You cannot audit OpenAI's losses. You can audit the electricity bill.

Almost every number you have read about how much money AI labs are losing is a leak. OpenAI, Anthropic and xAI are private; their revenue, burn and debt are not filed anywhere you can check. So this piece does not use those figures. What it uses instead is the part of the build-out that is public and auditable: capital spending in the national accounts, the physical power system, and the price households pay for electricity. Those three tell a coherent story, and it is not the one the leaks tell.

The headline finding is in the power data. US electricity generation was flat for fourteen years — an index reading of 99.4 in mid-2005 and 99.0 in mid-2019. Since then it has risen 11.2%, and 9.9 points of that came in the last three years alone. A system that did not grow at all through the smartphone era, cloud computing and the electrification of transport has started growing again, quickly.

Where the evidence stops

Not in this piece: any individual lab’s revenue, gross margin, cash burn or debt load, and the price at which inference is sold relative to what it costs to serve. Those companies are private. The widely-circulated figures come from reporters’ sources and may well be accurate, but they cannot be reproduced from a public series, so they are not stated here as fact.

In this piece: aggregate capital spending, the electricity system’s capacity and output, consumer and industrial power prices, semiconductor volumes and prices, credit spreads, and the stock of corporate debt.

A build-out of the claimed size has to appear in electricity and in capex whether or not anyone publishes a profit-and-loss statement. It appears in both.

The capex is back at dot-com proportions

Information-processing equipment and software together ran at $2,581bn a year in the most recent quarter — $1,783bn of equipment and $798bn of software. Set against total private nonresidential fixed investment of $4,490.5bn, that is 57.5% of everything American business spends on fixed capital.

Two things stand out. The share bottomed at 53.3% in 2024 and has climbed 4.2 points in two years — and in absolute terms, equipment and software spending is up $469bn a year over the same span. That is the build-out, measured without reference to anyone’s private accounts.

It is also worth noting what this is not. At 57.5% the share is still below the 62.1% of 2000. The dot-com peak was a higher-concentration event than this one, so far.

The power system is being rebuilt, and the last three years were absorbed

Capacity and output tell subtly different stories, and the difference is the interesting part.

Indices, 2017 = 100.

Since mid-2019, capacity is up 20.4% while generation is up 11.2%. On its own that looks like an overbuild — capacity added roughly twice as fast as output grew, with utilisation falling from 75.5% to 69.7%.

But split the period and it changes. Since mid-2023, capacity is up 10.9% and generation is up 9.9%. The recent additions are being consumed almost as fast as they are built. The long slide in utilisation is mostly an older story, and largely a compositional one: wind and solar have low capacity factors by nature, so a grid that adds renewables mechanically shows lower average utilisation without anything being wasted. That caveat matters, and anyone citing the utilisation decline as evidence of waste is over-reading it.

What is not compositional is generation growth resuming after fourteen flat years.

Who pays: households, measurably

This is the clearest answer in the data to the question of incidence.

Household electricity has outpaced headline inflation over every horizon, and over five years the gap is 14.7 percentage points. Electricity is not a discretionary purchase, and it is a larger share of spending for poorer households, so this is a regressive cost however it arose.

The third row is the one that deserves attention. Over three years, household electricity rose 14.9% while industrial power prices rose 5.8% — a 9.1-point asymmetry. Over ten years the two are close (48.2% versus 49.8%), so this is a recent divergence, not a structural feature. Large power buyers sign long-term contracts at negotiated rates; households pay regulated retail tariffs that absorb transmission and capacity costs. The build-out’s fixed costs land disproportionately on the customer with no bargaining position.

That gap cannot be attributed to data centres alone on this evidence. Grid hardening, fuel costs and state-level rate decisions all sit inside it. But the direction and the size are not in dispute.

The hardware layer already shows an overbuild signature

Semiconductors are the one part of this stack where volumes and prices are both public, and they are moving in opposite directions.

• Production is at a record 181.9, up 34.3% in three years.
• The producer price index for semiconductors is down 6.0% over the same three years.
• Capacity utilisation in semiconductors has fallen 11.8%, to 72.2%.

Record output, falling prices, and falling utilisation of capacity is the classic signature of a supply build that has outrun demand at the current price. It does not mean demand is weak — output is at an all-time high. It means capacity was added faster than demand grew, and the price is doing the adjusting.

