Sold to Enterprise. Rendered as Deepfake.

Every build-out keeps two ledgers. The first records what the capacity was sold as. The second records what the capacity is doing. For the AI build-out, the first ledger is the largest financing story ever told. The second is a feed.

The first paper in this series audited the liability side of the structure: who owes, and whether anyone is standing there at the exit. This paper audits the other side — not whether the demand is smaller than presumed, but what the demand is made of. The finding is not that the machines are idle. The machines are running constantly. The finding is what they are running.

The legitimate share is mostly deep, fake videos.

The capacity was sold to enterprise. That was the prospectus, in every deck, every filing, every roadshow: corporations would buy intelligence the way they buy electricity, metered and indispensable, and the cash flows would service the obligations. The obligations are documented — this series has the receipts. Anthropic’s amended S-1 carries aggregate commitments of roughly $895.4 billion: $447.5 billion by 2030 and $447.9 billion across 2031–45. The institutions underwriting the sector are quietly pushing $1.5 trillion. The payer of last resort is filed too: households hold a record 48.2 percent of their wealth in equities, through the index funds and pensions that own the build-out whether they chose it or not.

The enterprise customer was not a detail in this structure. It was the load-bearing assumption. Enterprise pays invoices. Enterprise signs multi-year contracts. Enterprise demand is the kind that can be taken to a bond desk. Everything else — the consumer, the curious, the crowd — was decoration in the model, upside, a rounding error with good engagement metrics.

So the sold ledger reads: capacity, financed at historic scale, against the most creditworthy demand class in the economy.

Now read the meter.

The only census of consumer usage published by a lab at scale is OpenAI’s, and it is candid. Of consumer ChatGPT requests, the work-related share fell from 51.6 percent in July 2024 to 32.2 percent in September 2026. The category OpenAI labels multimedia — image and media generation — more than quadrupled over the same window, from 2.0 to 8.8 percent of requests, the fastest-growing line in the file. Technical help — programming, mathematics, data analysis — collapsed from 13.5 to 4.6 percent. The earlier Harvard–OpenAI study of the same population put non-work usage above 70 percent of all consumer messages. The consumer product is drifting, measurably, away from work and toward media. And requests are the cheap unit. Media is where the compute is.

Because the units are not comparable, and the difference is the paper. A text turn costs about a quarter of a watt-hour: Google’s first-party measurement puts a median Gemini text prompt at 0.24 watt-hours; Epoch AI’s estimate for a typical ChatGPT query is about 0.3. Five seconds of generated video, measured on open models by Hugging Face researchers, cost between roughly 60 watt-hours on current efficient models and roughly 944 watt-hours on the 2024-era model MIT Technology Review metered — hundreds to thousands of text turns for five seconds. One finished minute of rendered video runs, on those measurements, from about 0.7 kilowatt-hours to about 11. The workload that is growing fastest is the workload that costs the most per unit, by orders of magnitude. The mix is shifting toward the expensive end of the meter at the same time as the meter itself is being financed.

What is the expensive end rendering? Sample one feed on one evening — November 5, 2026, this desk’s own, uncurated:

Roughly three million views in a single evening’s sample, and every item on the list is the same genre: real faces in fictional scenes. Public figures appear in this paper as data, the way they appear in the feed — named, not pictured; this desk does not traffic in the likeness it is auditing. The emblem of the sample is the dance number: the chief executive of the company that sells the build-out’s chips, rendered as a backup dancer behind a dancing president. The bet on the AI build-out is looking more like tech CEOs as backup dancers, while corporations struggle to figure out how to utilize the technology.

Nor is this merely one feed’s taste. In March 2026, OpenAI announced the discontinuation of Sora — the flagship consumer video product, the most famous renderer in the world — while the capacity build-out it was rendered on accelerated. The product could not carry its own compute. The dancing continued elsewhere, on other people’s models, at other people’s meter rates. The render is the demand. The render does not pay enterprise prices.

