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.

How the judgment is kept on the record

The question and its three assessments were fixed before the record began. A model then read up to ten developments for each assessment, first reported in the previous three weeks and linked to its own causal factors or to its rivals’, five separate times, and chose a term from the ICD 203 scale each time. A second model, from a different family, checked each development a reading relied on, the five readings were made again without the developments it disputed, and each assessment started at the average of its five terms. Since then each confidence has moved only as the rule below allows. The rule, and the rules that derive each daily state, are published here word for word, and their fingerprints are sealed in the record’s first day.

The record’s first day: Fingerprint . Fingerprint pending. Records are fingerprinted each day at 23:59 UTC.

Update rule, version v2

The rule, verbatim

Update rule, version 2

Each assessment carries a stated confidence. It starts at the confidence the system set for it when the record began, and it moves only as this rule allows. The rule was fixed before the record began. Changing any part of it creates a new version, fingerprinted and announced in the change log, and thresholds are never lowered to make an assessment move. Every change after the first day is made by this rule, and the daily check refuses any change the rule does not allow.

Evidence. Only a dated report with a web address from an outside source, or a reading of a public data series admitted for this purpose, can move an assessment. The project's own analysis never can.

Events. Reports of the same development are folded into one event, which counts once however many outlets repeat it.

What the record already knew. A development counts as known when the record began, and never moves an assessment, if any report of it had been retrieved by the time the starting point was read. A report retrieved later is new evidence, whatever date it carries.

Weight. An event bearing on a causal factor has the weight e = stance × sign × grade × materiality × diagnosticity. Stance is +1 when the event strengthens the factor and -1 when it weakens it. Sign is +1 when a stronger factor supports the assessment and -1 when it undercuts it. Grade is 1.0 for a source graded high, 0.6 for medium and 0.3 for unknown, taken from the best-graded outlet reporting the event, as the grade table fixed when the record began gives it. The table grades a web address by its domain, or failing that by the domain it belongs to, so news.example.com is graded as example.com. Materiality is 1.0 for a material development and 0.5 for a routine one. Diagnosticity is 1 when the event bears differently on the assessment than on at least one of its alternatives, and 0 when it fits them equally. An event with no diagnosticity carries no weight and is listed as considered and set aside.

Size. An assessment moves by 0.10 × Σe in log-odds, summed over the events applied to it. One high-grade, material, diagnostic event therefore moves it by 0.10 in log-odds, about 2.5 percentage points at even odds, and no single event can move it further.

Corroboration. An event moves an assessment only once at least 2 independent publishers have reported it within 7 days of its first report. Two reports come from one publisher when their web addresses share a registered domain or the grade table names them as one owner, and a report that copies another's text counts with the report it copies. Until then it is listed as noted and awaiting corroboration. After 7 days without a second publisher it is listed as uncorroborated and moves nothing by itself.

Cooldown. A causal factor moves its assessment at most once in 48 hours. Several events on the same factor in one update move it together, once. Evidence on a factor inside its cooldown waits until the cooldown ends.

Daily limit. The net move of one assessment within one UTC day stays within 0.20 in log-odds in either direction. When the events ready to apply would take it further, they are taken in the order they were first reported, each applied only if it keeps the net move within the limit, and the rest wait for the next update.

Challenge. Before any move is applied, a second model, from a different family than the one that read the event, looks for the strongest reason the reports do not support it. A move it disputes is not applied. It is recorded as set aside, with the challenge. A move whose challenge could not run waits for the next update.

Attention level. The attention level shown changes to the computed one once the computed level has held for 3 consecutive daily states. The activation of the contingent assessment changes the same way.

No drift. There is no decay and no pull toward even odds. A stated confidence stays where the evidence left it.

SHA-256 of the text above, as UTF-8: . This is the text of the rule sealed in the record.

State rules, version v2

How each daily state is derived

How the daily state is derived, version 2

Starting point. Before the record begins, the system gathers the developments it has linked to the causal factors of each assessment, and of its rivals, that were first reported in the previous 21 days. Reports of one development are folded into one. Those reported by at least 2 independent publishers come first, then the most widely reported, then the newest, up to 10 per assessment, of which at most 5 come through its rivals' factors, each side ranked the same way. Developments outside the record's scope are left out. A model reads them five times, separately, and each time chooses for each assessment one of five terms from the ICD 203 scale of likelihood, from very unlikely to very likely, and says which developments it relied on and in which direction. Where the developments shown are few, a reading may rest on what was widely known before them, and its reason says so. A second model, from a different family, then checks each development a reading relied on, in the direction it relied on it, and the five readings are made again without any development it disputed. Each assessment starts at the mean, in log-odds, of the five terms' midpoints: 12.5 per cent for very unlikely, 32.5 for unlikely, 50 for even chance, 67.5 for likely and 87.5 for very likely. Every reading and every check, the models that made them and the fingerprint of every reading made before the record began are kept in the first day's state.

Probability. A stated confidence held as log-odds L is shown as the probability 1/(1+e^-L), rounded to four decimal places, with its term from the ICD 203 scale of likelihood.

Thesis status. The alternative is the main assessment's rival: the two are answers that cannot both be right. The thesis is inverted when the main assessment is below the lower level set for it, when the alternative is more likely than the main assessment, or when the alternative is shown as active. It holds when the main assessment is at or above the level set for it, the assessment it rests on is at or above its level, the alternative is below even chance (50 per cent) and the alternative is shown as dormant. Otherwise it is strained. The levels were fixed with the question before the record began.

Trend. Up or down when today's stated confidence differs by at least 5 percentage points from the mean of the previous 6 daily states, and stable otherwise, or when fewer than 3 daily states exist.

Stability and attention level. Each day, moves of at least 0.10 in log-odds in either direction in the previous 48 hours raise an assessment's stability from stable to watch. The contingent assessment is also at watch while it is computed as showing early signals, and stressed while it is computed as active. The attention level is then read from the published ladder, using stability, stated confidence and trend. The attention level shown changes to the computed one once the computed level has held for 3 consecutive daily states.

