The single number the AI trade is priced off got cut by about a third, and the cut turned out to be accounting rather than business. OpenAI's annualized revenue is near $50 billion, not the roughly $70 billion that had circulated; the higher figure was a gross-up built to make OpenAI comparable to Anthropic, which books the whole cloud-partner sale as revenue and expenses the provider's cut, while OpenAI books only its share. Nvidia, Oracle and CoreWeave sold off on it, hours after Air Street's ninth State of AI Report put the two labs' combined annualized run rate at $105 billion, up from about $30 billion at the start of the year. Both of those are true at once, which is the whole lesson: the business is compounding faster than any quarterly report captures, and the metric everyone tracks it with is one month multiplied by twelve. Meanwhile the US government spent the day being two institutions at once — a major new buyer of AI-for-science, with $2.4 billion of industry compute credits pledged into the Genesis Mission inside a $6 billion package, and its newest investigator, with the FTC's career attorneys drafting subpoena-force demands for the same labs. Nobody is going to win this on run-rate multiples. The returns are still in which old problems you point the intelligence at.

🌊 The Tide

No shift. One confirmation, and it is the cleanest governance-as-market-structure evidence logged since the late-September accord: in a single day the US state appeared on both sides of the same ledger — as a named buyer of AI-for-science at $2.4 billion across 15-plus agencies, and as the agency drafting civil investigative demands against the sellers. cost-collapse held: the State of AI Report's measurement — minimum inference cost for a fixed benchmark score falling 47% per quarter, roughly 13-fold a year, per Epoch AI — restates the trend rather than bending it, so it is carried in the waves. distribution-rewrite held without new evidence strong enough to log. ai-as-worker also held: the report's 26% model-R&D figure was already carried in the previous edition against Epoch's InnovationEval correction, and nothing published since changed that balance.

Governance-as-market-structure, confirmed from both sides of the same ledger in the same day

At the White House AI Science Summit, billed as a $6 billion push for US-led scientific innovation, eleven companies pledged $2.4 billion in compute credits to the Genesis Mission Consortium, which the White House says will support over 15 federal agencies working on National Science & Technology Challenges. The named commitments: Nvidia $1 billion, AMD $500 million, OpenAI $200 million, Anthropic $150 million, Google $150 million, AMP $100 million, Emerald AI $100 million, and Amazon Web Services, Armada, Crusoe and Micron at $50 million each. The Energy Department added $159 million in phase II Genesis Mission project awards, $96 million for early-career researchers and a new graduate fellowship program; the Pentagon added an enterprise-wide framework for measuring its labs and up to $350 million for quantum computing. Trump awarded the National Medal of Technology and Innovation and the National Medal of Science to six executives, Elon Musk among them. Hours earlier the same day, Semafor reported that the FTC is close to sending civil investigative demands — subpoena-force, and described by a senior agency official as dozens of questions long — to frontier AI companies, covering what those companies have said publicly about the dangers of their own products along with other consumer-harm issues. Career attorneys in the Bureau of Consumer Protection are drafting them. Anthropic and OpenAI were named when the probe first surfaced around September 30; the independent evaluator METR may also be a target. Anthropic declined to comment and METR did not respond; OpenAI's Chris Lehane said the company is "trying to find out more about that." Former Republican chair William Kovacic called the early disclosure unusual but not improper; former chair Lina Khan doubts chair Andrew Ferguson is serious about it. The probe became public less than a day after Ferguson appeared alongside Trump and AI chief executives to announce a nonbinding industry self-regulation accord. The mechanism worth naming: compute credits are procurement and investigative demands are discovery, and one institution is now running both against the same counterparties on different clocks.

So what: Two concrete moves. If you sell into federal science, the Genesis Mission Consortium has stopped being a press release and become a channel with named dollar commitments and 15-plus agencies behind it — build the data-handling and authorisation path now, not after the first RFP lands. And treat every public claim you make about your own product's capability, safety or risk as discoverable text: the reported CID question set is specifically about what companies have said about their own dangers, which means marketing copy and model cards are now the same document class as a disclosure.

Sources: Nextgov/FCW — White House unveils billions in investments for Genesis Mission · Semafor — FTC advancing on its AI safety probe · Yahoo News — Tech firms pledge $2.4B in computing power to Trump's Genesis Mission

🌊 The Waves

Medium-term motion — weeks to quarters. These should move roadmaps, budgets and vendor selection.

