The AI Bill Finds the People Who Can't Say No

In Northern Virginia, the most crowded data-center market in the country, there is essentially nothing left to rent: vacancy is 0.3 percent. Inventory grew by a third last year anyway. Construction is running at twelve times its 2020 pace, and CBRE, the brokerage that tracks the market, expects the shortage to persist at least three more years ^[1]. The interesting change is not the demand. It is the unit of account. A CBRE director explained how the buildings are priced now.

My thoughts are we’re at least three years out before we catch up with the supply that equal the demand that's out there. — Pat Lynch

Commercial property has been quoted in square feet since there was commercial property. This one corner of it is now quoted in kilowatts, a physical thing rationed by availability. The product the buildings serve is quoted in tokens, and tokens keep getting cheaper. Megawatts, so far, do not. The re-pricing runs to the market's comic edges. Gorilla Technology Group sold $101 million of software last year. This week it announced a pivot into AI infrastructure around a contract for 25,856 GPUs, projecting $1.3 billion to $1.8 billion of revenue within two years, roughly eighteen times current sales, financed with debt that Fitch rates B-, which is to say junk: the grade below investment quality, reserved for borrowers with a live chance of default ^[2]. PwC, counting the whole boom rather than one company, puts global data-center investment at $1.1 trillion a year by 2030 and $31.6 trillion through 2050, which it calls the largest investment cycle in history, bigger than the railways or the electric grid ^[3]. The trouble with selling atoms is that they have to be repaid, and repayment happens at a meter: the per-seat subscription, the per-token API bill, the monthly invoice. The industry has never had more to repay. Tech companies committed roughly $740 billion to AI this year, 69 percent more than last ^[4]. At the meter, everyone who can say no is saying it, and each refusal comes with its stated reason. Enterprises were first. Pylon, a software maker, watched its projected annual Anthropic bill jump from $400,000 to $1.4 million after it crossed a 150-seat threshold: the introductory pricing expired, the new plans bill tokens separately, and every gain in usage arrives as a line item, which is why companies are installing internal spending caps and approval gates ^[5]. Uber ran through its entire 2026 AI tool budget in four months ^[4]. Swan AI, a four-person startup, reports a $113,000 monthly bill, and executives across the industry now describe computing costs that exceed the salaries of the people the software was supposed to replace ^[6]. The refusals are arriving before the adoption does. Only 23 percent of enterprises have scaled AI agents into even one business function, by McKinsey's count ^[7]. Customers are not refusing because they are sated. They are refusing because the price showed up. Microsoft is the hinge in miniature. It is canceling thousands of internal licenses for Claude Code, Anthropic's programming assistant, after a surge of experimentation pushed costs past what it would keep paying, while leaving untouched its $5 billion stake in Anthropic and a $30 billion commitment to buy Azure compute capacity ^[4]. One company, one budget. The metered intelligence, refused. The atoms, honored. The second refusal is to stop renting intelligence altogether and buy the atoms. Apple launched AI-capable Mac desktops this week, marketed against the per-token bill, and its hardware chief, Johny Srouji, made the ownership case in the plainest available terms.

There's no cost per token. You're just using the machine again and again. — Johny Srouji

Sam Altman is arguing the same exit from the other side of the meter, on privacy rather than price, for wearable devices that keep their processing on the device, a case that gained urgency after Meta, sued over human annotators reviewing users' private recordings, began moving toward camera-free smart glasses ^[8]. Altman put it plainly this week.

If a device is going to be listening to your whole life and watching your whole life, most people will not want that running on a cloud server. — Sam Altman

The labs themselves cannot charge what the atoms cost, and their answer has been the discount tag. OpenAI cut fees on its Luna model by 80 percent, and Anthropic launched a top-performing model at half its flagship's price, both to hold share against cheaper Chinese competitors, with xAI and Meta following; the one seller with demonstrated pricing power, China's DeepSeek, is raising prices ^[9]. What cannot be charged for gets rationed. Google imposed hard usage caps and tiered Gemini subscriptions in May. Anthropic quietly cut Claude's five-hour session limits during weekday peak hours, without formal notice, leaving power users watching quotas vanish in minutes. Salesforce moved its agent product to per-outcome pricing, billing for results rather than raw consumption, and a Dell director now warns corporate buyers against "tokenmaxxing," meaning agents set looping toward answers with no business rules attached, because the loop burns money and sinks the investment ^[10][11]. Analysts read the throttles as strategy rather than shortage: a way to move enterprises off fixed subscriptions and onto metered API use, where every token is sold rather than included.

