AI data centres (version 2) · All records

The numbers are stated judgments. The record began on 15 October 2026 with starting confidences that a model set as the average of five separate readings, from its general knowledge and up to ten developments for each assessment first reported in the three weeks before, after a second model had checked the developments the readings relied on, 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.

This record corrects the first record of the same question, AI data centres, which began on 12 October 2026 and continues unchanged under the rules it began with. What changed, and why is on this record’s method page.

Main assessment

Grid power and interconnection are the binding constraint on US AI data centre buildout through 2027.

67.5% stated confidence, Likely (55–80%). Starting point on the record’s first day: Likely (55–80%), 67.5 per cent, the average of five readings, all five of Likely. Attention level: For information. Trend: Steady.

Stated confidence in the one daily state so far: 15 October 2026, 67.5 per cent. A chart is drawn once there are 3 daily states.

Key judgment

The judgment as stated before the record began

The scarcest input for a new US AI data centre through 2027 is a firm grid connection of the size it needs, at the time it needs it. Accelerators and financing can usually be arranged within months. Transmission upgrades, interconnection studies and large power equipment take years, and the order in which projects receive them is set by utilities, grid operators and regulators rather than by the buyers. The assessment would weaken if power stopped being the reason given when projects slip, or if on-site generation and flexible-load agreements began to deliver large campuses on schedule.

Starting point

How the record started this assessment

On 15 October 2026, before the record began, a model read the 10 developments below five separate times, first reported in the three weeks before and linked to this assessment's causal factors or to its rivals', and chose a term each time. A second model, from a different family, checked each development a reading relied on, disputed 4 developments, and the five readings were made again without them. The record started the assessment at the average of the five final terms, taken in log-odds: 67.5 per cent, Likely (55–80%). Every reading and every check is kept in the first day's record.

The five final readings, in order: Likely, Likely, Likely, Likely, Likely. Their average, taken in log-odds, is 0.73, a stated confidence of 67.5 per cent. The readings were made by Claude Sonnet 4.6 (Anthropic) and checked by Gemini 4 Argon (Google).

Factors named above: M1 (Waits for large grid connections), M2 (Special terms for large loads: tariffs, deposits and pauses), M3 (Projects delayed, reduced or moved because of power), M4 (Lead times for transformers, switchgear and turbines), M5 (Ways around the grid: on-site power, co-location and flexible load) and C5 (Power still given as the reason for delays).