The two biggest revenue numbers in AI were published and corrected within hours of each other

Air Street's State of AI Report, now in its ninth edition, landed with OpenAI and Anthropic's combined annualized revenue run rate at $105 billion by late summer, up from about $30 billion at the start of the year. Later the same day the Financial Times reported that OpenAI's own annualized figure is closer to $50 billion — about $20 billion below the roughly $70 billion that had been widely reported, and some outlets had carried $68 billion. Axios has the reason, and it is not a business deterioration: Anthropic records the full value of a sale made through a cloud partner as revenue and books the partner's cut as an expense, while OpenAI records only its own share of certain partner sales, so the $70 billion was an attempt to gross OpenAI up to an apples-to-apples comparison. The market took it as news anyway. Nvidia, Oracle and CoreWeave shares fell on Thursday; one tracker had CoreWeave off about 7%, Nebius about 6% and Oracle about 5% intraday, in a session that tracker attributed to borrowing costs repricing debt-financed capacity rather than to OpenAI, while CNBC and Quartz tied the decline directly to the revenue news. The Information's Martin Peers, writing in The Briefing the same day, made the structural point plainly: annualized revenue is one month's revenue multiplied by twelve, the labs publish it in occasional blog posts rather than on any schedule, and much of what reaches reporters comes via investor communications. The rest of the report is worth the download for the same reason — it measures rather than asserts. Four US hyperscalers guide to roughly $733 billion of 2026 capex. Minimum inference cost for a fixed benchmark score is falling 47% per quarter, about 13-fold a year, per Epoch AI. Data Center Watch counted at least 45 projects, nearly $68 billion, blocked or delayed by local opposition in Q2, and Gallup found 71% of Americans opposed a local AI data center against 53% for a nuclear plant. Air Street scores its own prior-year predictions 2 hits, 5 partial, 3 misses, which is more honesty than most forecasters offer.

Roadmap implication: Stop underwriting anything off a counterparty's ARR without asking two questions: which month, and gross or net of channel. If you have multi-year compute or capacity commitments whose covenants reference a customer's revenue run rate, get the definition written into the document this quarter — Thursday showed that a definitional change alone can move the equity of your suppliers' suppliers by mid-single digits. For your own reporting, publish a trailing-twelve-month number alongside the annualized one; when the public comparison gets re-based, the firm with the duller, fuller disclosure is the one that does not have to explain itself.

Sources: Air Street Press — The State of AI Report 2026 · State of AI Report 2026 · Axios — OpenAI annualized revenue $20 billion less than previously reported · Quartz — OpenAI says its annualized revenue is $50 billion, not $68 billion · 24/7 Wall St. — CoreWeave Sinks 7% as Financed AI Buildout Names Sell Off Apart From Cloud Software; Nebius Drops 6%, Oracle Falls 5%

Google Cloud shipped a universal agent for work, and told buyers almost nothing they need to buy it

At Gemini at Work 2026, Google Cloud introduced the Gemini agent, described in its own announcement as "a universal agent for work." The design is the one the category is converging on: a single prompt box that takes knowledge work, question answering, content creation and coding; the agent draws on the organisation's business context, plans the work, uses skills and tools, connects to the customer's business systems, and returns finished output inside the documents, email and developer tools people already use. Two details are the differentiators Google chose to lead with — the agent selects a suitable model for each task rather than pinning one, and it ships with built-in cost controls alongside the usual enterprise security, administration and governance. That first one is the routing thesis arriving inside a vendor's own product: Google is now making the model choice on the customer's behalf, in the customer's tenancy, under a cost ceiling the customer sets. What the announcement does not contain is everything a procurement team needs: no availability date, no regions, no access tiers, no pricing, no named underlying model, no customer references and no performance numbers. For a product whose pitch is that it will run long tasks against your business systems, that absence is itself the information.

Roadmap implication: This is now a three-horse race for the enterprise agent layer, and all three are selling the same shape, so write your evaluation around the parts they are not advertising. Insist on three numbers before any pilot: cost per completed task rather than per token, task-completion reliability measured on your own work rather than a benchmark, and the blast radius when the agent is wrong. And since Google's own pitch is per-task model selection, require the routing log — if you cannot see which model did which job, you cannot reproduce an outcome, price it, or defend it later.