Contingent activation. Each day the contingent assessment is computed as dormant, showing early signals or active, from moves of at least 0.10 in log-odds toward it in the previous 48 hours. It steps down at most one level a day. The activation shown changes to the computed one once it has held for 3 consecutive daily states.

Record. One state is written for each UTC day, fingerprinted and linked to the one before it. A day the job missed is written later, marked as recorded late, and computed from the changes that had been recorded by that day's close.

Reading. Each state after the first also records what the update loop read in its window: how many reports it compared with the causal factors for the first time, how many it linked to a factor, how many repeated a development already reported, how many developments wait for a second outlet, and the developments it set aside because they fitted an assessment and each of its alternatives equally, listed up to 20 and counted beyond that. Reports excluded for scope are counted and never listed. This is a record of what was read. Nothing in it moves an assessment.

SHA-256 of the text above, as UTF-8: . This is the text of the state rules sealed in the record.

What is learned and what is fixed

Each step of the update loop, and how it is measured

Models find, fold and read the evidence. Only the rule moves a confidence, and the rule is not learned.

Step	What it does	How it is measured	Does it learn?
Starting point	Before the record begins, a model reads up to 10 developments for each assessment, first reported in the previous three weeks and linked to its own causal factors or to its rivals’, five separate times, and with its general knowledge chooses a term from the ICD 203 scale each time. A second model, from a different family, checks each development a reading relied on, and the five readings are made again without the ones it disputed. The assessment starts at the average of the five terms, taken in log-odds.	Every reading and every check, the developments shown and the fingerprint of every earlier reading, kept in the first day's record.	No. Before the record begins the loop reads a starting point once a day and keeps the fingerprint of each reading, and the first day seals the newest complete one.
Attention	Finds new outside reports that bear on a causal factor: a shortlist by text similarity, then a relevance check by a small model.	The reports read and linked each day, in each daily state. Recall is not measured.	No. The threshold is set in a parameter file, and every change records a fingerprint of that file.
Consolidation	Folds reports of the same development into one event, and counts the independent publishers that reported it. Reports on one web domain, or on domains the grade table names as one owner, are one publisher, and a report that copies another's text counts with the report it copies.	The repeated reports folded each day, in each daily state, and the publishers counted and copies folded for each event, in each change.	No. The thresholds are set in a parameter file, and every change records a fingerprint of that file.
Interpretation	Reads what an event says about the causal factor, the assessment and each alternative. Evidence that fits the assessment and its rival equally carries no weight.	The developments set aside as fitting an assessment and its rival explanations equally.	No.
Source grading	Grades each outlet high, medium or unknown from a published table fixed when the record began, by its web domain or, failing that, the domain it belongs to. Every outlet not in the table is graded unknown.	The share of evidence with a grade.	No. The table is fixed when the record begins, its fingerprint is sealed in the first day, and the table is published below.
Revision	Applies the published rule below to move a stated confidence.	How often a move is reversed within seven days.	No. A change to the rule is a new, fingerprinted version.
Challenge	A model from a different family argues against every move before it is applied. A disputed move is set aside, with the challenge on the record.	The share of moves it disputes.	No.

Changes to how the loop works

Dated changes outside the rule

The rule never changes within a version. The models and instructions the loop uses can, and each change is listed here with its date and reason.

15 October 2026. The relevance check and the challenge step moved from Gemini 2.5 Flash to Gemini 3.1 Flash-Lite, both made by Google, because the provider retires Gemini 2.5 on 20 October 2026. Their instructions did not change. The challenge step still comes from a different family than the model that reads each event, as the rule requires. Every change recorded since names the model that made it.

Version 2

What changed from version 1, and why

This record corrects the first record of the same question, AI data centres, which began on 12 October 2026. That record continues unchanged under the rules it began with. A sealed record cannot be edited, so the corrections below begin a new record with its own first day. Each answers a defect a review found in the first record.

Changes to the question

As the question file states them, sealed in this record’s first day.

Causal factor C4 of the alternative, AI revenue falling behind spending, now reads: spending cut or slowed because AI revenue lags it. It counts the largest spenders cutting or slowing their planned spending and giving lagging AI revenue as the reason. Spending that keeps rising while AI revenue lags behind it now counts against the alternative.
Why: The version 1 wording read both ways. AI revenue falling behind spending can be taken as a sign that cuts are coming, which supports the alternative, or as spending running ahead of revenue, which contradicts its claim that the largest spenders build as fast as the revenue AI earns justifies. The reason sealed with the first day's reading took it the second way, against the factor's own definition, and the update loop read the same report the same way.
Causal factor M5 of the main assessment, ways around the grid, now says that an arrangement which is approved, offered or delivers power to a large campus counts as power binding less even when the reason given for it is that the grid is slow, and that a report of a project waiting for the grid, with no such arrangement, bears on the factors on waits and delays instead.
Why: A workaround adopted because the grid is slow could be read either way: as proof that the grid binds, or as a way around it. The update loop read one such report both ways at once, saying it strengthened this factor and also made the main assessment more likely, which the factor's sign does not allow, so the reading was set aside.
The alternative's statement is added, word for word, to the main assessment's alternatives as MXC, and marked as the main assessment's rival.
Why: In version 1 the alternative that demand sets the pace was not among the main assessment's own alternatives. Evidence that favoured demand over power was never weighed against the main assessment, and nothing kept the two from both being rated likely although they cannot both be right.

Changes to the rules

The rule and the state rules above are version 1’s with these changes. The rule’s numbers are unchanged.

