The numbers are stated judgments. The record began on 12 October 2026 with starting confidences that a model set in a single reading, from its general knowledge and up to ten developments for each assessment retrieved in the three weeks before (some first reported earlier), and a published rule has made every change since. They are not measured probabilities. The question and the rule are fixed before the record begins, and each change shows its evidence and its reasons. The rule and what is not claimed are on the method page.
Demand for AI computing sets the pace of US AI data centre buildout through 2027, with the largest spenders building as fast as the revenue AI earns justifies.
32.5% stated confidence, Unlikely (20–45%). Starting point on the record’s first day: Unlikely (20–45%), 32.5 per cent. Attention level: For information. Trend: Steady. Activation: Dormant. No recent change has moved toward it.
The judgment as stated before the record began
This is the leading alternative to the main assessment. If the grid is not the binding constraint, the most plausible pace-setter is demand. The largest spenders would build as fast as AI revenue justifies, and connection capacity would go unused when revenue disappoints. Evidence for it would appear first as slowing capital spending, cancellations that cite demand or returns, and falling rental prices for accelerators while projects still find power.
How the record started this assessment
On 12 October 2026, before the record began, a model read the 10 developments below, picked from the reports retrieved in the three weeks before (some first reported earlier), and with its general knowledge chose Unlikely (20–45%) in a single reading. The record started the assessment at the middle of that range, 32.5 per cent. The reading is kept in the first day's record.
Its reason, machine-written by anthropic/claude-sonnet-4.6: Development 4 notes that hyperscaler capital spending has continued to outpace cloud revenue growth, which is the opposite of factor C4 supporting this scenario. Development 9 from the main assessment block and development 6 here both cite power delays as the stated reason for project slippage at Oracle, directly undercutting factor C5 and reinforcing that the grid rather than demand is the binding constraint. Development 10 on open-source price pressure is a mild signal for factor C3 but is not strong enough to shift the overall picture away from power as the pace-setter.
Note added after the record began: This reason reads factor C4 the other way from its definition. C4 is revenue from AI growing more slowly than the spending behind it, and a stronger C4 supports this assessment. So a report that spending is outpacing cloud revenue, as development 4 is, bears in favour of this assessment under C4, not against it. The sealed reason above is unchanged.
Note added after the record began: This reason also refers to the report of power delays at an Oracle project as development 9 from the main assessment. That report is development 9 in the list below. Development 9 on the main assessment is a different report. The sealed reason above is unchanged.
- 1. First reported 9 October 2026, 1 outlet.“The Pilot #73 - How to win friends and export AI” · pacforum.org · 9 October 2026 · Open the report
- 2. First reported 7 October 2026, 1 outlet.“AI Data Centers Hit a "Power Crunch": A Full Breakdown of the Electricity Gap and Which Semiconductor Stocks Face the Biggest Risk?” · odaily.news · 7 October 2026 · Open the report
- 3. First reported 6 October 2026, 1 outlet.“Cipher Mining Sees AI Demand Holding Strong as Power Scarcity Drives Data Center Leases” · marketbeat.com · 6 October 2026 · Open the report
- 4. First reported 6 October 2026, 1 outlet. Cited by the reading.“The AI capex ledger: hyperscaler spending versus cloud revenue, 2024 to 2026 | BlockWest Research” · blockwest.co · 6 October 2026 · Open the report
- 5. First reported 5 October 2026, 1 outlet.“Nvidia, Broadcom shielded as AI power crunch hits chip supply chain, says Morgan Stanley” · rmb.reuters.com · 5 October 2026 · Open the report
- 6. First reported 4 October 2026, 1 outlet. Cited by the reading.“PSC wipes out 10 months of progress on Oracle's 1.3 GW Wisconsin data center — pushing partial 310 MW power to late 2027 and full load to 2029” · techradar.com · 4 October 2026 · Open the report
- 7. First reported 3 October 2026, 1 outlet.“America's data center fight is a preview of what's to come for the rest of the world” · newsbreak.com · 3 October 2026 · Open the report
- 8. First reported 28 September 2026, 1 outlet.“The Financing Problem Behind the AI Infrastructure Boom” · mondaq.com · 28 September 2026 · Open the report
- 9. First reported 24 September 2026, 1 outlet. Cited by the reading.“Oracle triggers 'force majeure' on data center project over power delays, source says” · reuters.com · 24 September 2026 · Open the report
- 10. First reported 24 September 2026, 1 outlet.“Open Source AI Models Force Price War as Enterprises Slash Inference Bills” · webpronews.com · 24 September 2026 · Open the report
Causes of uncertainty
- Demand and power can bind in different places at the same time, so national figures can hide a regional constraint.
- Rental prices for accelerators are partly set by list prices that change rarely, so they may lag a change in demand.
- A slowdown in spending can reflect cheaper computation rather than weaker demand.
What would change the judgment
- Several of the largest spenders cutting capital spending while their projects are reported to have power available.
- Cancellations and lease exits that cite demand, returns or overcapacity rather than power.