Sources: Google — Google Cloud introduces the Gemini agent · CNBC — Google Cloud introduces Gemini agent for work as AI race heats up

Somebody counted China's model safety disclosures, and the number is 3.6%

SemiAnalysis published a census of the thing everyone argues about without data: every model release by nine leading Chinese developers from 2021 to 15 September 2026 — 857 releases, 741 product models and 116 research models — scored on whether a published developer safety result exists. Thirty-one, or 3.6%, have one at all. Nine, 1.1%, had it at or before launch. Sixteen were documented only afterwards, at a median lag of 42 days and a maximum of 349 for DeepSeek-R1. Eight hundred and thirteen, 94.9%, have no safety disclosure of any kind. The remainder of the 31 is the six whose timing or model match could not be established; separately, and outside that 31, ten claimed an evaluation without publishing figures and three are known only from press or investor accounts. The report pairs that with a text analysis of 102 expert documents from 2023 to September 2026 — 51 official, Party-theory and legal texts against 51 technical and scientific ones, from the 32 most influential voices — and the split is almost clean. Frontier or loss-of-control risk appears in 86% of technical texts, 44% of public-policy texts, 23% of official texts and 22% of law texts. Development-first or competition framing appears in 100% of official texts, 91% of law texts, 89% of public-policy texts and 45% of technical texts. Of the 42 technical texts, 22 say safety must precede or gate development; of the 60 non-technical texts, not one takes that position. Thirteen of the 102 call for binding frontier duties, and none has been adopted in a binding Chinese instrument. The constructive reading: Chinese technical staff are saying the same things their Western counterparts say, and the gap is institutional rather than cultural — which is the kind of gap that negotiation can actually close. The report's forward marker is the second round of US-China AI dialogue; the date this brief would set beside it is the APEC leaders' meeting in Shenzhen on 18–19 November.

Roadmap implication: If your risk register models Chinese open-weight adoption, swap the assumption. The constraint is not that Beijing will slow the frontier on safety grounds; the evidence says it will not, and 94.9% of releases arrive with nothing to read. So the diligence has to be yours: for any Chinese open-weight model you put in production, budget for running your own evaluation suite before deployment and treat the absence of a developer safety card as the default rather than the exception. Build that eval capability as a standing asset — it is the same asset that lets you adopt those models faster than competitors who are waiting for a disclosure that is not coming.

Sources: SemiAnalysis — Beijing Will Not Pace the Frontier: China's Speed-First AI Safety Regime

Nvidia's challengers are being bought into the ecosystem rather than beaten in it

The Information reported, citing three people with knowledge of the deal, that Nvidia plans to invest in d-Matrix, which builds AI server chips that compete directly with Nvidia's own — framed as part of an effort to make Nvidia's technology compatible with rival chip designers and pull them into its hardware ecosystem. Terms were not clear at publication. On the same day, The Information's Dealmaker newsletter reported that Eliyan, the five-year-old electro-optical interconnect startup moving the links between chiplets and high-bandwidth memory beyond copper, told some investors in the last few days that it had received a takeover bid; valued at $1 billion in July, it has been working with an investment bank since the summer and is likely to aim for $3 billion. SoftBank-owned Arm, an existing investor, held early talks but did not bid and would be unlikely to at that valuation. Eliyan has raised more than $250 million from AMD, Cisco and Meta alongside Seligman Ventures and Tiger Global. Set that against the comparables the same report cites — Nvidia's $20 billion licensing deal for Groq, Cerebras' $56 billion IPO, Etched's $700 million round at a $21 billion valuation, Marvell buying Celestial AI for up to $5.5 billion in February against no revenue, and Analog Devices buying Empower Semiconductor for $1.5 billion in July — and the pattern is unmistakable. The scarce asset is no longer the accelerator; it is everything that keeps the accelerator from idling, and the incumbent is buying optionality on all of it. This is the compute-financialization wave doing what it does: Nvidia's equity book and its product roadmap are now the same instrument.