The thesis status reads how likely the alternative is
The thesis now holds only while the alternative is below even chance, and it is inverted when the alternative is more likely than the main assessment. The rule still reads recent changes toward the alternative as well.
Why: The first records read the alternative only through recent changes toward it, never its level, so a thesis could read as holding beside an alternative the record itself rated likely. Microsoft's did from its first day.
A starting point that meets the standard later changes have to meet
The starting point is five separate readings instead of one. Each assessment is shown the developments linked to its rivals as well as its own, reports of one development are folded into one, and only developments first reported inside the three-week window are shown. A second model, from a different family, checks each development a reading relied on, the five readings are made again without the developments it disputed, and each assessment starts at the average of its five terms. Every reading made before the record began is listed by its fingerprint.
Why: Each first starting point was one reading, never challenged. Some counted one news story as several developments, included reports from before the stated window, and saw only the evidence on each assessment's own side. The challenge step later disputed six of the seven cited developments it reviewed, and earlier trial readings were not disclosed, so a reader could not rule out that a reading had been chosen.
Grades from a published table that knows primary sources
Outlets are graded from a table fixed when the record begins, published below with its fingerprint, which is sealed on the first day. Besides news outlets it grades official government and regulator sites, companies' own newsrooms and investor pages, and trade press that does its own reporting, each entry with its reason. A web address is graded by its domain or, failing that, the domain it belongs to.
Why: The first records graded from a list of news outlets matched on the exact web address. Every Microsoft and Cohere report, and about nine in ten of the AI data centre reports, were graded unknown, government sites and company filings included, so every move they could make was as small as the rule allows.
Question files that say what they mean
Each question file names the record it corrects and what it changed, listed above, and states the alternative word for word among the main assessment's own rivals, marked as its rival.
Why: In the first records the alternative was not among the main assessment's own rivals, so evidence for one was never weighed against the other, and some wording let both be rated likely at once.
Evidence the starting point never saw can move the record
A development counts as known when the record began, and never moves an assessment, only if one of its reports had been retrieved by the time the starting point was read. A report retrieved later is new evidence, whatever date it carries. The rule now says so.
Why: The first records pass over any development first reported before they began, by a rule no published text stated. Reports often carry a date without a time and arrive a day or more later, so some developments were neither shown to the starting point nor allowed to move the record. A regulator's decision on a grid operator's plan for large new loads, reported by four publishers, fell into that gap.
Two publishers means two publishers
A development moves an assessment only once two independent publishers have reported it. Reports on one registered domain, or on domains the grade table names as one owner, are one publisher, and a report that copies another's text counts with the report it copies.
Why: The first records count distinct web addresses, so one firm's alert posted on two of its sites, or a press release repeated word for word by a local station, counted as two outlets, while independent reports of one development were often not grouped at all.
Every move is challenged
Before any move is applied, a second model, from a different family than the one that read the event, looks for the strongest reason the reports do not support it. In the first records only material moves were challenged.
Why: The challenge step disputed 10 of the 11 material moves it saw, often on grounds its own instructions ruled out, while routine moves were never challenged. A routine development therefore carried more weight in practice than a material one, the reverse of the rule's own weights.

Source grading

The grade table, and how publishers are counted

Each report's outlet is graded high, medium or unknown from the table below, fixed when the record began. The table grades a web address by its domain, or failing that by the domain it belongs to, so news.example.com is graded as example.com unless the table grades news.example.com itself. A leading www is ignored. Every outlet not in the table is graded unknown, and no grade is entered by hand.

A development's grade is that of the best-graded outlet reporting it, and under the rule a development graded unknown weighs less than a third as much as one graded high.

The table’s fingerprint is the SHA-256 of its RFC 8785 canonical JSON: 162ff1a37351f22af0675a256c4bd6aec9399585b559a43aedf32685a591ffa2. This is the fingerprint sealed in the record's first day, so the table below is the one the record grades by.

The table below lists 146 domains graded high and 194 domains graded medium. It is version 2 of the table.