- Sustained falls in rental prices for accelerators across the major providers.
The causal factors, and which way each one cuts
| Factor | Which way it cuts | What evidence looks like | Last moved |
|---|---|---|---|
| C1. Slowing capital spending A slowdown in the growth of the largest spenders' capital spending, quarter on quarter or year on year. Slowing spending while power is available points to demand as the pace-setter. Public data named for this factor, not a test until admitted: sec.xbrl.microsoft.ppe_acquisitions_usd, sec.xbrl.meta.ppe_acquisitions_usd. | Supports. A stronger factor supports the assessment. |
| Not yet |
| C2. Cancellations and deferrals that cite demand Projects paused, cancelled or deferred where the stated reason is demand, returns or overcapacity rather than power. | Supports. A stronger factor supports the assessment. |
| Not yet |
| C3. Falling rental prices for accelerators Price cuts for rented accelerator capacity at the major cloud and specialized GPU providers. Falling prices point to spare capacity, which a market paced by demand would show. Public data named for this factor, not a test until admitted: azure.Standard_ND96isr_H100_v5.eastus.usd_per_hour, azure.Standard_NC80adis_H100_v5.eastus2.usd_per_hour. | Supports. A stronger factor supports the assessment. |
| Not yet |
| C4. AI revenue falling behind spending Revenue that the largest spenders attribute to AI services growing more slowly than the spending behind it, as reported in earnings releases. | Supports. A stronger factor supports the assessment. |
| Not yet |
| C5. Power still given as the reason for delays Continued reports that projects slip because of power. These mean the grid still binds, which weakens this scenario. | Weakens. A stronger factor weakens the assessment. |
| Not yet |
What the assessment assumes, and what would break it
- CA1. The largest spenders pace AI capacity to the returns they can see.Linchpin.
Invalidated if Microsoft, Alphabet, Amazon and Meta together raise quarterly payments for property and equipment year over year in two consecutive quarters in which their combined operating income falls year over year, as filed with the SEC (XBRL concepts PaymentsToAcquirePropertyPlantAndEquipment for Microsoft, Alphabet and Meta, and PaymentsToAcquireProductiveAssets for Amazon, with year-to-date figures differenced into quarters and the fourth quarter taken as the annual figure less nine months, and OperatingIncomeLoss).
Source: SEC EDGAR XBRL company facts. The criterion’s fingerprint was fixed on 9 October 2026 and is part of the record’s first day. - CA2. A slowdown led by demand would show in signed commitments before it shows in spending.Bridge.
Invalidated if the combined quarterly payments for property and equipment of Microsoft, Alphabet, Amazon and Meta (XBRL concepts PaymentsToAcquirePropertyPlantAndEquipment for Microsoft, Alphabet and Meta, and PaymentsToAcquireProductiveAssets for Amazon, with year-to-date figures differenced into quarters and the fourth quarter taken as the annual figure less nine months) fall year over year in a quarter that follows two consecutive quarters in which the combined total of the amounts each company discloses in its 10-Q or 10-K as leases not yet commenced at the balance-sheet date (for Amazon, the “Leases not yet commenced” row of its commitments table) grew.
Source: SEC EDGAR, quarterly and annual filings. The criterion’s fingerprint was fixed on 9 October 2026 and is part of the record’s first day.
The rival explanations
Evidence moves the assessment only when it bears differently on the assessment than on at least one of these. Evidence that fits them equally carries no weight.
- CX1. Demand softens but power still binds locally: a slowdown shows up as fewer new connection requests rather than as faster completion of those already waiting.No development was recorded on this assessment in the last 7 days.Favoured by:
- Connection queues shrink while waiting times stay long.
- Spending slows in regions with spare power but not where power is scarce.
Ruled out by:- Waiting times for connection fall as spending slows.
- CX2. Cheaper computation slows spending, because a falling cost per unit of AI output lets the largest spenders build less while use of AI services keeps growing.No development was recorded on this assessment in the last 7 days.Favoured by:
- Reported growth in the use of AI services alongside slower capital spending.
- The largest providers announce a falling cost per unit of AI output.
Ruled out by:- Use of AI services and capital spending fall together.
A development is recorded when the rule applies it to the assessment or sets it aside. Each count comes from the readings sealed with the changes, and a development set aside as fitting every alternative equally tells no alternative apart.
Evidence behind changes in the last 30 days
Each item is a report or data series cited by a sealed change on this assessment. The update loop reads headlines and summaries, and no summary is shown as the evidence.
No evidence has been cited in the last 30 days.
Evidence considered and set aside
Developments the update loop read in the last 30 days and set aside because they fitted this assessment and each of its rival explanations equally. Under the rule such evidence carries no weight, so none of it moved anything.
Nothing on this assessment was set aside in the last 30 days.
Changes to this assessment
No change to this assessment yet.
The rest of the judgment
The question’s text on this page is as sealed on the record’s first day. Fingerprint . In the fingerprint of 12 October 2026, published on GitHub. Check it in your browser.