AI Hit a Wall of Physics. The Industry Borrowed Against It.

The U.S. Treasury's career analysts — not outside critics, not short-sellers, not the usual Cassandras — have drafted a report identifying electricity bottlenecks as a risk factor in a systemic AI market bubble, one they describe as more deeply entrenched than the dotcom era. ^[1] Senator Elizabeth Warren has her own phrase for what is funding it.

The official position of the Secretary and the U.S. Treasury is that Artificial intelligence will be a key driver of America’s new Golden Age. — United States Department of the Treasury

The report is a draft. But the pattern it names is already visible in the quarter-by-quarter decisions of the companies building AI. The industry has run into a wall of physics — power plants, water tables, transformer backlogs — and its response has been to borrow against it, not adapt to it. The result is a compounding risk for which no regulatory instrument exists. The wall is real and it is already here. Goldman Sachs now states plainly that compute, not model quality, has become the binding constraint on scaling AI. ^[2] Datadog's 2026 engineering report found that roughly 5% of AI model requests fail in production, and nearly 60% of those failures stem from capacity limits, not model intelligence. ^[3] The bottleneck is not that the models are too dumb. It is that there are not enough chips, not enough power, not enough cooling to serve them. Anthropic quietly throttled Claude session limits during peak weekday hours — 5 a.m. to 11 a.m. Pacific — with roughly 7% of Pro users hitting limits they would not have encountered before. ^[4] Amazon CEO Andy Jassy put the situation plainly: his company cannot install capacity fast enough to meet the demand that is already there. ^[5] The demand is there. The capacity to meet it is not. And the gap between them is measured in years: PJM Interconnection, the grid operator serving 65 million people across the eastern United States, launched an emergency plan for 15 gigawatts of new generation targeting June 2031. ^[6] GE Vernova carries a $76 billion backlog for gas turbines and grid equipment — the industry is selling hardware faster than it can be installed. ^[7] The Department of Energy's Lawrence Berkeley National Laboratory projects data centers will consume 9.5% to 15% of total U.S. electricity by 2030, with hardware efficiency gains more than offset by computational demand growth. ^[8] The industry's response to this physical limit has not been to slow down. It has been to escalate financially. Morgan Stanley reports that hyperscalers doubled their gross leverage from 0.9x to 1.8x in two quarters, while UBS tracked $800 billion in additional U.S. credit over the past year to fund AI projects. ^[9] The borrowing compresses into quarters. The power arrives in years. That is the mismatch. And the trap closes from both sides. The physical constraint that triggers the borrowing also delays the returns needed to service the debt. Yann LeCun, Meta's chief AI scientist, has been blunt about where this leads.

Labs like OpenAI and Anthropic are going to have to increase prices, they're going to have to cut costs, or there's going to be a big bubble explosion. — Yann LeCun

The counter-evidence is real and should be stated honestly. Nvidia posted $215.9 billion in annual revenue, a 73% year-over-year surge, with CEO Jensen Huang forecasting $1 trillion by 2027. ^[10] This is not a bubble built on vapor — the revenue exists. But revenue can exist while costs outrun it, and LeCun's warning is precisely that the spread between the two is not closing fast enough. Cushman & Wakefield, the commercial real estate firm, argues that physical constraints like power and land shortages naturally prevent a data center overbuilding bubble — the limits act as a natural regulator. ^[11] That is a useful distinction, but it confirms the constraint, not the thesis that the constraint is manageable. The absence of overbuilding is evidence that the wall is real. What is missing is any mechanism to address the financial fragility the wall is creating. In June, the Federal Energy Regulatory Commission issued show-cause orders to six grid operators, determining that existing large-load tariffs are unjust and unreasonable given AI-driven load growth and requiring data centers to pay full grid upgrade costs. ^[12] The order addresses who pays. It cannot make power plants appear faster. Grid operators have 60 days to propose tariff changes; new generation takes years. The Treasury report identifies a systemic bubble risk for which no instrument exists — a risk the career analysts inside the government can see but no agency is equipped to defuse. ^[1] The physical world imposed the speed limit Washington never set. The industry's response to that limit has been to make itself more fragile, not less. And the gap between what regulators can do and what the moment demands is not closing — it is widening with every quarter of leverage and every year of delayed generation.