Graded high: 146 entries

Entries graded high, in order of domain.
Domain	Why it has this grade
afp.com	Agence France-Presse, a news agency with original reporting. From the credibility seed.
alberta.ca	Government of Alberta. An official source.
apnews.com	Associated Press, a news agency with original reporting. In version 1's list.
arxiv.org	arXiv, the research preprint server run by Cornell University. In version 1's list.
azcc.gov	Arizona Corporation Commission, Arizona's utility regulator. An official source.
bbc.co.uk	BBC, the UK public broadcaster. In version 1's list.
bbc.com	BBC, the UK public broadcaster. In version 1's list.
bis.org	Bank for International Settlements, an international financial institution. An official source.
bloomberg.com	Bloomberg News, a news agency with original reporting. In version 1's list.
bls.gov	US Bureau of Labor Statistics. In version 1's list.
brookings.edu	Brookings Institution, a research institution. In version 1's list.
bundesnetzagentur.de	Bundesnetzagentur, Germany's federal network regulator. An official source.
bundesregierung.de	The German federal government. An official source.
ca.gov	California state government, including the Public Utilities Commission and the Energy Commission. An official source.
cam.ac.uk	University of Cambridge. In version 1's list.
cbo.gov	US Congressional Budget Office. An official source.
cdc.gov	US Centers for Disease Control and Prevention. In version 1's list.
census.gov	US Census Bureau. In version 1's list.
cfr.org	Council on Foreign Relations, a research institution. In version 1's list.
chathamhouse.org	Chatham House, a research institution. From the credibility seed.
commerce.gov	US Department of Commerce. In version 1's list.
congress.gov	The official record of the US Congress, from the Library of Congress. In version 1's list.
corriere.it	Corriere della Sera, an Italian newspaper of record. From the credibility seed.
cru.ie	Commission for Regulation of Utilities, Ireland's energy regulator. An official source.
csis.org	Center for Strategic and International Studies, a research institution. From the credibility seed.
ct.gov	Connecticut state government. An official source.
cyber.gov.au	Australian Cyber Security Centre, a government agency. An official source.
defense.gov	US Department of Defense. An official source.
doc.gov	US Department of Commerce, including its bureaus' sites. An official source.
ec.europa.eu	European Commission. In version 1's list.
ecb.europa.eu	European Central Bank. In version 1's list.
economie.gouv.fr	French Ministry of the Economy. An official source.
economist.com	The Economist, a news magazine with original reporting. In version 1's list.
edb.gov.sg	Singapore Economic Development Board, a government agency. An official source.
eia.gov	US Energy Information Administration, the federal energy statistics agency. In version 1's list.
elpais.com	El País, a Spanish newspaper of record. From the credibility seed.
elysee.fr	Office of the President of France. An official source.
energy.gov	US Department of Energy. In version 1's list.
epa.gov	US Environmental Protection Agency. In version 1's list.
eur-lex.europa.eu	EUR-Lex, the official database of EU law. In version 1's list.
europa.eu	The European Union's official domain, covering its institutions and agencies. In version 1's list.
europarl.europa.eu	European Parliament. In version 1's list.
factcheck.org	FactCheck.org, a fact-checking organisation. From the credibility seed.
fcc.gov	US Federal Communications Commission, a regulator. An official source.
federalregister.gov	Federal Register, the US government's official journal. In version 1's list.
federalreserve.gov	US Federal Reserve Board. In version 1's list.
ferc.gov	Federal Energy Regulatory Commission, the US regulator of wholesale electricity and transmission. In version 1's list.
frbsf.org	Federal Reserve Bank of San Francisco, part of the US Federal Reserve System. An official source.
ft.com	Financial Times, a newspaper with original reporting. In version 1's list.
ftc.gov	US Federal Trade Commission, a regulator. An official source.
fullfact.org	Full Fact, a fact-checking organisation. From the credibility seed.
ga.gov	Georgia state government, including the Public Service Commission. An official source.
gao.gov	US Government Accountability Office. In version 1's list.
georgia.gov	Georgia state government. An official source.
gouvernement.fr	The French government. An official source.
gov.ie	The Government of Ireland's official site. An official source.
gov.sg	The Government of Singapore's official site. An official source.
gov.uk	The UK government's official site. In version 1's list.
govinfo.gov	The US Government Publishing Office's official documents. In version 1's list.
harvard.edu	Harvard University. In version 1's list.
house.gov	US House of Representatives, its committees and members. An official source.
idpc.org.mt	Information and Data Protection Commissioner, Malta's data protection regulator. An official source.
iea.org	International Energy Agency, an intergovernmental organisation. In version 1's list.
illinois.gov	Illinois state government, including the Illinois Commerce Commission. An official source.
imda.gov.sg	Infocomm Media Development Authority, Singapore's digital regulator. An official source.
imf.org	International Monetary Fund, an intergovernmental organisation. In version 1's list.
in.gov	Indiana state government, including the Utility Regulatory Commission. An official source.
itu.int	International Telecommunication Union, a UN agency. An official source.
japantimes.co.jp	The Japan Times, a newspaper. From the credibility seed.
justice.gov	US Department of Justice. In version 1's list.
korea.kr	The Korean government's official news portal. An official source.
lbl.gov	Lawrence Berkeley National Laboratory, a US Department of Energy laboratory that publishes grid interconnection data. An official source.
legislation.gov.uk	The official text of UK legislation. In version 1's list.
lemonde.fr	Le Monde, a French newspaper of record. From the credibility seed.
loudoun.gov	Loudoun County, Virginia, county government. An official source.
louisiana.gov	Louisiana state government, including the Public Service Commission. An official source.
maryland.gov	Maryland state government, including the Public Service Commission. An official source.
meity.gov.in	India's Ministry of Electronics and Information Technology. An official source.
meti.go.jp	Japan's Ministry of Economy, Trade and Industry. An official source.
michigan.gov	Michigan state government, including the Public Service Commission. An official source.
mit.edu	Massachusetts Institute of Technology. In version 1's list.
motie.go.kr	Korea's Ministry of Trade, Industry and Energy. An official source.
mro.net	Midwest Reliability Organization, a regional entity enforcing NERC reliability standards. An official source.
msit.go.kr	Korea's Ministry of Science and ICT. An official source.
nature.com	Nature and its peer-reviewed journals. In version 1's list.
nerc.com	North American Electric Reliability Corporation, the reliability regulator certified by FERC. An official source.
nhk.or.jp	NHK, Japan's public broadcaster. From the credibility seed.
nist.gov	US National Institute of Standards and Technology. In version 1's list.
nj.gov	New Jersey state government, including the Board of Public Utilities. An official source.
npcc.org	Northeast Power Coordinating Council, a regional entity enforcing NERC reliability standards. An official source.
npr.org	NPR, US public radio with original reporting. In version 1's list.
nrc.gov	US Nuclear Regulatory Commission. In version 1's list.
nrel.gov	National Renewable Energy Laboratory, a US Department of Energy laboratory. An official source.
nsf.gov	US National Science Foundation, including its statistics centre. An official source.
ntia.gov	US National Telecommunications and Information Administration. In version 1's list.
ny.gov	New York state government, including the Public Service Commission. An official source.
nytimes.com	The New York Times, a newspaper with original reporting. In version 1's list.
oecd.ai	The OECD's AI Policy Observatory. An official source.
oecd.org	OECD, an intergovernmental organisation. In version 1's list.
ofgem.gov.uk	Ofgem, Great Britain's energy regulator. An official source.
ohio.gov	Ohio state government, including the Public Utilities Commission of Ohio. An official source.
oireachtas.ie	Houses of the Oireachtas, Ireland's parliament. An official source.
ontario.ca	Government of Ontario. An official source.
ox.ac.uk	University of Oxford. In version 1's list.
pa.gov	Pennsylvania state government, including the Public Utility Commission. An official source.
parliament.uk	UK Parliament. In version 1's list.
pbs.org	PBS, US public broadcaster. From the credibility seed.
pewresearch.org	Pew Research Center, survey research. In version 1's list.
pib.gov.in	Press Information Bureau, the Government of India's official news service. An official source.
politifact.com	PolitiFact, a fact-checking organisation. From the credibility seed.
pubmed.ncbi.nlm.nih.gov	PubMed, the US National Library of Medicine's index of biomedical research. In version 1's list.
quebec.ca	Government of Quebec. An official source.
rand.org	RAND Corporation, a research institution. In version 1's list.
regulations.gov	Regulations.gov, the US federal rulemaking docket. In version 1's list.
reuters.com	Reuters, a news agency with original reporting. In version 1's list.
rfirst.org	ReliabilityFirst, a regional entity enforcing NERC reliability standards. An official source.
sam.gov	SAM.gov, the US government's official system for federal contract opportunities and awards. An official source.
science.org	Science and its peer-reviewed journals. In version 1's list.
sciencedirect.com	ScienceDirect, Elsevier's peer-reviewed journals. In version 1's list.
sec.gov	US Securities and Exchange Commission, including company filings. In version 1's list.
senate.gov	US Senate, its committees and members. An official source.
serc.org	SERC Reliability Corporation, a regional entity enforcing NERC reliability standards. An official source.
snopes.com	Snopes, a fact-checking organisation. From the credibility seed.
speaker.gov	Office of the Speaker of the US House of Representatives. An official source.
spiegel.de	Der Spiegel, a German news magazine. From the credibility seed.
stanford.edu	Stanford University. In version 1's list.
state.gov	US Department of State. In version 1's list.
texas.gov	Texas state government, including the Public Utility Commission of Texas. An official source.
texasre.org	Texas Reliability Entity, a regional entity enforcing NERC reliability standards. An official source.
theguardian.com	The Guardian, a newspaper with original reporting. In version 1's list.
thetimes.co.uk	The Times, a UK newspaper of record. From the credibility seed.
transportation.gov	US Department of Transportation. An official source.
treasury.gov	US Department of the Treasury. In version 1's list.
un.org	United Nations. In version 1's list.
usgs.gov	US Geological Survey, a federal agency. An official source.
ustr.gov	Office of the US Trade Representative. An official source.
virginia.gov	Virginia state government, including the State Corporation Commission and the legislature. An official source.
washingtonpost.com	The Washington Post, a newspaper with original reporting. In version 1's list.
wecc.org	Western Electricity Coordinating Council, a regional entity enforcing NERC reliability standards. An official source.
whitehouse.gov	The White House. In version 1's list.
who.int	World Health Organization. In version 1's list.
wi.gov	Wisconsin state government, including the Public Service Commission. An official source.
worldbank.org	World Bank, an international financial institution. In version 1's list.
wsj.com	The Wall Street Journal, a newspaper with original reporting. In version 1's list.
wto.org	World Trade Organization, an intergovernmental organisation. From the credibility seed.
zeit.de	Die Zeit, a German newspaper. From the credibility seed.