The developments shown

1. First reported 30 September 2026, 5 independent publishers. Linked to this assessment's own causal factors.
“FERC asks grid operator PJM to revise plan to shield homes from data center costs” · reuters.com · 30 September 2026 · Open the report · archived copy
“Biggest US Grid's Data-Center Plan Accepted With 5-Month Pause” · bloomberg.com · 30 September 2026 · Open the report · archived copy
“Grid manager delays data center power auction over concern on who gets the bill” · cleveland.com · 1 October 2026 · Open the report · archived copy
“Biggest US Grid Suspends Plan for Data Center Power Auction” · energyconnects.com · 5 October 2026 · Open the report · archived copy
“PJM Targets 6.8-GW Reliability Gap With Special Backstop Auction — EnergyPulse” · opensentiment.org · 2 October 2026 · Open the report · archived copy
And 10 more reports of the same development.
First round: relied on by 5 readings as support for the assessment. Second round: relied on by 5 readings as support for the assessment.
- Check of its use as support for the assessment: Upheld by the second model, run on Gemini 4 Argon (Google).
  Its reason, as the model wrote it: The reports confirm that FERC paused PJM's large-load power procurement plan for five months and ordered revisions so that data centers bear the costs attributable to their demand. This directly aligns with factor M2 regarding regulators setting special terms and pauses for large new loads to ration scarce capacity and allocate costs, supporting the assessment that grid power is the binding constraint.
  - Report 1, quoted: “FERC accepted key elements of PJM’s proposed power procurement but delayed implementation for five months and ordered revisions to ensure data centers and other large new users bear costs attributable to their demand.”
  - Report 3, quoted: “FERC suspended PJM’s planned emergency-power auction for up to five months, saying PJM had not adequately ensured that large new loads driving costs would pay their share rather than shifting costs to existing customers.”
2. First reported 8 October 2026, 4 independent publishers. Linked to this assessment's own causal factors.
“PJM Fast-Tracks 2.1 GW Capacity to Meet Rising Demand — EnergyPulse” · opensentiment.org · 10 October 2026 · Open the report · archived copy
“PJM accepts two ENGIE batteries and a gas uprate on its fast track | The Duck Curve” · theduckcurve.com · 9 October 2026 · Open the report · archived copy
“PJM fast track approves first 2.1 GW: Engie storage, LS Power gas uprate — Grid Brief” · gridbrief.org · 10 October 2026 · Open the report · archived copy
“PJM approves first 3 expedited interconnection projects” · globalenergyreport.com · 8 October 2026 · Open the report · archived copy
First round: relied on by 3 readings as support for the assessment and 1 reading as counting against it. Removed from the second round, because the second model disputed it.
- Check of its use as support for the assessment: Disputed by the second model, so the second round was read without it, run on Gemini 4 Argon (Google).
  Its reason, as the model wrote it: The reports state that PJM expedited the interconnection of 2.1 GW of generation capacity (two battery storage projects and a gas plant upgrade) into the grid, not large load connections (M1) or ways for a data centre campus to bypass the grid (M5). Furthermore, fast-tracking new power capacity relieves grid constraints rather than tightening them, so this development does not represent a change in M1 or M5 that supports the assessment that grid power is the binding constraint.
  Its ground: what the reports state is not a change in this causal factor.
  - Report 4, quoted: “PJM cleared two battery storage projects and a gas plant upgrade worth more than 2,100 MW under its new fast-track interconnection process.”
  - Report 2, quoted: “PJM expects projects to sign a Generation Interconnection Agreement within 10 months of submis”
- Check of its use as counting against the assessment: Disputed by the second model, so the second round was read without it, run on Gemini 4 Argon (Google).
  Its reason, as the model wrote it: The reports state that PJM fast-tracked two battery storage projects and a gas plant upgrade under a Generation Interconnection Agreement process for new capacity resources. Neither factor M1 (which covers how long a large new load waits for a grid connection) nor factor M5 (which covers on-site power, co-location, or flexible load arrangements that let a data centre campus draw power sooner) encompasses utility-scale grid generation interconnection queues.
  Its ground: what the reports state is not a change in this causal factor.
  - Report 2, quoted: “PJM expects projects to sign a Generation Interconnection Agreement within 10 months of submis”
  - Report 4, quoted: “PJM cleared two battery storage projects and a gas plant upgrade worth more than 2,100 MW under its new fast-track interconnection process.”
3. First reported 7 October 2026, 4 independent publishers. Linked to this assessment's own causal factors.
