SecondSourceJudgment rebuilt from primary sources
Claims & Receipts · Nov 15, 2026

Checking the answer: will Google sign an outside TPU long-term deal with a named customer?

From the Nov 15, 2026 daily brief

This issue checks no new claim. It goes back to three calls SecondSource made earlier and asks whether they held up now that they have come due.

Will Google sign an outside TPU long-term deal with a named customer?

① What the call said then: SecondSource's standing judgment is that whether TPUs, Google's in-house AI chips, take share from NVIDIA turns on externalization, meaning Google supplying TPUs over the long term to customers outside Google, not on price-performance. On September 7, the semiconductor research firm SemiAnalysis said on X that TPUs deliver better inference performance per dollar than NVIDIA's chips. The question SecondSource posed on September 13 was "will Google sign a verifiable outside TPU long-term deal (a named customer, with capacity or term that can be checked)", due October 31. No odds were recorded at the time.

② What the record shows: the call was set up in advance to count only official announcements and statutory filings from Google and the customer. Page 170 of the registration statement Anthropic filed for its listing on October 26 states that Anthropic has committed to buy additional TPU compute from suppliers that include Google, for delivery from 2028 to 2031, with estimated payments of US$125 billion to US$150 billion. That figure is a total across several TPU suppliers. The filing does not break out Google's share.

③ Ruling, and at which layer: yes, it came true, but the starting point SecondSource wrote at the time was wrong. Anthropic had already announced on October 23, 2025 that it would use up to 1 million Google TPUs. That was a one-sided announcement from Anthropic and states a ceiling, not the purchase commitment in the filing, but it points the same way. When SecondSource posed the question on September 13, it described this kind of long-term deal as having no precedent. That was wrong. The error sits at the data layer: an announcement made public a year earlier was missed, and the reasoning behind the call was sound. The filing shows only that Google is already supplying outside customers over the long term. It does not show that this began in September or October.

④ Which tracked judgments moved: none yet. The main judgment holds that whether TPUs can take share from NVIDIA depends on whether Google is willing to supply TPUs to outside customers over the long term, not on whether the chips themselves are good value. The answer this time was yes, and the judgment still does not move, because this piece of evidence cannot separate two explanations. If Google was always willing to supply outsiders, customers would sign long-term deals. If TPUs are simply good value, customers would sign long-term deals too. Strengthening the judgment would take third-party market-share figures showing share moving with the long-term deals. Weakening it would take share actually moving for reasons that are plainly price-performance. SecondSource set both of these conditions. The other two judgments have not moved either: one holds that cloud providers' in-house chips will keep expanding; the other tracks SemiAnalysis's performance-per-dollar claim. That claim rests on a single SemiAnalysis post, and no one else has reproduced the test method.

⑤ Does the test change: no. The conditions were clearly written and could be ruled on. What was missed was the check back on an announcement made public a year earlier, when the question was posed on September 13; the test itself was not at fault.

Check the call →

Will a frontier lab publish the absolute layer count of its flagship model?

① What the call said then: Geoffrey Irving, chief scientist at the UK AI Security Institute, argued on September 2 that there is no need to wait for legislation: an independent person could be allowed to look at a frontier model's architecture now. The test Irving himself set calls for absolute numbers. OpenAI's Pachocki gave only a relative statement, "within 2x of GPT-4", and no actual figure. The question SecondSource posed on September 6 was whether any of OpenAI, Anthropic or Google DeepMind would, before November 1, publish the absolute computational-graph depth or layer count of its current flagship (how many layers of computation the model chains from input to output), or have a named independent party vouch for it. No odds were recorded at the time.

② What the record shows: the call was set up in advance to count only the three labs' system cards, the capability and safety documents published when a model ships, and reports from named independent parties. None of the OpenAI GPT-6 Astra system card, the Anthropic Claude Fable 5.1 and Mythos 5.1 system card or the Google Gemini 4 Argon model card gives a layer count, and the UK AI Security Institute's signed report carries no architecture figures either. Anthropic's October 15 update also states that even the access given to METR, an outside evaluator that tests frontier models, excludes model architecture information.

③ Ruling, and at which layer: no, it didn't come true. None of the three labs' public documents gives an absolute layer count. What Irving asked for, an independent person looking at the architecture, has not happened in current public disclosures. This result gives no direct evidence on whether an independent person can see the architecture privately. The case agreed in advance that would flip the answer was "an independent party could already see the architecture", and no evidence of that turned up.

④ Which tracked judgments moved: none of the four related judgments. Three of them are statements industry figures have already made, and what is tracked is whether those statements hold up, not whether the people made them: Irving's argument that an independent person could be allowed to see the architecture now; the claim that GPT-6 Astra's written reasoning has become harder to monitor from outside, where the cause is disputed in the industry; and, after a message-board incident, the observation from people in the industry that the safeguards did not catch it and the model did not report it. This result does not bear on any of the three, so none moves. The fourth is SecondSource's own judgment about that same message-board incident: the AI safety community calls it a detection failure, the security community calls it an authorization-design error, and the two framings send money to different places. With none of the three labs disclosing architecture, both framings still fit, so this result cannot tell which is right. Moving the judgment either way would take OpenAI disclosing how its permissions are designed and how its monitoring actually works.

⑤ Does the test change: no.

Check the call →

Is the share of general-purpose enterprise AI spending going to standalone platforms that let employees build their own agents still small? Can't judge. The call bet that in Menlo Ventures' annual enterprise AI report, the venture firm's yearly survey of corporate AI spending, these platforms would take under 25% of "horizontal AI" spending (general AI tools usable across every department); the 2025 report put it at 10%. Menlo Ventures' 2026 enterprise AI report, published November 3, changed its categories and no longer lists "agent platforms" separately, so the share cannot be computed and, as agreed in advance, the call is not scored. Check the call →

Every call SecondSource has made, with its ruling date → the judgment index

One more step: confirm by email

A confirmation email is on its way to the address just entered; clicking the link inside completes the subscription. Not in the inbox? Check spam or Promotions. Wrong address? Enter it again

Subscribe free — first issue lands tomorrow morning

Just an email address, unsubscribe anytime. Subscribing finishes with one click on the link in the confirmation email; if it is not in the inbox, check spam or Promotions.

More in this section