To manage growing demand for Claude, we’re adjusting our 5 hour session limits for free/pro/max subscriptions during on-peak hours. — Anthropic

None of which means the demand is invented. The neoclouds, smaller operators that rent raw compute, report demand from the frontier labs outstripping supply, with revenue per megawatt expected to rise as contracts renew ^[12]. Anthropic is on pace for more than $100 billion of revenue this year, the largest application-layer figure the industry has produced ^[13]. But the insatiable customer in that demand report is the labs themselves: the order book for atoms measures the labs' spending, not the end-user revenue that must eventually cover it. And the industry's best number would fund about a tenth of the $1.1 trillion the industry plans to spend in 2030 alone ^[3]. Refused at the meter, the bill does not vanish. It moves down the chain, and the chain now carries macro weight: data-center investment passed consumer spending as a contributor to U.S. GDP growth in 2025, for the first time ^[14]. The first stop is the household meter. The Stargate buildout has contributed to rising electricity bills in 13 states, alongside an estimated $5.7 billion to $9.2 billion in public health costs from the fossil-fueled plants powering it ^[15]. Altman's response is a request one link further down: he has formally asked Washington to underwrite the $500 billion project, projecting that $1 trillion of AI infrastructure would add more than 5 percent to GDP within three years and requesting 100 gigawatts of new energy capacity a year ^[15]. California, this week, told the industry where its costs stop. Seven new laws create a utility rate classification requiring data centers to pay for their own grid and water upgrades instead of passing them to residents, a reversal for Governor Gavin Newsom, who vetoed earlier disclosure bills on the ground that they would slow the industry ^[16]. He was blunt about what changed.

With these laws, we are ensuring that Californians remain in the driver’s seat — and that those profiting from data centers aren’t doing so at our expense. — Gavin Newsom

India's refusal was made for it. When U.S. export controls forced Anthropic to disable its models for foreign nationals, Delhi tasked its planning agency with building sovereign AI on domestic GPU clusters, atoms bought for security with no revenue test anywhere in the purchase ^[17]. Sarvam AI, one of the country's own model builders, states the logic.

For AI users, it is clear that you should not confuse access with ownership, or adoption itself as an advantage. — Sarvam AI

And OpenAI bridges its share of the gap with the only parties still saying yes. Projections reported as it seeks funding at a $1.2 trillion to $1.5 trillion valuation show it burning $278 billion more cash than it generates from 2026 through 2030, $856 billion of that on computing, against roughly $840 billion of cumulative revenue, with the record $122 billion round from March exhausted by 2028 and no positive cash flow until 2030 ^[18]. The last refusal is underway in the market for the atoms themselves. Nvidia has doubled its revenue and projects 70 percent growth, and its stock now trades below 17 times forward earnings, the forward multiple being the share price divided by the next year's expected profit, down from more than 25 in May and half its level of a year ago ^[19]. The one company in the boom whose growth is not in dispute is being priced as if the earnings cannot last. Jensen Huang has a name for what the market has done to him.

the world's first and only growth value stock — Jensen Huang

A growth value stock is not a category anyone trades in. It is what you call a growth stock once the market starts grading it on cash. The arithmetic behind the re-rating was already public this month. Morgan Stanley counts the best case, a fully optimized data center running the newest Nvidia chips, as costing $25 billion a year to rent while producing $23 billion of output, with profitability hinging on holding the price of AI at $8.50 an hour in a market the bank expects to keep getting cheaper ^[20]. BCA Research sizes the requirement from the other end: to profitably carry a steady-state $1.4 trillion of annual spending, the four hyperscalers would need $7.4 trillion of revenue a year at a 30 percent EBITDA margin, EBITDA being earnings before interest, taxes and the depreciation that measures how fast these buildings wear out. Their depreciation alone is projected to double, from $255 billion next year to $581 billion by 2029 ^[21]. The next reading on the meter already has a date circled. Anthropic has filed confidentially for an IPO that may seek $2 trillion, a sum equal to about twenty years of this year's projected revenue ^[13]. When it prices, the boom will present its first bill to buyers who are simply free to walk. Nvidia, doubled revenue and all, already trades at seventeen times earnings. That is what walking looks like.