to handle the first 80% of a task can cost pennies; eliminating the final slice of errors needed to remove a human from the loop entirely tends to require an exponential increase in compute, fine-tuning, and monitoring — the same kind of exponential cost curve, incidentally, that shows up when a utility tries to add the last few gigawatts of firm capacity to a grid that's already tight. Keeping a human in the loop remains, for now, the cheapest error-correction mechanism available — which is precisely the relationship the calculator to the right is built to make visible.

Figures reflect U.S. facilities and U.S. grid regions (PJM, ERCOT, TVA) as of mid-2026; see References for source-by-source detail.

As the physical realities of American power grids, cooling water, and silicon scarcity collide with corporate balance sheets, the "AI will replace everyone" narrative is beginning to fracture — not because the models stopped improving, but because the electrons, water, and copper required to run them in Virginia, Texas, and Tennessee refuse to get cheaper on command. Rather than a future of empty U.S. offices, the more likely near-term economy belongs to organizations that master a budget-conscious synergy between human judgment and optimized machine assistance, deploying full automation only where the physical cost stack — measured rack by rack, substation by substation — genuinely clears the bar.

None of this is a permanent verdict. Accelerator efficiency, grid buildout, and model compression will keep chipping away at the physical cost stack over time, the same way earlier waves of American automation eventually got cheaper than the labor they displaced. But for the specific window the U.S. labor market is living through right now — with Dominion filing rate cases, PJM's auction hitting its price cap, and Loudoun County adding substations as fast as it can permit them — the humble sandwich, and the roughly 145 watts it takes to power the person eating it, is doing more competitive work against the American data center than most 2023-era forecasts ever priced in.