“Duke Energy protects customers from data center costs” · wowktv.com · 7 October 2026 · Open the report · archived copy
“Duke Energy strikes deal with data center giants” · axios.com · 7 October 2026 · Open the report · archived copy
“Duke Energy Corporation - Duke Energy protects customers from data center costs” · investors.duke-energy.com · 7 October 2026 · Open the report · archived copy
“FERC approves Duke large load transmission rules despite cost-shift concerns” · utilitydive.com · 13 October 2026 · Open the report · archived copy
First round: relied on by 5 readings as support for the assessment. Second round: relied on by 5 readings as support for the assessment.
- Check of its use as support for the assessment: Upheld by the second model, run on Gemini 4 Argon (Google).
  Its reason, as the model wrote it: The reports state that Duke Energy proposed and secured regulatory approval for special tariff provisions, minimum payments, and financial security requirements for large-load data center customers. This directly matches factor M2, which covers utilities and regulators setting special terms like minimum-demand charges and collateral for large new loads. Because a stronger M2 factor supports the main assessment and distinguishes it from the listed alternatives, the reading is upheld.
  - Report 2, quoted: “Duke Energy and large customers proposed North Carolina rules requiring major electricity users to cover connection costs and make minimum payments, addressing who bears the cost of data-center-driven grid expansion.”
  - Report 4, quoted: “FERC approved Duke Energy Carolinas’ tariff provisions for large-load transmission upgrades and financial security, while urging stronger safeguards against shifting data-center-related costs to existing customers.”
4. First reported 25 September 2026, 4 independent publishers. Linked to this assessment's own causal factors.
“FERC rejects Oklo complaint seeking to reinstate project in PJM interconnection study cycle” · utilitydive.com · 25 September 2026 · Open the report · archived copy
“FERC rejects Oklo's bid to put a 750 MW project back in PJM's queue — Grid Brief” · gridbrief.org · 26 September 2026 · Open the report · archived copy
“FERC Rejects Oklo Bid to Restore 750 MW Project to PJM Queue – American Buildout” · americanbuildout.com · 25 September 2026 · Open the report · archived copy
“FERC Backs PJM on Oklo Virginia Project, Finding No Rules Breach — EnergyReader.io” · energyreader.io · 29 September 2026 · Open the report · archived copy
First round: relied on by 5 readings as support for the assessment. Removed from the second round, because the second model disputed it.
- Check of its use as support for the assessment: Disputed by the second model, so the second round was read without it, run on Gemini 4 Argon (Google).
  Its reason, as the model wrote it: The reports state that a 750-MW power generation project by nuclear power company Oklo was removed from PJM's generator interconnection study cycle because Oklo failed to provide a workable model for its facility. Factor M1 concerns waits for large new loads requesting a grid connection, and Factor M3 concerns data centre projects delayed or moved because of power availability. A generator project being dropped from an interconnection queue due to an application deficiency by the developer does not represent a wait for a large load connection (M1) or a delayed data centre project (M3).
  Its ground: what the reports state is not a change in this causal factor.
  - Report 3, quoted: “The Federal Energy Regulatory Commission (FERC) has turned thumbs down on Oklo Inc.’s bid to get back in line for a 750-megawatt power project to PJM Interconnection’s first reformed study cycle.”
  - Report 3, quoted: “FERC’s said it found that that the company failed to provide a workable model for the full facility after receiving a chance to correct its application.”
5. First reported 8 October 2026, 3 independent publishers. Linked to this assessment's own causal factors.
“Energy Department Supports FERC’s Call for PJM to Protect Ratepayers and Build More Reliable Power | Department of Energy” · energy.gov · 8 October 2026 · Open the report · archived copy
“DOE presses PJM on ratepayer protections from large load costs | Utility Dive” · utilitydive.com · 8 October 2026 · Open the report · archived copy
“DOE Presses PJM On Ratepayer Protections From Large Load Costs” · zerohedge.com · 9 October 2026 · Open the report · archived copy
First round: relied on by no reading. Second round: relied on by 3 readings as support for the assessment.
No reading relied on it, so it was not checked.
6. First reported 5 October 2026, 3 independent publishers. Linked to this assessment's own causal factors.
“FERC Accepts and Suspends PJM’s Reliability Backstop Procurement Proposal, Establishes Paper Hearing Procedures, and Encourages a New Section 205 Filing | Troutman Pepper Locke - JDSupra” · jdsupra.com · 7 October 2026 · Open the report · archived copy
