Cognition snubbed SpaceX, doubled to a $1 billion run rate, and priced itself at 48x

Devin doubled the run rate in four months

Cognition is on track to book $1 billion in annualized revenue this month, roughly double its May pace, Bloomberg reported on 25 September, citing a person familiar with the numbers↗; the company declined to comment ¹². The company itself had said its run rate topped $900 million in September, up from $492 million in May, as Devin, its autonomous coding agent, spread through Nvidia, Citi and Mercedes-Benz ¹⁴.

SpaceX came back five days after buying Cursor

SpaceX closed its $60 billion, all-stock purchase of Cursor's parent Anysphere on 14 August ⁵. Five days later it approached Cognition about a buyout↗, and Cognition did not engage. Chief executive Scott Wu answered in public: "not for sale and we haven't been talking" ³. Within three weeks Cognition had raised more than $2 billion at $48 billion, co-led by Andreessen Horowitz and Accel, up from $26 billion in May ⁴. Musk has told staff he wants AI revenue to out-earn rockets by this month ³.

Index funds paid for the consolidation

The Cursor money traces to a June IPO that raised a record $85 billion↗ ⁶, and passive funds had no choice about holding the acquirer: after Nasdaq rewrote its entry rules on 1 May↗, Nasdaq-100 funds executed about $4.3 billion of mandatory SPCX purchases on 6 July ⁷. Cognition's holders took the other exit, private cash, and were marked from $26 billion to $48 billion in three months ⁴. Cofounder Steven Hao, 30, became the youngest member of the Forbes↗ 400 at $5.8 billion ⁸.

Independence was marked at 48 times revenue

The new round prices Cognition at 48 times its on-track run rate. Cursor fetched about 30 times annualized revenue when it sold, and even that took a compute shortage to force the deal ⁹. Public software carries nothing like it: PTC, Zoom and Akamai trade at 5.8, 5.3 and 3.8 times forward sales↗ ¹⁰. The bulls' answer is growth: reporting points to $4 billion to $5 billion in annualized revenue by year-end, which would compress the mark toward 12 times ⁹. The bill is real either way, an Nvidia cluster costing hundreds of millions a year and possible 2026 cash burn near $800 million ⁹.

The labs underneath cut prices the same month

Days before the run-rate report, OpenAI shipped GPT-6 Sol at $2 per million input tokens↗, half its predecessor, aimed squarely at building features, reviewing code and debugging; Anthropic cut Opus 5.5 by 20 percent ¹¹. Cheaper tokens cut both ways for an agent vendor: they shrink what Devin can charge, and they shrink what Cognition pays to run it, which is why the company is training its own model on open-source bases ⁹¹¹.

The tell is a compute negotiation

The refused buyer and the refusenik are still talking about Cognition renting SpaceX's GPUs ³. Avenir, which co-led the round, sees a path from a $26 billion market to $800 billion ⁴. Until then, the company that would not be bought is negotiating to rent its engines from the company it refused.