The AI build is already contracted; its payoff is not yet measured
Signed backlogs at Microsoft and Google Cloud sit ahead of any revenue the economy has yet shown.
Drafted with Claude Fable 5.1; revised with Claude Opus 5.5 · read and committed by the publisher How we work →
Bottom line
- Do we really need this much AI?
- Not yet, on the evidence, but the buyers think so: the companies building the capacity have already sold it, in signed contracts Microsoft put at $678 billion and Google Cloud at $514 billion , and NVIDIA's data-center revenue grew 117 percent .
- Grant Bishop's view: What would change our mind: AI's economy-wide productivity lift does not arrive before 2028 or Spending on the AI build is running ahead of its revenue being proved wrong, or a major buyer walking away from capacity it has contracted.
What would change our mind
AI's economy-wide productivity lift does not arrive before 2028 or Spending on the AI build is running ahead of its revenue being proved wrong, or a major buyer walking away from capacity it has contracted.
The question
Do we need this much AI? Asked that way it is a mood, not a question. Asked properly it splits. What is the build buying now, in numbers someone has filed? And what is it about to buy, on a date someone can check? The House View holds a position on both: Spending on the AI build is running ahead of its revenue and AI's economy-wide productivity lift does not arrive before 2028. This piece tests that position against the filings.
Spending on the AI build is running ahead of the revenue it produces.
What is known: the build has customers
The clearest fact in the whole debate is that the capacity being built is already sold. Microsoft reported $678 billion of contracted commercial revenue not yet delivered at the end of June, up 84 percent from a year earlier . Alphabet put Google Cloud's backlog at $514 billion, up more than $50 billion in a single quarter, with just over half of it expected to turn into revenue within two years .
The suppliers to that build confirm it from the other side. NVIDIA reported on 26 August 2026 data-center revenue of $89.0 billion for the quarter to July, up 117 percent from a year earlier . Its next quarterly report shows whether that pace holds. TSMC, which makes the chips, reported on 15 October 2026 third-quarter revenue of $46.72 billion, up 41.2 percent from a year earlier . Its fourth-quarter report, for which it guided revenue of $52.3 billion to $54.3 billion , is the next check. ASML, which makes the machines that make the chips, reported on 14 October 2026 record quarterly sales of €11.7 billion and said its planned supply for 2027 and 2028 will not meet what customers want . Its fourth-quarter results are the next check on that claim.
Whatever this is, it is not a market that has stopped wanting the product.
What is known: the economy has not caught up
Now the other half. Output per hour in the U.S. nonfarm business sector rose 2.2 percent over the year to June, and 1.4 percent at an annual rate in the second quarter alone .
As of early May, 19.8 percent of U.S. businesses used AI in any part of their work. The number was 37 percent for firms with 250 or more employees. In retail it was about 14 percent .
The Federal Reserve's own researchers disagree about how much of the productivity gain AI explains. The Dallas Fed found that industries more exposed to AI have improved faster . The Kansas City Fed, looking at the same economy, concluded that AI adoption "explains little of the aggregate gain" in productivity . The Fed Board's own tracking shows what the build adds to growth and what it costs: in the first quarter, spending on software, computers, data centers and power facilities added to growth, and net exports of computers and peripherals subtracted from it . Much of the hardware is bought abroad, so the gross spending overstates what it does for the home economy.
Where AI shows up most clearly is at the level of the task, not the country. A study of customer-support agents found that an AI assistant raised the number of problems resolved per hour, with the largest gains for the least experienced workers and small gains for the best . Klarna told the Securities and Exchange Commission (SEC) in a 2025 registration filing (Form F-1/A) that its assistant handled 69 percent of customer chats in the year to June 2025, doing the work of more than 700 agents, and that revenue per employee had risen from $344,000 in 2022 to $972,000 . Its next annual filing shows whether those figures hold. Google said on 22 April 2026 that 75 percent of its new code is now written by AI and approved by an engineer, up from 50 percent last autumn . Alphabet's late-October quarterly report is the next dated disclosure against which to read it. Real gains, in a few places, in the kinds of work where the tool drops straight into the job.
Names that bear on this
Microsoft reported $678 billion of contracted revenue not yet delivered on 29 July 2026, up 84 percent . This cuts toward demand being real. Its next quarterly report, in late October, is the disclosure that would show whether the backlog keeps growing faster than revenue.
