Accel is in talks to lead a $1B round for Thinking Machines at $40B
The AGI-era headlines have quieted down, so the money is doing the talking instead. Tonight: Mira Murati's lab tests whether pedigree still clears a $40 billion price, Microsoft undercuts the transcription market, and Zed turns a sixty-year-old hypertext vision into an agent feature.
Accel is reportedly in talks to lead a $1 billion round valuing Mira Murati's Thinking Machines Lab at $40 billion — a step down from the $50 billion the company was said to be seeking late last year. TechCrunch, citing a report from The Information, says the startup's annual revenue run rate currently stands at over $100 million, according to a source with knowledge of its financials. Neither Accel nor Thinking Machines responded to requests for comment, and the deal has not closed.
The gap between those two numbers is the whole story. A $40 billion price on roughly $100 million of recurring revenue is an extreme multiple by any standard, and it lands after the lab lost co-founders to OpenAI — one of whom has since moved on to Google. The bull case is real but narrow: in July, Thinking Machines shipped Inkling, an open-weight model that earns through usage-based compute fees on its Tinker platform, where customers adapt models to their own proprietary data. That is a genuine product with a genuine billing model, not a research preview — but it is also a customization business competing with every lab that now bundles the same capability for free.
Our read: the down-round framing misses what actually changed. Investors aren't paying less for Murati's team, they're paying for revenue that now exists, at a multiple that assumes Tinker compounds. The same $40 billion number attached itself to Devin maker Cognition eyes $40B valuation in new round this month — when $40 billion becomes the frontier-lab default, that's a signal about the market's appetite, not about any single company.
Microsoft says its MAI-Transcribe-2 speech model now leads the FLEURS benchmark across 60 languages, with an average word error rate of 5.2%. The company claims it beats Gemini 3.5 Transcribe, GPT-Transcribe, Whisper V3-Large, and ScribeV2 while running faster, and it is pricing the model at 10 cents per audio hour as a promotional rate through the end of the year. New features include diarization, configurable transcription styles, and word-level timestamps.
The 5.2% figure matters less than the shape of the claim: Microsoft is arguing that one model holds a consistent accuracy bar across all 60 languages, which kills the per-language model roulette that anyone transcribing multilingual audio has lived with for years. Take the pricing seriously as a competitive weapon rather than a launch discount — at ten cents an hour, transcription stops being a line item worth architecting around.
Zed thinks Ted Nelson's Xanadu finally found its user, and it's an AI agent. In a post this week, Zed's Nathan Sobo argues that Nelson's 1960s hypertext vision didn't die of mismanagement — it died of a missing dependency tree. DeltaDB, the storage layer under Zed's Delta workspace, represents a file as fragments with stable identities rather than a string of characters, so references stay attached to the same span even as the code around them moves.
The payoff is provenance: the files and symbols earlier agents read, edited, or cited become landmarks for the next agent, resolved against the code as it exists now and linked back to the conversations where that understanding was formed. A line number tells you where text sat in one snapshot; an anchor tells you which span you meant across all of them. It's the same pressure Harness rebuilds source control for the agent-written code flood is responding to from the other end — agents generate far more context than git was built to hold, and the tooling is finally catching up.
What to watch: whether the Thinking Machines round actually closes at $40 billion, or gets repriced before the ink dries.
Is a $40 billion valuation on $100 million of revenue a bet on the team or a bet on the market? Tell us in the comments.
Sources: TechCrunch · The Information · Investing.com · Microsoft AI · Unite.AI · VentureBeat · Zed · Zed — Introducing Delta