Graded medium: 194 entries

Entries graded medium, in order of domain.
Domain	Why it has this grade
abc.net.au	ABC, Australia's public broadcaster. In version 1's list at medium, which is kept, though the credibility seed rates it high.
abc.xyz	Alphabet's investor relations: a primary source for what the company reports, and an interested one.
abcnews.go.com	ABC News, US broadcast news. In version 1's list.
about.fb.com	Meta's newsroom: a primary source for what the company reports, and an interested one.
about.meta.com	Meta's corporate site: a primary source for what the company reports, and an interested one.
aboutamazon.com	Amazon's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
aep.com	American Electric Power's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
aeso.ca	Alberta Electric System Operator, reporting on its own grid.
agenceurope.eu	Agence Europe, EU affairs news agency with original reporting.
ai.meta.com	Meta's AI blog: a primary source for what the company reports, and an interested one.
aleph-alpha.com	Aleph Alpha's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
aljazeera.com	Al Jazeera, international news. In version 1's list.
alliantenergy.com	Alliant Energy's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
amd.com	AMD's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
ameren.com	Ameren's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
amkor.com	Amkor's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
anthropic.com	Anthropic's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
appliedmaterials.com	Applied Materials' own newsroom and investor relations: a primary source for what the company reports, and an interested one.
aps.com	Arizona Public Service's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
arm.com	Arm's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
arstechnica.com	Ars Technica, technology press. In version 1's list.
aseglobal.com	ASE's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
asml.com	ASML's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
aws.amazon.com	Amazon Web Services' own site and blogs: a primary source for what the company reports, and an interested one.
axios.com	Axios, news with original reporting. In version 1's list.
barrons.com	Barron's, financial news from Dow Jones. From the credibility seed.
blog.google	Google's own blog and newsroom: a primary source for what the company reports, and an interested one.
bloomberglaw.com	Bloomberg Law, legal and regulatory news from Bloomberg with original reporting.
bostonglobe.com	The Boston Globe, a newspaper. In version 1's list at medium, which is kept, though the credibility seed rates it high.
breakingdefense.com	Breaking Defense, defence trade press with original reporting.
broadcom.com	Broadcom's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
businessinsider.com	Business Insider, business news. In version 1's list.
caiso.com	California ISO, the California grid operator, reporting on its own grid and decisions.
canarymedia.com	Canary Media, energy trade press with original reporting.
cbsnews.com	CBS News, US broadcast news. In version 1's list.
chicagotribune.com	Chicago Tribune, a newspaper. In version 1's list.
cloud.google.com	Google Cloud's own site and blog: a primary source for what the company reports, and an interested one.
cnbc.com	CNBC, business news. In version 1's list.
cnet.com	CNET, technology press. From the credibility seed.
cnn.com	CNN, cable news. In version 1's list.
cohere.com	Cohere's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
comed.com	ComEd's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
computerweekly.com	Computer Weekly, enterprise technology trade press with original reporting.
computerworld.com	Computerworld, enterprise technology trade press with original reporting.
constellationenergy.com	Constellation Energy's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
consumersenergy.com	Consumers Energy's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
coreweave.com	CoreWeave's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
crn.com	CRN, technology channel trade press with original reporting.
dailymail.co.uk	Daily Mail, a UK newspaper. In version 1's list.
datacenterdynamics.com	DatacenterDynamics, data centre trade press with original reporting.
datacenterfrontier.com	Data Center Frontier, data centre trade press with original reporting.
datacenterknowledge.com	Data Center Knowledge, data centre trade press with original reporting.
defensescoop.com	DefenseScoop, defence technology trade press with original reporting.
denverpost.com	The Denver Post, a newspaper. In version 1's list.
dominionenergy.com	Dominion Energy's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
dteenergy.com	DTE Energy's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
duke-energy.com	Duke Energy's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
dw.com	Deutsche Welle, Germany's international broadcaster. In version 1's list at medium, which is kept, though the credibility seed rates it high.
eenews.net	E&E News, energy and environment policy press with original reporting.
eetimes.com	EE Times, electronics industry trade press with original reporting.
eirgrid.ie	EirGrid, Ireland's grid operator, reporting on its own grid.
eirgridgroup.com	EirGrid Group, Ireland's grid operator, reporting on its own grid.
energy-storage.news	Energy-Storage.news, storage trade press with original reporting.
engadget.com	Engadget, technology press. From the credibility seed.
engineering.fb.com	Meta's engineering blog: a primary source for what the company reports, and an interested one.
enr.com	Engineering News-Record, construction trade press with original reporting.
entergy.com	Entergy's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
ercot.com	ERCOT, the Texas grid operator, reporting on its own grid and decisions.
euractiv.com	Euractiv, EU policy news. In version 1's list.
euronews.com	Euronews, European broadcast news. From the credibility seed.
evergy.com	Evergy's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
exeloncorp.com	Exelon's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
fedscoop.com	FedScoop, US government technology trade press with original reporting.
fierce-network.com	Fierce Network, telecoms and cloud trade press with original reporting.
firstenergycorp.com	FirstEnergy's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
forbes.com	Forbes, business news. In version 1's list.
foreignaffairs.com	Foreign Affairs, a policy journal. From the credibility seed.
foreignpolicy.com	Foreign Policy, a policy magazine. From the credibility seed.
fortune.com	Fortune, business news. From the credibility seed.
foxnews.com	Fox News, cable news. In version 1's list.
fpl.com	Florida Power & Light's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
france24.com	France 24, France's international news channel. In version 1's list at medium, which is kept, though the credibility seed rates it high.
georgiapower.com	Georgia Power's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
gf.com	GlobalFoundries' own newsroom and investor relations: a primary source for what the company reports, and an interested one.
googleblog.com	Google's product and research blogs: a primary source for what the company reports, and an interested one.
govtech.com	Government Technology, trade press on public-sector technology. In version 1's list.
heatmap.news	Heatmap, energy and climate press with original reporting.
heise.de	heise online, German technology press with original reporting.
hindustantimes.com	Hindustan Times, an Indian newspaper. In version 1's list.
hpcwire.com	HPCwire, high-performance computing trade press with original reporting.
huffpost.com	HuffPost, news and opinion. In version 1's list.
iapp.org	IAPP, the privacy professionals' association, and its news service. In version 1's list.
ieso.ca	Independent Electricity System Operator, Ontario's grid operator, reporting on its own grid.
independent.co.uk	The Independent, a UK newspaper. In version 1's list at medium, which is kept, though the credibility seed rates it high.
insidecybersecurity.com	Inside Cybersecurity, US cybersecurity policy trade press with original reporting.
intel.com	Intel's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
