The AI labs got faster; the market stopped paying for it

At 12:16 on November 5, prediction markets cut the odds that Anthropic files its IPO from 84 percent to 60 percent, and the talk around the valuation slid toward $1 trillion ^[1]. At 18:02 that afternoon, METR published its measurement of the same company: development speed running at 1.9 times the prior year's pace, attributed to the models improving their own training, with three advanced models trained in six months and withheld from the public ^[2]. The repricing came first, hours before the disclosure. The clock says what it says. That ordering carries the whole week. Two ledgers ran in parallel, and they never once agreed. On the technical side, the record-keeping was remarkable. OpenAI's autonomous vulnerability pipeline turned up 948 working Linux kernel exploits for $1.21 million in API credits — about $1,279 each — with human review reduced to the final stage of a five-stage process ^[3]. The same week the company shipped GPT-6 Cyber, a vulnerability-research model rated Critical, aimed at a $51.3 billion security market ^[4]. And the CEO who built his public identity on deliberate slowness is the one METR clocked at double its own prior pace ^[5][2]. On the money side, over the same window, the verdict ran the other way, in strict order. On November 4 the market handed the tech giants roughly $1 trillion in a single day on hopes that a Democratic Congress would curb AI spending and preserve free cash flow, rewarding Apple for spending the least ^[6]. A day later a $2 trillion rout swept across the AI supply chain, Nvidia alone down more than 9 percent, the selling blamed on regulation fears stacked on a correction that had been running since October ^[7][8]. That evening an anonymous seller put a $500 million data-center loan portfolio to auction, including a slice backing a Stargate site ^[9]. By the next morning three data-center IPOs had been postponed, and Anthropic's revenue growth quarter over quarter had slid from about 145 percent to 43 percent, with roughly 20 percent projected next ^[10]. The two ledgers meet on one arithmetic: the variable the labs optimized for — model speed — has stopped being the variable the market buys. Three strands land on the same number. The acceleration turned private. The three advanced models METR counted were trained and kept from the public, and the benchmark data behind the measurement came redacted — capability the market cannot price because it cannot see it ^[2]. The shipped product still fails. Visa's first study of the matter found AI agents complete only about 57 percent of basic purchases, with Google and Meta's best near 82 percent and OpenAI's agent at 64 ^[11]. Visa's head of growth put the gap plainly.

We’re not there yet from a trust or technology perspective. But we’ll soon get into a world where consumers can trust and delegate. — Rubail Birwadker

OpenAI's new Dots agent, for its part, routes tasks to a smaller, cheaper model to hold down compute costs ^[12]. And the balance sheet has stopped balancing. Annualized hyperscaler investment now approaches $1 trillion against recurring revenue of roughly one-fifth that, with the token price war commoditizing the frontier ^[13].

There are limits to how much investment hyperscalers can tap from capital markets. The trend has to be for deceleration — Citrinitas Capital Management Inc.

Private capital did not flee, to be clear. Goldman Sachs and Apollo arranged a record $40 billion off-balance-sheet GPU sale-leaseback for SpaceX on November 4 — money migrating into structured private credit even as public AI equity fell ^[14]. And exits now route elsewhere: Meta bought the agent startup Conductor, whose cloud product shuts down November 20 with its users' data deleted, the product dying so the team can join a giant ^[15]. November 17 tests both ledgers on the same day. Nvidia reports against roughly $108 billion in guidance ^[7], and Anthropic's advisers meet on whether to file, delay, or downsize ^[1]. The reading above is wrong if Anthropic files near its $1.5-to-$2 trillion target and the offering holds, if fourth-quarter sequential growth re-accelerates, or if a capability announcement moves the labs' valuations again.