Accountability

Corrections and version history

FAIR Labs reports are revised when the evidence changes. This page records every substantive revision, with its date, so that a reader can see exactly what a report said when they read it.

The policy. A revision is substantive if it alters a finding, adds or withdraws evidence, or changes a figure the report relies on. Every substantive revision gets a version number, a date, and an entry here and in the report's own back matter. Copy-editing, wording and layout revisions are not logged and never alter a finding. The unversioned address of a report always serves the current text; from version 1.1 onward each superseded edition stays permanently available at its own address, so anything cited from an earlier version can still be retrieved and checked. If you believe something in a FAIR Labs report is wrong, write to us through the contact form and we will check it and, if it is wrong, correct it here.

The Backlash Against Data Centers

FAIR Labs Frontier Series · Special Report · FL-2026-SR2 · current version 1.1, September 8, 2026 · current PDF

VersionDateTypeChange
1.0Sept 4, 2026First publicationPublished at fairlabs.ai. 47 pages, 212 sources.
1.1Sept 8, 2026New evidenceSection 02 gained the summer's foreign-funding paper trail: Senator Cotton's June 10 letter to the Justice Department and the European provenance of the "$2 billion" it cites, Representative Gooden's July 21 letter, the Ways and Means subpoenas of July 24, the absence of the claim from the 2026 Annual Threat Assessment, the absence of any foreign-agent registration or charge, and statements by President Trump and Senator Ron Johnson. Exhibit S.3 gained two rows. Eight sources added. Download v1.1

Version 1.0 was not separately archived, because this policy postdates it.

The Open Frontier

FAIR Labs Frontier Series · Special Report · FL-2026-SR1 · current version 1.1, September 8, 2026 · current PDF

VersionDateTypeChange
1.0July 20, 2026First publicationPublished at fairlabs.ai.
1.1Sept 8, 2026CorrectionExhibit S.3 described Claude Fable 5 and GPT-5.6 Sol as settled shipped products. Both were newly available. Fable 5 was released on 9 June, withdrawn from every customer between 12 and 30 June under a US export directive, and restored on 1 July; GPT-5.6 reached general availability on 9 July after a vetted limited preview announced on 26 June. The rows, the exhibit caption and a new paragraph in Section 02 now carry those dates. Four sources added. Download v1.1
1.1Sept 8, 2026CorrectionSection 03 put frontier capability rented through a proprietary API at fifteen to twenty dollars per million output tokens. The rate at the time was thirty to fifty: fifty for Fable 5, thirty for GPT-5.6 Sol. The passage and Exhibit S.4 now state the correct range, give Kimi K3's published rates of three dollars input and fifteen output, and carry two cautions the report had omitted: that per-token price is not per-task cost because K3 runs at maximum reasoning effort by default, and that the launch coverage, pricing a blended token, put K3 at roughly twelve dollars per million and called it no bargain. Download v1.1
1.1Sept 8, 2026CorrectionThe 45 percent figure for Chinese-developed models was described as a share of enterprise API traffic. It is a share of tokens routed on OpenRouter, a developer platform, and the underlying series runs from a 4.5 percent average in the first half of 2025. Corrected in the key points, the statistics band, Section 05 and Exhibit S.6. Download v1.1
1.1Sept 8, 2026EditorialThree citations that pointed at publication landing pages now point at the articles themselves; publication dates were added to six source entries.

Version 1.0 was not separately archived, because this policy postdates it.

Other reports. Version histories for the other reports in the series appear here as they are published.