“FERC Suspends PJM's Reliability Backstop. The September 30 Window Does Not Open — Hyperscape Digital” · hyperscapedigital.com · 5 October 2026 · Open the report · archived copy
“FERC Accepts and Suspends PJM’s Reliability Backstop Procurement Proposal, Establishes Paper Hearing Procedures, and Encourages a New Section 205 Filing” · lexblog.com · 6 October 2026 · Open the report · archived copy
“FERC opens investigation into PJM Reliability Backstop Procurement tariff proposal | Headwaters Wire” · headwaterswire.com · 6 October 2026 · Open the report · archived copy
First round: relied on by 3 readings as support for the assessment. Second round: relied on by 5 readings as support for the assessment.
- Check of its use as support for the assessment: Upheld by the second model, run on Gemini 4 Argon (Google).
  Its reason, as the model wrote it: The reports confirm that on September 29, 2026, FERC accepted and suspended PJM's Reliability Backstop Procurement proposal for five months, preventing the September 30 window from opening and establishing a hearing on cost allocation, exit provisions, and collateral requirements. This directly supports the reading under factor M2, which covers regulators and grid operators setting special terms, collateral requirements, or pauses related to grid capacity and connection procurement.
  - Report 1, quoted: “On September 29, 2026, the Federal Energy Regulatory Commission (“FERC”) accepted PJM Interconnection, L.L.C.’s (“PJM”) Reliability Backstop Procurement proposal (“RBP Proposal”), suspending the proposal for five months and establishing a paper hearing on three elements of the RBP Proposal: (1) cost allocation; (2) Transmission Owner (“TO”) exit provisions; and (3) Load-Serving Entity (“LSE”) collateral requirements.”
  - Report 2, quoted: “FERC Suspends PJMs Reliability Backstop. The September 30 Window Does Not Open”
7. First reported 14 October 2026, 1 independent publisher. Linked to a rival's causal factors.
“Data Center Physical Infrastructure Market Grows 27 Percent Y/Y To $14 Billion In 2Q 2026, According To Dell'Oro Group” · menafn.com · 14 October 2026 · Open the report · archived copy
First round: relied on by no reading. Second round: relied on by no reading.
No reading relied on it, so it was not checked.
8. First reported 11 October 2026, 1 independent publisher. Linked to a rival's causal factors.
“Big Tech GPU supplies run dry as neoclouds expand” · g-enews.com · 11 October 2026 · Open the report · archived copy
First round: relied on by 1 reading as counting against it. Removed from the second round, because the second model disputed it.
- Check of its use as counting against the assessment: Disputed by the second model, so the second round was read without it, run on Gemini 4 Argon (Google).
  Its reason, as the model wrote it: The linked factor C5 concerns reports that data centre projects are delayed because of grid power, which means the grid still binds and would support—not count against—the main assessment that grid power is the binding constraint. While the report's title about GPU supplies running dry counts against the main assessment by favoring the accelerator supply alternative (MX1), GPU supply tightness is not a change in factor C5 as defined.
  Its ground: what the reports state is not a change in this causal factor.
  - Report 1, quoted: “Big Tech GPU supplies run dry as neoclouds expand”
  - Report 1, quoted: “intense competition for NVIDIA GPUs, rising rental prices, and supply constraints that could persist as data-center power infrastructure lags.”
9. First reported 5 October 2026, 1 independent publisher. Linked to a rival's causal factors.
“Nokia CEO says data centers would be built '2x faster' without supply constraints” · cnbc.com · 5 October 2026 · Open the report · archived copy
First round: relied on by 3 readings as counting against it. Removed from the second round, because the second model disputed it.
- Check of its use as counting against the assessment: Disputed by the second model, so the second round was read without it, run on Gemini 4 Argon (Google).
  Its reason, as the model wrote it: The development was linked only to factor C5, which belongs to assessment if_wrong (MXC) and represents reports that power delays continue to bind, thereby undercutting MXC and supporting the main assessment. Citing energy shortages as slowing construction does not count against the main assessment that grid power is the binding constraint, and even the mention of memory shortages is not covered by the linked factor C5.
  Its ground: what the reports state is not a change in this causal factor.
  - Report 1, quoted: “Nokia’s CEO says energy and memory shortages are slowing data-center construction, underscoring that infrastructure constraints are limiting the pace of AI expansion.”
10. First reported 28 September 2026, 1 independent publisher. Linked to a rival's causal factors.
“The Financing Problem Behind the AI Infrastructure Boom” · mondaq.com · 28 September 2026 · Open the report · archived copy
First round: relied on by no reading. Second round: relied on by no reading.
No reading relied on it, so it was not checked.

