Is “AI demand” one number, or a customer backlog and a building budget that only look alike?

Customer backlog and in-house capital spending both go by the name of AI demand, and the latest filings show why they have to be counted apart.

Grant Bishop · data as of 2026-10-21 · bears on W-6 · Show sourcesHide sources

Drafted with Claude Opus 5.5; counter-case by GPT-6 Astra · read and committed by the publisher

Every number shows where it came from: hover or tap a figure for its source, date and filing.

The question

In artificial intelligence (AI), one phrase has carried most of the weight. For most of this boom, “AI demand” has stood for a single quantity: more chips, more data centers, more power, and the assumption that all of it was spoken for. The latest disclosures, from the June quarter for three companies and the September quarter for Meta, no longer describe one quantity. They describe compute contracted to outside customers, counted as backlog, and compute a company builds for its own products, counted as capital spending.

The question is whether AI compute is now genuinely two businesses, each with its own measure of health, or one business reported through two sets of accounts. If it is two, a single “demand” figure conceals more than it reveals. The question bears on W-6.

W-6: Spending on the AI build is running ahead of the revenue it produces.

What is known

The national accounts register the build but cannot see the split. Private investment in computers and peripheral equipment ran at a seasonally adjusted annual rate of $420.074 billion in the second quarter, up from $399.615 billion in the first, on Bureau of Economic Analysis data published in the Federal Reserve Economic Data (FRED) database ¹. Investment in information processing equipment and software as a whole grew in real terms at a 10.7% annual rate from the preceding quarter ². None of these series separates compute deployed for a company’s own use from compute rented to others, and none is specific to AI.

The company disclosures do separate them, imperfectly. Cloud providers report remaining performance obligations (RPO), revenue contracted but not yet recognized, and that is where external demand shows up. Companies building for themselves report capital expenditure and guidance for it. The two measures answer different questions: one counts what customers have committed to pay, the other what a company has committed to spend. Meta’s latest release contains no backlog figure at all ³, so the two cannot be set against each other line by line.

Names that bear on this

Meta. On 21 October 2026 Meta filed its third-quarter results ³. Revenue rose 26% from a year earlier. Capital expenditures, including principal payments on finance leases, were $34.91 billion for the quarter, and Meta narrowed its full-year outlook to $135–145 billion from a prior $130–145 billion ³. The release carries no backlog figure; instead the company says its AI investments “are already delivering substantial incremental advertising revenue” ³. Bearing: the clearest instance of in-house compute measured by spending, and by returns on its own products, rather than by customer contracts. Next: its fourth-quarter results, which no filing has yet dated.

CoreWeave. Its quarterly filing (Form 10-Q) for the period ended 30 June 2026 said revenue recognized from customer commitments was 98% of the quarter’s total, and reported $103.7 billion of unsatisfied RPO, 41% of it expected to be recognized over the initial 24 months ⁴. It spent $14,117 million in cash on property and equipment in the first half of the year ⁴. No per-megawatt rental price from the filing is available to cite. Bearing: the strongest evidence that external compute has its own ledger: nearly all of its revenue is contracted in advance. Next: its third-quarter report, which no filing has yet dated.

Microsoft. On 29 July 2026 Microsoft reported, for the quarter ended in June, Microsoft Cloud revenue of $59.3 billion, up 27%, and commercial RPO of $678 billion, up 84% ⁵. Additions to property and equipment were $35,802 million in the quarter ⁵. Bearing: a contracted ledger and a heavy capital budget in the same company. Next: its next quarterly results, which no filing has yet dated.

Alphabet. Its quarterly filing for the period ended 30 June 2026 reported $519.5 billion of revenue backlog, $513.9 billion of it related to Google Cloud ⁶. It said that in the first quarter of 2026 it began to include contracts with an original expected term of one year or less, so earlier backlog figures are not directly comparable ⁶. Purchases of property and equipment were $80,598 million in the first half of the year ⁶. Bearing: a large external ledger whose growth cannot be read off earlier filings, held by a company that also builds heavily. Next: its third-quarter results, which no filing has yet dated.

The case for

“AI demand” now covers two trades. One is compute contracted to outside customers, booked as backlog. The other is compute a company builds for its own products, visible in capital spending. The latest evidence supports reading them separately, because they are measured, funded and paid for differently.