Meta reported on 29 July 2026 capital spending of $31.08 billion in the second quarter against free cash flow (cash from operations less capital spending) of $784 million, and guided 2026 spending of $130–145 billion, narrowed from $125–145 billion . This cuts toward Spending on the AI build is running ahead of its revenue. The next quarterly report shows whether that gap narrows.
Alphabet reported on 22 July 2026 second-quarter free cash flow of negative $5.9 billion, operating cash flow of $39.1 billion less capital spending . Its late-October quarterly report shows whether free cash flow turns positive.
CoreWeave, which rents AI computing to others, reported in its second-quarter filing, for the quarter ended 30 June 2026, net interest expense of $640 million, up 140 percent from a year earlier, on $35.6 billion of debt . This is where the price of money meets the build first. Its next quarterly report shows whether the interest bill keeps outrunning the business.
Just over the horizon
The longer-run uses of AI have dated markers now.
Medicine. In July, Insilico Medicine started a trial of rentosertib, a drug for a scarring disease of the lungs whose target was found by AI and whose molecule was designed by AI, in Phase III, the last stage of testing before approval, the first of its kind to reach that stage . The trial treats patients for 52 weeks, so no answer comes for at least a year.
Power. The Energy Information Administration expects U.S. electricity sales this year, 4,135 billion kilowatt-hours, and names data centers as a main reason. The same release noted "a pause in new data center projects in Texas" . The build is large enough to show up in the national grid, and the first sign of hesitation is in the same sentence.
Neither is an economic result yet. Both are checkable on a date.
The case for
The build is sized to contracts, not to hope. ASML says its planned supply will not meet what customers want . Where the tools have been applied to the right work, the measured gains are large . The evidence fits AI's economy-wide productivity lift does not arrive before 2028.
The case against
The strongest honest case against is not that demand is fake. It is Spending on the AI build is running ahead of its revenue. Alphabet's free cash flow was negative $5.9 billion in the second quarter . About one business in five uses the tools , and the Kansas City Fed finds adoption explains little of the gain . One widely cited estimate finds AI's total addition to productivity small . If that is right, the build is a very large capital program aimed at a gain the economy may barely register. The contracts are real, but a contract is a promise to pay.
What we don't know, and when
Whether the backlog converts on schedule (Spending on the AI build is running ahead of its revenue): the late-October reports from the cloud companies, where remaining performance obligations (contracted revenue not yet delivered) are a required disclosure.
Whether the lift is broadening (AI's economy-wide productivity lift does not arrive before 2028): the Census Bureau's survey, every two weeks, and the Bureau of Labor Statistics (BLS) productivity release for the third quarter.
Whether the first AI-designed drug works: not before the trial's 52-week readout.
Whether the leveraged edge holds: CoreWeave's next quarterly report.
What would change our mind
The evidence says the capacity is sold and the payoff is not yet measurable, which is what the frameThe frame: what we think is going on in the world, written down, dated, and revised when the evidence asks. The House View is the frame in full; this picture is the frame at a glance. already holds. This sharpens Spending on the AI build is running ahead of its revenue and AI's economy-wide productivity lift does not arrive before 2028; it moves nothing.
What would move them: AI's economy-wide productivity lift does not arrive before 2028's or Spending on the AI build is running ahead of its revenue's proved-wrong test being met (House View v3.0, AI's economy-wide productivity lift does not arrive before 2028; Spending on the AI build is running ahead of its revenue), or a major buyer walking away from capacity it had contracted, which would mean the demand under the backlog was softer than the filings say. Either would show up in filings, and we will say so when it does.
How this piece was made
Sources. 19 numbered sources, in the Record.
Related claims
Tell the publisher
Something we should look at? Suggest a company, a topic or a feature.
Record
Written according to the Grant Bishop standard — machine-researched, machine-checked, human-judged; then read and committed by the publisher. Every number traces to a filing or a dated market record; every claim carries a date and what would prove it wrong.
- Drafted by
- Claude Fable 5.1; revised with Claude Opus 5.5 under Grant Bishop's methodology
- Verification
- multi-source verified, machine-drafted, human-gated
- As of
- 2026-10-19
- Published
- Oct 20, 2026, 09:30 ET · 2026-10-20 13:30 UTC · changelog
- Bears on
- What would change this
- AI's economy-wide productivity lift does not arrive before 2028 or Spending on the AI build is running ahead of its revenue being proved wrong, or a major buyer walking away from capacity it has contracted.
- Position disclosure
- The publisher holds index-level U.S. equity exposure and no direct position in any single name discussed.
- Not advice
- General commentary for a general audience. Not advice for your situation.