The readings, and their reasons

The first round, before the check

Reading 1: Likely (55–80%), from 10 developments shown. It relied on development 1 as support for the assessment, development 6 as support for the assessment, development 3 as support for the assessment, development 2 as support for the assessment, development 4 as support for the assessment, development 9 as counting against the assessment.
Its reason, as the model wrote it: Developments 1 and 6 directly support the assessment: FERC paused PJM's large-load procurement plan citing ratepayer cost allocation (M2), and Oracle's Wisconsin data centre lost ten months of progress and faces power grid approval delays pushing full load to 2029 (M3). Development 3 shows Duke Energy imposing special cost-protection terms on data centre customers (M2 supports). Development 2 shows PJM fast-tracking only 2.1 GW, a modest amount relative to the queue, which does little to relieve the constraint. Development 4 shows FERC rejecting Oklo's attempt to re-enter the PJM queue, confirming queue congestion (M1 supports). The rival alternative MX1 gains some color from development 9 reporting GPU supplies running dry, but that is a single source and the weight of multi-source evidence points to power as the binding constraint.
Reading 2: Likely (55–80%), from 10 developments shown. It relied on development 4 as support for the assessment, development 1 as support for the assessment, development 3 as support for the assessment, development 6 as support for the assessment, development 8 as counting against the assessment, development 9 as counting against the assessment.
Its reason, as the model wrote it: Development 4 shows FERC rejecting a bid to restore a 750 MW project to PJM's interconnection queue, reinforcing that the queue itself is a hard gate. Development 1 shows PJM's data centre power plan paused over cost allocation, directly bearing on M2 special terms and rationing of connection capacity. Development 3 shows Duke Energy negotiating special cost-protection terms with data centre operators, again consistent with M2. Development 6 shows FERC suspending PJM's reliability backstop procurement, signalling ongoing capacity adequacy stress. Developments 8 and 9, linked through the rival, report GPU supply tightness and Nokia's CEO saying centres would be built faster without supply constraints, which weakly support alternative MX1 but are single-source reports that do not clearly displace the multi-source grid evidence. On balance the grid-constraint case is well supported and the alternatives are not dominant.
Reading 3: Likely (55–80%), from 10 developments shown. It relied on development 1 as support for the assessment, development 4 as support for the assessment, development 3 as support for the assessment, development 2 as support for the assessment, development 9 as counting against the assessment.
Its reason, as the model wrote it: Development 1 shows FERC intervening in PJM's data centre connection plan and imposing a five-month pause over ratepayer cost allocation, directly bearing on factor M2 (special terms and pauses for large loads). Development 4 shows FERC rejecting Oklo's attempt to re-enter the PJM interconnection queue, reinforcing that queue discipline is tight and waits are real (M1). Development 3 shows Duke Energy striking a deal with data centres that protects other customers from costs, another instance of special tariff terms (M2). Development 2 shows PJM fast-tracking only 2.1 GW on its reliability path, which is a modest relief against a large queue and does not negate the binding character of M1. Development 9 (linked through rival) notes Nokia's CEO saying data centres would be built twice as fast without supply constraints, which could point toward accelerators as a rival constraint, but the same reports are consistent with power also binding; it modestly counts against the main assessment without overturning it. The pre-existing evidence that interconnection queues in the US run to years, that transformer lead times remain long (M4), and that no large-scale on-site alternative (M5) has emerged to dissolve the constraint keeps the assessment firmly likely rather than very likely.
Reading 4: Likely (55–80%), from 10 developments shown. It relied on development 1 as support for the assessment, development 2 as support for the assessment, development 3 as support for the assessment, development 4 as support for the assessment.
Its reason, as the model wrote it: Development 1 shows FERC pausing PJM's large-load procurement plan over cost-allocation disputes, pointing to ongoing interconnection friction (supports via M2). Development 3 shows Duke Energy imposing special financial terms on data centre customers to protect other ratepayers, a direct example of utilities rationing connection capacity (supports via M2). Development 4 shows FERC refusing to reinstate a 750 MW project in the PJM study queue, illustrating how crowded the interconnection pipeline remains (supports via M1). Development 6, linked through the rival, reports Oracle's Wisconsin campus pushed to 2029 for full power due to grid approval delays, a direct project delay caused by power availability (supports via M3). Development 2 shows PJM fast-tracking only 2.1 GW of new supply, which is modest relative to reported demand, providing limited relief (supports via M1). The rivals' GPU-supply developments 8 and 9 note chip tightness but do not show connected power sitting idle waiting for chips, leaving the grid constraint intact.
Reading 5: Likely (55–80%), from 10 developments shown. It relied on development 4 as support for the assessment, development 1 as support for the assessment, development 3 as support for the assessment, development 6 as support for the assessment, development 2 as counting against the assessment.
Its reason, as the model wrote it: The developments show continued grid friction bearing on factors M1 and M2. Development 4 shows FERC rejecting a 750 MW project's reinstatement in PJM's interconnection queue, reinforcing that queue backlogs persist. Development 1 shows FERC imposing a five-month pause on PJM's large-load plan over cost allocation disputes, adding regulatory delay to the connection process. Development 3 shows Duke Energy negotiating special cost terms with data centre operators, consistent with M2's rationing signals. Development 6 reflects ongoing regulatory debate over backstop procurement that keeps connection timelines uncertain. Development 2, which shows PJM fast-tracking only 2.1 GW, is a modest offset but does not dissolve the broader queue constraint. Rivals MX1 and MX2 lack comparable corroboration in the developments shown.