The external business has a ledger. Third-party compute now comes with contracts that can be counted. Microsoft’s commercial RPO rose 84% to $678 billion ⁵. Alphabet’s backlog, almost all of it tied to Google Cloud, is a second ledger of the same kind ⁶. CoreWeave, which rents out capacity, is the purest case: 98% of its second-quarter revenue came from customer commitments ⁴, so its income statement is close to a reading of its contracts. Each of these is a count of what customers have contracted to pay, which no capital budget can claim to be. A business whose revenue is almost entirely agreed in advance has its own measure of health, and that measure is not a spending plan.

The internal business has a budget and a return. Meta reports no backlog at all ³. What it reports is spending, a full-year outlook narrowed to $135–145 billion from $130–145 billion ³, and a return measured in its own advertising revenue, which the company says its AI investments are already lifting ³. That is a return on internal use, not a contract with anyone. No backlog figure captures that commitment, and no customer can enforce it.

The national accounts agree, with a caveat. The official series set out above show the build rising in nominal and in real terms ¹², and the Bureau of Economic Analysis names information processing equipment among the leaders of second-quarter investment ⁷. They cannot tell the two businesses apart, which is why the company ledgers matter: only there does the split appear.

Where the case is thinnest. Backlog is not uniform. Alphabet now counts contracts of a year or less in its backlog ⁶, so a short contract and a long one sit in the same total. Meta’s return is the company’s own description ³. These are arguments for the split, not against it. One word that covers a short contract, a multi-year commitment and a self-funded build hides exactly what a reader needs to price.

The case against

The strongest case against “AI compute is now two businesses” is that the disclosures establish different purchasing and accounting arrangements, not two demonstrably separate economic businesses. The distinction is useful for organizing evidence, but too strong as a conclusion about demand.

Variable 3: AI compute has split into a rented market and a government-funded one.

The boundary runs through companies, not between them. Microsoft holds a contracted ledger and, in the same quarter, added $35,802 million of property and equipment ⁵. Alphabet bought $80,598 million of property and equipment in the first half ⁶. CoreWeave, the purest renter, spent $14,117 million in the same half ⁴. Meta’s capital expenditures were $34.91 billion in its third quarter ³. Every seller of contracted compute is also a heavy builder, and nothing in these filings shows which machines serve a company’s own products and which serve its customers. That looks more like one build, booked in several ways, than a clean split into two businesses.

The proposed measures also answer different questions. Meta’s outlook measures planned spending, not customer revenue or returns from internal deployment ³. CoreWeave’s backlog measures future contracted revenue, with 41% expected within the initial 24 months ⁴. Alphabet’s change of method ⁶ means its backlog cannot be read against its own earlier figures. These disclosures cannot, by themselves, establish comparable demand strength, profitability or utilization. Spending commitments and revenue commitments are not interchangeable evidence.

Nor is a backlog as firm a measure as it looks. Microsoft’s commercial backlog grew far faster than the 27% rise in Microsoft Cloud revenue ⁵: commitments are running ahead of delivery, and revenue not yet recognized is a promise still to be kept, as a capital budget is. No per-megawatt rental price from CoreWeave’s filing is available to cite ⁴, so the claim that external compute has its own pricing rests on no filed figure.

The national accounts do not validate the structural claim either. The Bureau of Economic Analysis’s advance estimate says increases within second-quarter investment were widespread, led by industrial and transportation equipment as well as information processing equipment, and it attributes part of the quarter’s slowdown in output to decelerating investment ⁷. Those series cover all buyers ¹².

The defensible conclusion is narrower: AI infrastructure has multiple uses, counterparties and financing structures. These disclosures justify separating metrics. They do not yet demonstrate two distinct businesses, or establish what makes that division newly true “now.”

What we don't know, and when

Microsoft’s backlog. Whether Microsoft’s commercial RPO keeps growing faster than Microsoft Cloud revenue ⁵, and whether that growth is broad across customers or carried by a few large contracts. That is the cleanest available test of whether external demand broadens or narrows, and it bears on W-6. Microsoft’s next quarterly results will say; no filing has yet dated them.
CoreWeave’s pricing. Whether a rental price per megawatt can be cited from a CoreWeave filing at all ⁴, and whether the share of revenue from customer commitments holds. Its third-quarter report will test the claim that external compute has its own pricing.
Meta’s return. Whether Meta’s narrowed range holds through the year, and whether the company attaches a measure to the advertising revenue it says its AI investments deliver ³; the range bears on W-3. Its fourth-quarter results will show both; no filing has yet dated them.
Alphabet’s short contracts. How much of Alphabet’s backlog comes from the shorter contracts it began counting under its new method ⁶. Its third-quarter results will show whether it breaks them out.
The national accounts. Whether investment in computers and peripheral equipment keeps rising ¹. FRED gives the series’ next release as 29 October 2026 ¹.