Perhaps the enterprise ledger is simply late. The record says it is not late; it is absent, and the absence has been measured by the friendliest possible measurers.

MIT’s NANDA initiative, in The GenAI Divide: State of AI in Business 2025, reviewed more than 300 public enterprise deployments, interviewed 52 organizations, surveyed 153 senior leaders, and found that about 95 percent of organizations were getting zero measurable return on generative AI, against an estimated $30–40 billion of enterprise spending. The honesty requirements of this desk oblige the caveats, and the caveats are real: the report calls itself preliminary, its success bar was a marked and sustained P&L impact judged roughly six months out, and its pilot figures are, in its own words, “directionally accurate based on individual interviews rather than official company reporting.” Cite it correctly and it still stands. McKinsey’s State of AI 2025, surveying nearly 2,000 organizations, found about 6 percent reporting AI contributing five percent or more of EBIT. RAND’s 2024 study of practitioners put the broader AI project failure rate above 80 percent — roughly double the failure rate of non-AI IT projects. BCG’s 2024 survey of 1,000 executives found 74 percent of companies had yet to show tangible value.

Four instruments, four methodologies, one shape. The demand that was sold has not materialized at anything like the financed scale. The demand that materialized was never in the prospectus. No deck from the two-trillion-dollar summer carried a slide titled synthetic video, likenesses of public figures, entertainment, because no bond desk would have underwritten it — not because the render is shameful, but because the render is unbillable at the required rate. Entertainment revenue is real; it is advertising and engagement revenue, priced per eyeball per second, and it does not service $895.4 billion of commitments. The gap between the two ledgers is not a rounding error. It is the whole structure, restated: capacity financed against one customer, consumed by another, at a price the first customer’s absence makes impossible.

The public has begun the audit on its own, from the only line of the ledger it can see: its bill.

In Gallup’s March 2026 polling, 71 percent of Americans opposed a data center in their community — 48 percent strongly — up from 47 percent opposed in late 2025. In the same survey, 53 percent opposed a local nuclear plant. The build-out is now less welcome in an American town than a reactor. The mechanism is filed, not vibes: on PJM, the grid serving 65 million people, capacity auction prices rose from $28.92 per megawatt-day for 2024/25 to $329.17 for 2025/26, and the grid’s independent monitor attributed roughly 63 percent of the increase to data-center demand. Wholesale electricity prices in areas of high data-center concentration have risen as much as 267 percent in five years. By late 2026, more than 550 American jurisdictions had enacted moratoriums or bans; in the first quarter of 2026 alone, 75 projects worth roughly $130 billion were blocked or delayed. The Tennessee Valley Authority, the country’s largest public utility, voted a separate, higher rate class for data centers effective October 1, 2026 — the principle of the load pays for the load, converted from a pledge into a tariff. A bipartisan Senate package would write the same principle into law.

People are wondering where the power bill is going. It is a better question than they know, because it is actually two questions. The first — why is the power being bought — has an answer on file: the build-out. The second — what is the power doing — has an answer on the feed: it is rendering. Some share of the most expensive electricity expansion in a generation terminates, measurably, in a president dancing, a prosecutor confessing for a defendant who never stood trial, a chief executive keeping time in the back row. The ratepayer was told the load was the future of enterprise productivity. The ratepayer is not wrong to ask for the meter readings. The meter readings are the rendered ledger, and the rendered ledger does not match the sold one.

No claim is made here that synthetic media is worthless, or that entertainment is an illegitimate use of electricity — the republic renders its fictions at scale every night and always has. The claim is narrower, and it is the series’ claim restated at the level of the workload: a financing is only as honest as the demand it names. This one named enterprise. It got the audience. An audience is not a customer class. It does not sign, it does not commit, it does not arrive at the exit with a balance sheet — and when the bill for its entertainment arrives socialized through a capacity auction, it does the one thing audiences have always done when the show stops being free.

It asks where the power bill is going.