The second round, after the check, which set the start

Reading 1: Likely (55–80%), from 6 developments shown. It relied on development 1 as support for the assessment, development 3 as support for the assessment, development 5 as support for the assessment, development 6 as support for the assessment.
Its reason, as the model wrote it: Developments 1 and 6 directly support factor M1 and M3: FERC suspended PJM's data centre power plan pending revisions and Oracle's Wisconsin campus faces years of delays due to grid approvals, both showing power as a binding constraint. Development 3 supports factor M2, with Duke Energy striking special-terms deals with data centres to protect other ratepayers. Development 5 supports M2 further, with DOE pressing PJM on ratepayer protections for large loads. No development strongly supports alternative MX1, MX2 or MX3 as the dominant constraint, and M5 is not meaningfully strengthened by any development shown.
Reading 2: Likely (55–80%), from 6 developments shown. It relied on development 6 as support for the assessment, development 3 as support for the assessment, development 1 as support for the assessment.
Its reason, as the model wrote it: Development 6 (FERC suspending PJM's reliability backstop procurement and opening a paper hearing) shows continued regulatory friction around large-load grid access, supporting M2. Development 3 (Duke Energy striking a deal to protect customers from data centre costs with special terms for large loads) directly supports M2 by showing utilities imposing conditions on new large connections. Development 1 (PJM delaying its data centre power auction with a five-month pause citing cost allocation concerns) further supports M2. No development shown materially strengthens the alternatives MX1, MX2 or MX3, and M5 is not meaningfully bolstered. Prior knowledge of multi-year interconnection queues and long equipment lead times already set a high prior for grid as binding constraint, and the developments reinforce it.
Reading 3: Likely (55–80%), from 6 developments shown. It relied on development 1 as support for the assessment, development 3 as support for the assessment, development 5 as support for the assessment, development 6 as support for the assessment.
Its reason, as the model wrote it: Development 1 shows FERC pausing PJM's data centre load plan over cost allocation, pointing to ongoing grid-capacity rationing that bears on M2. Development 3 shows Duke Energy negotiating special terms with large data centre customers to protect ratepayers, also bearing on M2. Development 6 reports Oracle's Wisconsin campus pushed to 2027 at the earliest for partial power and 2029 for full load because of power-line approvals, directly illustrating M3 and M1. Development 5 shows the Energy Department pressing PJM on ratepayer protections, reinforcing that grid cost and capacity constraints are active policy concerns. No development shows a large bypass of the grid constraint through on-site power at scale that would strengthen M5 enough to offset these signals.
Reading 4: Likely (55–80%), from 6 developments shown. It relied on development 1 as support for the assessment, development 3 as support for the assessment, development 5 as support for the assessment, development 6 as support for the assessment.
Its reason, as the model wrote it: Development 1 reports that PJM delayed its data centre power auction and FERC asked for revisions over cost-allocation concerns, supporting M2 by showing regulators imposing special terms on large loads. Development 3 shows Duke Energy reaching a deal with data centres that includes protective terms for other ratepayers, also supporting M2. Developments 5 and 6 further reinforce that FERC and DOE are scrutinising PJM's capacity mechanisms for large loads, pointing to ongoing grid-capacity rationing. No strong developments emerged for the alternatives MX1 or MX2, and the waits-and-terms picture fits what was already widely known about multi-year interconnection queues, keeping the assessment likely rather than very likely because workarounds under M5 are also active in the market.
Reading 5: Likely (55–80%), from 6 developments shown. It relied on development 1 as support for the assessment, development 3 as support for the assessment, development 6 as support for the assessment.
Its reason, as the model wrote it: Development 1 shows FERC suspending PJM's large-load procurement plan and asking for revisions to protect ratepayers, signaling that grid capacity allocation for data centers remains contested and slow (factor M2). Development 3 shows Duke Energy requiring data centers to bear their own grid upgrade costs under special tariff terms, reinforcing that utilities are rationing connection capacity through financial conditions (factor M2). Development 6 shows FERC's backstop procurement mechanism suspended, further tightening the timeline for new grid capacity in the largest US grid region (factor M1). The Oracle Wisconsin project in the rival's development 6 also confirms a major project pushed to 2029 for full power due to grid approvals, directly illustrating factor M3. These developments consistently point to power and interconnection as the binding constraint, with no strong evidence that alternatives MX1 or MX2 are tighter.