W-3: Spare rented computing capacity moves work around; it does not cancel new power and cooling construction, which follows capital spending, not chip scarcity.

What would change our mind

The proposition that AI compute is now two businesses survives as a rule of measurement and is not yet proven as a fact about economics. That is where this piece stands against House View v3.0, variable 3, with the caveat that this piece tests a contracted-versus-in-house cut rather than the row's own split. The latest disclosures make one thing clear: no reader should accept a single “AI demand” figure again. Backlog at Microsoft, Alphabet and CoreWeave and capital spending at Meta are different instruments, and adding them together produces a number that means nothing.

But the case against is right that the boundary leaks. The companies that sell contracted compute also build on a heavy scale; Alphabet’s backlog changed what it counts; no rental price from CoreWeave’s filing is available to cite. A backlog is a promise from a customer; a capital budget is a promise a company makes to itself, and only one of them can be enforced.

The two-business reading becomes structural only when the two measures diverge: external backlog growing while internal spending is cut, or the reverse. Meta’s third-quarter outlook narrowed by raising its floor, not by cutting ³, so on the evidence so far the two still move together and may simply be two views of the same cycle. The next round of backlog disclosures from Microsoft, Alphabet and CoreWeave is the test.

This sharpens W-6; it moves nothing.

45 Degree Research

Members get the monthly deep dive, earnings-season previews with a post-print verdict, and a note when a filing moves one of our claims. Membership, $149 a year; 30-day full refund, no questions.

See what members get

Was this useful?

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 Opus 5.5; counter-case by GPT-6 Astra under Grant Bishop's methodology
Verification: multi-source verified, machine-drafted, human-gated
As of: 2026-10-21
Published: Oct 21, 2026, 17:46 ET · 2026-10-21 21:46 UTC
Bears on: W-6
What would change this: If the next disclosures show the two measures moving in opposite directions (contracted backlog at Microsoft, Alphabet and CoreWeave still growing while Meta cuts its capital expenditure range, or the reverse), the two-business reading would graduate from a measurement rule to a structural fact; if they keep moving together, as Meta’s narrowed range and the June-quarter backlogs did, it would not.
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.

Sources (7)

1Private fixed investment: Nonresidential: Information processing equipment and software: Computers and peripheral equipment (B935RC1Q027SBEA), FRED / BEA, updated Sep 30, 2026 — https://fred.stlouisfed.org/series/B935RC1Q027SBEA

2Real Private Fixed Investment: Private Fixed Investment in Information Processing Equipment and Software (A679RL1Q225SBEA), FRED / BEA, updated Sep 30, 2026 — https://fred.stlouisfed.org/series/A679RL1Q225SBEA

3Meta Reports Third Quarter 2026 Results (Form 8-K, Exhibit 99.1), Meta Platforms / SEC, Oct 21, 2026 — https://www.sec.gov/Archives/edgar/data/1326801/000162828026066700/meta-09302026xexhibit991.htm

4CoreWeave, Inc. Form 10-Q for quarter ended June 30, 2026, SEC, 2026 — https://www.sec.gov/Archives/edgar/data/0001769628/000176962826000366/crwv-20260630.htm

5FY26 Q4 Press Release: Microsoft Cloud and AI Strength Fuels Fourth Quarter Results, Microsoft, Jul 29, 2026 — https://www.microsoft.com/en-us/investor/earnings/fy-2026-q4/press-release-webcast

6Alphabet Inc. Form 10-Q for quarter ended June 30, 2026, SEC, 2026 — https://www.sec.gov/Archives/edgar/data/0001652044/000165204426000071/goog-20260630.htm

7GDP (Advance Estimate), 2nd Quarter 2026, Bureau of Economic Analysis, 2026 — https://www.bea.gov/news/2026/gdp-advance-estimate-2nd-quarter-2026

Show sourcesHide sources