For anyone buying compute, that is good news and it is the strongest verifiable argument that unit costs are falling. For anyone who financed the capacity expecting to price at yesterday’s levels, it is the opposite.

On the cost-versus-price question

This is where the public data runs out, so the distinction between measurement and reasoning matters.

What can be said from the data: the input cost of the hardware layer is falling, not rising. What cannot be said from the data: whether any given lab sells inference above or below what it costs to serve. That requires disclosed unit economics, and there are none.

What is worth stating plainly is the structure of the question. Selling below cost to capture a market is not irrational and not unusual — it is the standard playbook where a business believes costs fall with scale and switching costs rise with use. It becomes a problem in exactly two circumstances: when the cost curve flattens before the market consolidates, or when the financing that funds the gap reprices. Both are observable in advance, and both are in the watchlist below.

Debt and financing: calm on the surface, more expensive underneath

Aggregate credit conditions do not currently look stressed.

• The Baa corporate spread over Treasuries is 1.58 points, tighter than the 1.72 of a year ago. Lenders are not demanding more compensation for risk.
• But the 10-year Treasury is 4.71%, up from 4.26% a year ago. The risk-free rate against which every one of these projects is discounted has risen.
• Nonfinancial corporate debt stands at $8.98tn, against $6.82tn in 2019 — up 31.7%.
• The Nasdaq is at 24,976, up 23.6% year on year.

Put together: spreads say no one is worried, but the projects are being financed at a higher base rate, on a corporate debt stock a third larger than before the pandemic, with equity valuations that have already priced a good outcome. That is not a crisis. It is a configuration with less room for error than the spread alone suggests.

Our own Trade Stress Index reads 49.9 — normal, at the 62.9th percentile of the last five years. Nothing in the trade and funding data we track is flashing. That is a genuine finding and not a reassuring one: stress indices are contemporaneous, not predictive.

Who wins, who loses

Winning now, verifiably: anyone buying compute or chips, because the price index is falling while volume sets records. Equity holders in the hardware layer, on the Nasdaq’s showing. Large industrial power buyers, whose input costs rose 5.8% over three years against households’ 14.9%.

Losing now, verifiably: household electricity customers, by 14.7 points against headline inflation over five years. That is the one unambiguous transfer in this dataset.

Winning or losing later, unverifiable from here: the labs themselves, their private creditors, and the owners of generation capacity built on demand forecasts that assume the last three years continue.

The honest summary of the winners question is that the losses so far are diffuse and socialised through utility bills, while the gains are concentrated and visible in asset prices. Whether that reverses depends on facts that are not public.

Is it sustainable

Economically: on this evidence the build-out is not obviously a bubble and not obviously sound. The capex share is elevated but below 2000. Credit spreads are tight. But the hardware layer is already discounting, the risk-free rate has risen, and the debt stock is a third larger than 2019. The scenario that breaks it is not a demand collapse — it is a repricing of the financing while the cost curve is still flattening.

Environmentally: the electricity data supports one firm conclusion and refuses another. Firm: US generation is growing again after fourteen flat years, and capacity is up 20.4% since 2019. Refused: any claim about the emissions consequence. The most recent US total CO2 figure available on the series used here is from 2021, which is far too stale to say anything about a build-out that accelerated after 2023. Anyone telling you confidently what this has done to emissions is not reading it off public data.

That gap is itself the finding worth reporting. We are five years into a measurable expansion of electricity demand with no timely public emissions accounting to set against it.

What to watch

• Generation growth in the next two prints. It has run 9.9% over three years. If it keeps pace with capacity additions, the demand is real; if it stalls while capacity keeps rising, the overbuild case strengthens.
• The household-versus-industrial power gap, currently 9.1 points over three years. If it widens, the cost incidence is getting more regressive, not less.
• Semiconductor PPI, at −6.0% over three years. A flattening here would mean the cost curve is losing its slope, which is the first of the two conditions that turns loss-leading pricing into a problem.
• The Baa spread, at 1.58. This is the cleanest early indicator of the second condition. It is tight now; the move that matters is the widening, not the level.
• Tech share of capex, at 57.5%. The 2000 comparison is 62.1%. Above that, the concentration argument becomes much harder to dismiss.

Every figure above is drawn from the linked series and can be reproduced by downloading them. Where a number could not be verified that way, it is not in this piece.