Readings made before this one

No starting point was read before this one, as the first day records.

Uncertainty

Causes of uncertainty

Utilities and grid operators report requests for large connections in different ways, and a lengthening queue can reflect speculative or duplicate requests rather than real projects.
Most of the evidence arrives as news reports of announcements. An announced delay or an announced workaround is weaker evidence than a completed project.
On-site generation and co-location could relieve the constraint faster than grid upgrades, and their scale is not yet visible in public data.

What would change it

What would change the judgment

Several large campuses energized on or ahead of schedule without new transmission, using on-site generation or flexible-load agreements.
Utilities in the leading data centre markets reporting that large new loads are studied and connected within a year.
The largest spenders attributing their delays to the supply of accelerators or memory rather than to power.

Causal factors

The causal factors, and which way each one cuts

Factor	Which way it cuts	What evidence looks like	Last moved
M1. Waits for large grid connections How long a large new load waits between its request for a grid connection and energization, including the studies and network upgrades that stand in between. Longer waits mean power binds more tightly.	Supports. A stronger factor supports the assessment.	A utility or grid operator reports a longer study backlog or later in-service dates for large-load requests. A state regulator publishes the volume of data centre load waiting to be connected. A developer reports that its connection date has moved.	Not yet
M2. Special terms for large loads: tariffs, deposits and pauses Utilities and state regulators setting special terms for large new loads, such as minimum-demand charges, collateral, study deposits or pauses on new connections. These terms ration scarce connection capacity, and more of them means power binds more tightly. A pause that cites grid capacity bears on this factor. A pause that cites land use, water or noise bears on the local-permitting alternative instead.	Supports. A stronger factor supports the assessment.	A regulator approves a large-load tariff with minimum-demand or exit-fee provisions. A utility or grid operator requires deposits or collateral before it studies a large load. A state or utility pauses new large connections and cites grid capacity.	Not yet
M3. Projects delayed, reduced or moved because of power Announced delays, reductions in size or changes of site where the stated reason is the availability or timing of power.	Supports. A stronger factor supports the assessment.	A developer or cloud company says a campus will open later because its grid connection has slipped. A project moves to another state or region to obtain power sooner. A first phase is cut back to the power that will be available on opening.	Not yet
M4. Lead times for transformers, switchgear and turbines How long it takes to obtain the large equipment that delivers power to a campus. Longer lead times mean power binds more tightly, whatever the state of the grid itself.	Supports. A stronger factor supports the assessment.	A manufacturer or utility reports longer delivery times or sold-out capacity for large power transformers or gas turbines. A manufacturer announces new production capacity for this equipment, which bears on the factor in the other direction.	Not yet
M5. Ways around the grid: on-site power, co-location and flexible load Arrangements that let a campus draw power sooner than a conventional grid upgrade would allow, such as generation on site, siting beside an existing power plant or agreeing to curtail load at peak times. The factor is stronger when such an arrangement is approved, offered or delivers power to a large campus, and power then binds less, even when the reason given for the arrangement is that the grid is slow. A report that a project is waiting for the grid, with no such arrangement, bears on the factors on waits and delays instead.	Weakens. A stronger factor weakens the assessment.	A regulator approves co-location of a data centre at an existing power plant. A campus is energized on on-site generation ahead of its grid connection. A utility offers faster connection in exchange for load that can be curtailed.	Not yet