investopedia.com	Investopedia, financial explainers. In version 1's list.
investor.atmeta.com	Meta's investor relations: a primary source for what the company reports, and an interested one.
iso-ne.com	ISO New England, the New England grid operator, reporting on its own grid and decisions.
kla.com	KLA's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
lamresearch.com	Lam Research's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
latimes.com	Los Angeles Times, a newspaper. In version 1's list at medium, which is kept, though the credibility seed rates it high.
latitudemedia.com	Latitude Media, energy transition trade press. In version 1's list.
lawfaremedia.org	Lawfare, national security law and policy. In version 1's list.
lightreading.com	Light Reading, telecoms trade press with original reporting.
marketwatch.com	MarketWatch, business news from Dow Jones. In version 1's list.
marvell.com	Marvell's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
miamiherald.com	Miami Herald, a newspaper. From the credibility seed.
micron.com	Micron's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
microsoft.com	Microsoft's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
misoenergy.org	MISO, the grid operator for the US Midwest and South, reporting on its own market and decisions.
mistral.ai	Mistral AI's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
mlex.com	MLex, regulatory and competition news with original reporting.
msnbc.com	MSNBC, cable news. In version 1's list.
nationalgrid.com	National Grid's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
nbcnews.com	NBC News, US broadcast news. In version 1's list.
neso.energy	National Energy System Operator, Great Britain's system operator, reporting on its own grid.
networkworld.com	Network World, networking trade press with original reporting.
newscientist.com	New Scientist, a science magazine. From the credibility seed.
newyorker.com	The New Yorker, a magazine. In version 1's list.
nexteraenergy.com	NextEra Energy's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
nextgov.com	Nextgov/FCW, US government technology trade press with original reporting.
nextplatform.com	The Next Platform, computing infrastructure trade press with original reporting.
nrg.com	NRG Energy's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
nvenergy.com	NV Energy's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
nvidia.com	NVIDIA's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
nyiso.com	New York ISO, the New York grid operator, reporting on its own grid and decisions.
nzherald.co.nz	New Zealand Herald, a newspaper. In version 1's list.
oncor.com	Oncor's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
openai.com	OpenAI's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
oracle.com	Oracle's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
pge.com	Pacific Gas and Electric's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
pjm.com	PJM Interconnection, the grid operator for the US mid-Atlantic and Midwest, reporting on its own market and decisions.
politico.com	Politico, policy news with original reporting. In version 1's list.
politico.eu	Politico Europe, EU policy news with original reporting. In version 1's list.
powermag.com	POWER magazine, power generation trade press with original reporting.
pplweb.com	PPL Corporation's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
pv-magazine-usa.com	pv magazine USA, solar and storage trade press with original reporting.
pv-magazine.com	pv magazine, solar and storage trade press with original reporting.
qualcomm.com	Qualcomm's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
rdworldonline.com	R&D World, research and development trade press with original reporting.
renewableenergyworld.com	Renewable Energy World, energy trade press with original reporting.
rtoinsider.com	RTO Insider, trade press on US grid operators and wholesale power markets. In version 1's list.
salon.com	Salon, news and opinion. In version 1's list.
samsung.com	Samsung's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
scientificamerican.com	Scientific American, a science magazine. From the credibility seed.
sdxcentral.com	SDxCentral, networking and cloud trade press with original reporting.
searchengineland.com	Search Engine Land, search industry trade press with original reporting.
seattletimes.com	The Seattle Times, a newspaper. From the credibility seed.
seekingalpha.com	Seeking Alpha, investment analysis and news. In version 1's list.
semafor.com	Semafor, news with original reporting. In version 1's list.
semiengineering.com	Semiconductor Engineering, chip industry trade press with original reporting.
sempra.com	Sempra's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
servethehome.com	ServeTheHome, server and data centre hardware trade press with original reporting.
sfchronicle.com	San Francisco Chronicle, a newspaper. In version 1's list.
siliconangle.com	SiliconANGLE, enterprise technology press with original reporting.
sky.com	Sky News, UK broadcast news. In version 1's list.
slate.com	Slate, a magazine. In version 1's list.
smh.com.au	The Sydney Morning Herald, a newspaper. In version 1's list at medium, which is kept, though the credibility seed rates it high.
southerncompany.com	Southern Company's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
spglobal.com	S&P Global, including its commodity and market news services. In version 1's list.
spp.org	Southwest Power Pool, a US grid operator, reporting on its own market and decisions.
straitstimes.com	The Straits Times, a Singapore newspaper. In version 1's list at medium, which is kept, though the credibility seed rates it high.
talenenergy.com	Talen Energy's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
tdworld.com	T&D World, electricity transmission and distribution trade press with original reporting.
techcrunch.com	TechCrunch, technology press. In version 1's list.
techpolicy.press	Tech Policy Press, technology policy news. In version 1's list.
techtarget.com	TechTarget, enterprise technology trade press with original reporting.
telegraph.co.uk	The Telegraph, a UK newspaper. In version 1's list at medium, which is kept, though the credibility seed rates it high.
theatlantic.com	The Atlantic, a magazine. In version 1's list.
theaustralian.com.au	The Australian, a newspaper. In version 1's list at medium, which is kept, though the credibility seed rates it high.
thehill.com	The Hill, US policy news. In version 1's list.
theinformation.com	The Information, technology business press with original reporting.
thenextweb.com	The Next Web, technology press. From the credibility seed.
therealdeal.com	The Real Deal, real estate trade press with original reporting.
theregister.com	The Register, technology trade press. In version 1's list.
theverge.com	The Verge, technology press. In version 1's list.
timesofindia.indiatimes.com	The Times of India, a newspaper. In version 1's list.
tomshardware.com	Tom's Hardware, computer hardware trade press with original reporting.
trendforce.com	TrendForce, a market research firm that publishes its own estimates of memory and foundry markets.
tva.com	Tennessee Valley Authority, a federally owned utility, reporting on its own business: a primary source, and an interested one.
usatoday.com	USA Today, a newspaper. In version 1's list.
utilitydive.com	Utility Dive, utility and power sector trade press. In version 1's list.
venturebeat.com	VentureBeat, technology press. From the credibility seed.
vice.com	Vice, news and culture. In version 1's list.
vistracorp.com	Vistra's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
vox.com	Vox, news and explainers. In version 1's list.
wecenergygroup.com	WEC Energy Group's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
windows.com	Microsoft's Windows blog: a primary source for what the company reports, and an interested one.
wired.com	Wired, technology press. In version 1's list.
x.ai	xAI's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
xcelenergy.com	Xcel Energy's own newsroom and investor relations: a primary source for what the company reports, and an interested one.
yahoo.com	Yahoo, a news portal. In version 1's list.
zdnet.com	ZDNET, technology press. From the credibility seed.