Roadmap implication: Two roadmap consequences. First, if you are planning around a credible second source for inference silicon, price in that your alternative may end up inside Nvidia's ecosystem rather than opposite it — write exit and portability terms into any multi-year commit rather than relying on competitive pressure that keeps getting capitalised away. Second, the real signal for infrastructure investors is where the money is going: interconnect, optics and power delivery, not more accelerators. If half the value of a GPU hour is lost to data movement, the vendors solving movement are where the next margin sits, and they are being bid for at 3x marks inside a quarter.

Sources: The Information — Nvidia to Invest in Chip Rival d-Matrix as Its Challengers Choose Partnership · Investing.com — Nvidia plans to invest in AI chip rival d-Matrix- The Information

🌊 The Ripples

Immediate and tactical. Each one ends with something to do now.

Anthropic rewrote Claude's usage policy for machines that move, and dropped one election-era prohibition

The updated Usage Policy was published Thursday and takes effect 12 November. The headline change for anyone building physical systems is new: where Claude is connected to hardware that takes autonomous physical actions, a qualified operator must be able to observe and stop the equipment, and the equipment must hold a safe state if Claude is disconnected. The weapons prohibition now explicitly covers software and components that make weapons function, and arming drones or other autonomous vehicles. The surveillance section was rewritten to bar tracking people without consent, in real time or by analysing previously collected data, and to bar using Claude to decide or recommend whom to investigate, arrest or charge — with consented fraud monitoring, content moderation, journalism and legal research expressly permitted. Anthropic says the weapons and surveillance rewrites describe how it was already enforcing. Two changes are genuinely new in substance: a consolidated section against deceptive campaigns and artificial activity covering fake accounts, fabricated news sites and influence operations, political and commercial alike; and the removal of the blanket ban on personalised vote and campaign targeting, narrowing the elections section to voter deception and election disruption so that legitimate civic work is allowed. Deceptive or personal-data-misusing targeting stays barred elsewhere. There is also a new prohibition on sustained, needless cruelty toward Claude, limited to extreme cases with no discernible purpose.

Do this now: If you ship anything with an actuator — a robot, a drone, a vehicle, a piece of plant equipment — put the 12 November date in your release calendar today and verify two things in code: that a named operator can halt it, and that it fails to a safe state when the model connection drops. Both are good engineering regardless of whose policy requires them. And if you do political or civic work, re-read the elections section before 12 November: personalised targeting moved from prohibited to conditional, and the condition is your data provenance.

Sources: Anthropic — 2026 Usage Policy update

Manus closed more than $500 million, six months after Beijing unwound its sale to Meta

Butterfly Effect, the parent of the agent startup Manus, said on WeChat that it had raised more than $500 million, led by Boyu Capital and IDG Capital with existing shareholders Tencent, HSG and ZhenFund following. It is the first round since the company's separation from Meta, which announced a $2 billion acquisition in December 2025 before Chinese authorities ordered the deal unwound in April 2026 over concerns about losing AI talent westward; Manus resumed independent operations in August 2026 and said it had to delete some user data as part of the split. The company did not disclose a valuation, and did not answer questions about it; TechCrunch reported the previous month that it had been in talks at $4 billion. No revenue figure has been disclosed since the annual recurring revenue of more than $100 million reported at the time of the Meta deal. A Hong Kong listing is reportedly under consideration.

Do this now: The practical read for anyone evaluating agent vendors: an acquisition can be reversed by a government, and the counterparty that survives it is the one whose product is not dependent on the acquirer's infrastructure. Ask your shortlisted vendors where their weights, data and inference actually sit, and what happens to your tenancy if their cap table changes hands — Manus had to delete user data to get its independence back.

Sources: TechCrunch — China's Manus raises over $500M in first funding round since split with Meta · TechNode — Manus parent Butterfly Effect completes more than $500 million funding round

Isomorphic Labs is in early talks at a valuation of at least $40 billion, five months after a $2.1 billion round

Bloomberg reported that Isomorphic Labs, the drug-discovery company spun out of Alphabet's Google DeepMind, is in early talks to raise at a valuation of at least $40 billion, with some coverage putting the range as high as $50 billion. Terms could change; the amount being raised was not stated. The company closed a $2.1 billion Series B about five months ago with substantial Alphabet backing, and counts Eli Lilly and Novartis as partners. Demis Hassabis' stated target for the company is compressing drug development from years into months. Read it next to the same day's federal science package and the picture is coherent rather than coincidental: AI-for-science is being capitalised on both the public and private side in the same week, at marks that assume the compression works.