Assumptions

What the assessment assumes, and what would break it

MA1. Requests for large grid connections in ERCOT, the US grid with the largest public queue of large-load requests, keep outrunning what can be connected.
If this fails, the assessment fails.
Invalidated if, in three consecutive monthly Large Load Interconnection Status updates published by ERCOT, the megawatts shown as approved to energize rise from the previous update while the megawatts shown as having planning studies approved fall. If ERCOT stops publishing these figures, the assumption is reviewed rather than invalidated.
Source: ERCOT, Large Load Interconnection Status update in the monthly report to the Technical Advisory Committee. The criterion’s fingerprint was fixed on 15 October 2026 and is part of the record’s first day.
MA2. Large power equipment stays scarce enough that delivery dates, rather than orders, limit construction.
If this fails, the assessment weakens.
Invalidated if the Bureau of Labor Statistics producer price index for power, distribution and specialty transformer manufacturing (series PCU335311335311) is lower than a year earlier in two consecutive quarterly averages.
Source: US Bureau of Labor Statistics, Producer Price Index by industry, via FRED. The criterion’s fingerprint was fixed on 15 October 2026 and is part of the record’s first day.
MA3. Ways around the grid do not yet deliver large campuses at scale.
If this fails, the assessment weakens.
Invalidated if, before the end of 2027, at least three US AI campuses first energized after 4 October 2026 are stated by their owner or operator (in an SEC filing, a press release, or a filing with a state utility or environmental regulator) to be operating at 500 MW or more, mainly on generation on the site or at an adjacent power plant.
Source: SEC EDGAR full-text search, Owners' and operators' press releases, and state utility and environmental regulator filings. The criterion’s fingerprint was fixed on 15 October 2026 and is part of the record’s first day.

Other explanations

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.

MX1. The supply of advanced packaging and high-bandwidth memory is the binding constraint: accelerators cannot be produced fast enough, so connected power waits for chips.
This record can rarely test this explanation. Its scope leaves out reports that name the companies that make most advanced packaging and high-bandwidth memory, and the filter reads a report's summary, web address and outlet as well as its headline. So most reports that could favour this explanation are excluded before anything reads them, and the record can rarely see evidence in its favour.
No development was recorded on this assessment in the last 7 days.
Favoured by:
- The largest spenders say capacity is limited by the allocation of accelerators or memory rather than by power.
- Energized data centre space is reported as waiting for hardware.
- Accelerator makers report that they cannot meet orders into 2027.
Ruled out by:
- Lead times for accelerators shorten while projects keep slipping for power.
- Packaging and memory suppliers report inventory building up.
MX2. Capital is the binding constraint: the cost and availability of financing set the pace, and power follows money.
No development was recorded on this assessment in the last 7 days.
Favoured by:
- Projects are paused or cancelled citing the cost of financing.
- Data centre debt is repriced or pulled from the market.
- Spending outlooks are cut while power is available.
Ruled out by:
- Financing closes on schedule while energization dates slip.
MX3. Local permitting and community opposition, rather than grid capacity, is what slows buildout.
No development was recorded on this assessment in the last 7 days.
Favoured by:
- Pauses and zoning refusals cite land use, water or noise rather than grid capacity.
- Projects that have secured power are held up by local approval.
Ruled out by:
- Projects with local approval wait for a grid connection.
MXC. 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.
This is the alternative assessment, stated word for word as the main assessment’s rival: the two cannot both be right.
No development was recorded on this assessment in the last 7 days.
Favoured by:
- The largest spenders cut or slow capital spending while their projects are reported to have power available.
- Projects are paused, cancelled or deferred citing demand, returns or overcapacity rather than power.
- Rental prices for accelerators fall across the major providers while new campuses still find power.
Ruled out by:
- The largest spenders keep raising capital spending while their projects keep slipping for power.

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

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.

Considered, set aside

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

Changes to this assessment

No change to this assessment yet.

The other assessments

The rest of the judgment

The question’s text on this page is as sealed on the record’s first day. Fingerprint . Fingerprint pending. Records are fingerprinted each day at 23:59 UTC.