21 more entries are not listed on this page, because this record’s publication filter keeps certain topics off its pages. They are part of the table and of its fingerprint. The same table, with the same fingerprint, is listed whole on the method page of Microsoft (version 2).

Publishers and copies

A development moves an assessment only once two independent publishers have reported it within seven days of its first report. A publisher is a registered domain: reports from news.example.com and example.com come from one publisher. The table below also names groups of domains that belong to one owner, and each group counts as one publisher.

A report that copies another counts with the report it copies. A copy is a report with the same headline, once letter case, punctuation and spacing are set aside, or one whose text the update loop measures as all but identical. Each change lists the publishers it counted and the copies it folded, so the count can be checked from the record.

Domains the table names as one owner:

tokenpost.com, tokenpost.kr
microsoft.com, windows.com
google.com, blog.google, googleblog.com, abc.xyz
amazon.com, aboutamazon.com
fb.com, meta.com, investor.atmeta.com, facebook.com
bloomberg.com, bloomberglaw.com, bloombergtax.com, bnef.com
wsj.com, dowjones.com, marketwatch.com, barrons.com
politico.com, politico.eu, eenews.net
yahoo.com, aol.com

What the record already knew

Developments reported before the record began

Point 3 of the rule above says what the record already knew when it began: a development none of whose reports had been retrieved when the starting point was read is new evidence, whatever date it carries. The first day records when the starting point was read, and the fingerprint of the web address of every report it was shown, so the daily check can refuse a change on a development the record already knew.