Do this now: If you are building in AI-for-science, the financing window is open and the comparable is now a mark of at least $40 billion five months after a $2.1 billion round, on a company with no approved drug. Use it: raise against the capability milestone you can actually demonstrate rather than waiting for a clinical readout. If you are buying, the inverse holds — ask which part of the timeline the AI actually compresses, because target identification and trial enrolment are different businesses and only one of them is a modelling problem.

Sources: Bloomberg — Alphabet's Isomorphic Labs in Funding Talks for at Least $40 Billion Value · Yahoo Finance — Alphabet's Isomorphic Labs seeks valuation of at least $40 billion in new funding round

Epoch pointed six frontier models at its own job and found the capability index is not the reliability index

Epoch AI published the companion to the week's self-reported-research question by running GPT-6 Astra, Claude Fable 5.1, Grok 4.6, Gemini 3.8 Flash, Kimi K3 and Qwen 3.8 Max — each on its own agent harness at its highest reasoning setting — across 11 of Epoch's real tasks in five categories: graphic design, data insight generation, data explorer generation, AI data centre research and research design. One run per model per task, graded by a single human against rubrics built with Epoch staff, so the scores are noisy and the authors say so. Fable 5.1 and GPT-6 Astra are broadly tied for the lead and neither can automate the work. The sharpest finding is a measurement one: Kimi K3 scores 158 on the Epoch Capabilities Index, roughly tied with Grok 4.6, yet produced factual errors the closed-weight frontier models did not — filtering all arXiv papers rather than physics papers for a physics chart. Models also converge: three of six picked the same data-insight topic from the polling data. And a methodology footnote worth stealing — in GPT-6 Astra's pilot a 4,096-token output limit truncated 61 of 280 responses before the model named an input, which Astra's own write-up framed as "sensitivity to the acquisition budget" and Epoch calls misleading.

Do this now: Do not let an index score into your vendor decision. Take 10 tasks your team actually did last quarter, run the candidates once each on your own rubric, and have a human grade them — Epoch just showed that a day of that work separates models an index ranks as tied. And when you read any agent write-up, including your own team's, check the output token limit before you believe the failure analysis.

Sources: Epoch AI — Can AI automate Epoch?

The three fired OpenAI safety researchers published an open letter, and OpenAI answered with an internal memo

Jasmine Wang, Tomek Korbak and Mikita Balesni, dismissed by OpenAI the previous week after the company said they violated policy by "accessing and handling sensitive company information," published an open letter Thursday addressed to OpenAI's Safety and Security Committee, Safety Advisory Group and Mission Advisory Council. They deny mishandling information outside established procedures, deny involvement in a leak to The Information about less monitorable architectures in OpenAI's newest models, and deny engaging with outside parties beyond their mandates — writing that "terminations such as ours, executed and communicated so abruptly, are chilling the open culture OpenAI has prized in the past." Wang says OpenAI told her she was fired for accessing an executive's email, which she describes as an inbox she had been given delegated access to for recruiting, opened accidentally after IT failed to remove the access she had asked it to remove, and reported to the executive within minutes. Their three asks are specific: embed third-party safety auditors, preserve monitorability of frontier models, and keep an open dialogue with the safety ecosystem. OpenAI has not formally responded, but shared an internal memo from a research leader praising their contributions, stating "we do not terminate employees for raising concerns," and saying the company agrees with their recommendations.

Do this now: The forward-looking item here is the first ask, not the dispute: embedded third-party auditors with standing access is the arrangement the FTC's reported question set and the embedded-evaluation-market wave are both converging on. If you run an AI programme of any size, decide now whether your evaluator is internal, contracted or named-and-independent, because that is the distinction a regulator will ask about. Operationally, the smaller lesson is free: delegated mailbox access that survives a revocation request is an access-control failure, and it is in your estate too.

Sources: TechCrunch — Fired OpenAI safety researchers dispute misconduct claims, warn of chilling effect

Read this edition and the full archive at excelsiorgroup.ai/insights/signal.

The Signal — The Excelsior Group. Old problems, new physics.