Limits of the scope

What the scope keeps this record from testing

This record can rarely test one of the rival explanations on the main assessment, MX1, that the supply of advanced packaging and high-bandwidth memory sets the pace. The record's scope leaves out reports that name the companies that make most of both, and the filter reads a report's summary, web address and outlet as well as its headline. So most reports that could favour MX1 are excluded before anything reads them, and the record can rarely see evidence in its favour.

Starting points

Every starting point read before the record began

Before the record began, the update loop read a starting point once a day. The first day keeps the fingerprint of every reading made before the one it sealed, so a reading made again and replaced would show. The sealed reading, with every reading and check in it, is on each assessment’s page.

The sealed reading was made on 15 October 2026, 04:10 UTC.

No starting point was read before it.

Quality monitors

How the record's own behaviour is watched

Each monitor is measured from the sealed record alone and published whatever its value, with its n and the dates its data span.

The rates cover the last 28 days.

Reversals within 7 days
Not yet measured.
Moves the challenge step disputed
Not yet measured.
Restatements
Not yet measured.
Evidence set aside
Not yet measured.
Possible echo
An assessment is flagged when straining evidence stays under 10 per cent of its evidence in each of the last 3 complete weeks. A record that only ever hears agreement may be hearing one voice.
- Main assessment: not yet measured, as the record does not yet cover 3 complete weeks.
- The condition it rests on: not yet measured, as the record does not yet cover 3 complete weeks.
- The alternative: not yet measured, as the record does not yet cover 3 complete weeks.

Scope

What the judgment covers, and what it leaves out

Through the end of 2027, what sets the pace at which AI data centres are built and connected to power in the United States? Geography: United States. Horizon: 31 December 2027.

In scope:

Connection of large new loads to the US grid, and the queues, studies and network upgrades that connection requires
How utilities and state regulators treat large new loads, including special tariffs, deposits and pauses on new connections
Lead times for the equipment that delivers power to a campus, such as large transformers, switchgear and gas turbines
Capital spending by the largest US cloud companies, as filed with the Securities and Exchange Commission
Debt, lease and joint-venture financing for data centre campuses
Signs that demand for AI computing, rather than power, sets the pace, such as slowing capital spending, cancellations that cite demand and falling rental prices for accelerators

Out of scope:

Data centre buildout outside the United States
Cross-border trade and technology policy
The capabilities of AI models and questions of AI safety
Investment advice about any company

Reports on topics the project keeps out of this record are excluded before they can move anything, and the public pages refuse to show a record that carries one. Only the count is published: no excluded report is listed, linked or quoted anywhere on these pages.

Tests we're watching

When a data series may count as a test

No data series is admitted as a test in this version of the rule. A later version may admit one only if, over all the readings it will get, its chance of firing on noise alone is 5 per cent or less. Hourly regional electricity demand, for example, fails that bar, because weather and the daily cycle move it more than any buildout would. The arithmetic and a replay are on the page When a resolution criterion is a real test.

Fingerprints

How records are fingerprinted

One daily state is recorded for each day in UTC, and one record for each change. Each names the one before it by its fingerprint. That link fixes the order of the records, but it does not by itself stop a record being replaced, because a replacement chain can be built to link just as well. What does is publication outside this site. At 23:59 UTC every such record not yet fingerprinted joins that day’s fingerprint of all of them, and the day’s fingerprint is pushed to a public repository on GitHub. When a day is missed, its records join the next day’s, and the pages give the date of the fingerprint that contains each record.

The first daily state and every new change are also published at once. As soon as one is recorded, the newest fingerprint of each of the two chains is pushed to the same repository, in a folder of its own, so the wait is minutes rather than until 23:59 UTC. If that push fails, the record waits for the daily one. Until its fingerprint is published one way or the other, a record could still be replaced without trace. From then on a replacement would no longer match what GitHub holds.

The record is published as data: the current state, the changes and every daily state, the last two as JSON Lines. Each record comes in full with its fingerprint, so anyone can recompute it. Once its day’s fingerprint is published, it also comes with the path to that fingerprint and the commit on GitHub, so it can be checked offline. The project’s sealed protocols and pre-registrations are listed in the methods ledger.

What we do not claim

No accuracy, skill, calibration, Brier score, hit rate or track record of any kind is claimed. Nothing here predicts events, and nothing here is said to beat experts or markets.
The numbers are not measured probabilities. They are stated judgments. A model set each starting confidence 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 the record began, after a second model had checked the developments the readings relied on. Since then only a published rule has revised them.
The three stated confidences are set and moved separately. They are not shares of one whole, and they need not add up to 100 per cent.
The system did not choose the question. The thesis, its assessments and the rule were fixed before the record began, and the system works within them.
Nothing is learned from outcomes. There is no learned model of source quality and no reliability score for the challenge step.
The record is not real-time: it is updated once a day. It is not full-text: the update loop reads headlines and summaries. It is not comprehensive: it covers one slice of the United States, drawn from a web-search service and public datasets. It does not establish causation.
A timestamped fingerprint does not show that a judgment was right. It shows that the record existed no later than the day its fingerprint was published. When a day's fingerprint is missed, that day's records join the next day's, and the pages give the date of the fingerprint that contains each record.
Nothing is reconstructed after the fact. The only material from before the first day that is part of the record is the starting point: the developments it was shown, every reading and every check, and the fingerprint of every earlier reading, kept in the first day's record. The update loop's working notes are shown as it recorded them, and they are not part of the fingerprinted record.