Bloomberg reported on September 2 that Cognition — the company behind the Devin coding agent and, since June, the Windsurf editor rebranded as Devin Desktop — is in talks to raise roughly $1 billion at a $47 billion valuation. Per people familiar with the discussions, investor demand is running close to $10 billion against that $1 billion target, meaning the final raise could land higher than the number currently being reported. None of this is closed. Cognition didn’t comment for Bloomberg’s story, and “talks are ongoing” is doing real work in that sentence — terms, size, and even the headline valuation could shift before anything is signed. Corroborated independently via Yahoo Finance’s mirror of the same reporting and Investing.com, both citing the same Bloomberg sourcing rather than independent confirmation.
Treat the specific number as provisional, then. The trajectory underneath it is the actual story.
The numbers that aren’t in dispute#
Three months earlier — roughly June 2026 — Cognition closed a round at a $26 billion valuation. If the reported $47 billion figure holds, that’s nearly a doubling in a single quarter, in a funding environment where even AI-sector “hot money” rounds don’t usually reprice that fast without a real revenue signal backing them up. That signal appears to be there: Cognition’s annualized revenue run-rate has reportedly grown from around $492 million in late May to more than $900 million by September — also close to doubling in the same window. Revenue growth roughly tracking valuation growth is a healthier pattern than the alternative (valuation running ahead of revenue on vibes alone), even accounting for the fact that all of these figures come from unnamed sources rather than an audited filing.
Bloomberg’s reporting also draws an explicit connective line: investor interest reportedly accelerated after SpaceX closed its $60 billion all-stock acquisition of Cursor’s parent company, Anysphere, on August 14 — a deal this blog covered in detail when OpenAI responded by cutting off Cursor’s model access three weeks later. Put the two deals side by side and the read is straightforward: Cursor just got bought for $60 billion by a company that makes rockets, and the AI-coding-tools category’s other marquee independent player is now being valued at a number that would have been unthinkable for a coding startup eighteen months ago. Money is flowing into this category at a pace that makes the underlying architecture choices worth scrutinizing rather than waving through as “AI hype, film at eleven.”
What $47 billion is actually buying#
Cognition’s product is no longer just “Devin,” singular. Since rebranding Windsurf as Devin Desktop in June, the company has organized its whole lineup around a cloud-first, delegate-and-review workflow: Devin Desktop (the IDE, now built around an Agent Command Center — a Kanban board for tracking every running local and cloud agent), Devin Cloud (autonomous agents running in isolated VMs that open pull requests and run their own QA), Devin CLI, and an open Agent Client Protocol that lets third-party agents — Codex, Claude Agent, OpenCode — run inside the same interface as Cognition’s own model. That last piece is a tacit admission that Cognition doesn’t expect Devin’s own model to always be the best available option, so it built the plumbing to swap in whichever one is.
The in-house model, SWE-1.7, shipped July 8 — RL post-trained on top of Moonshot’s Kimi K2.7 Code inside Devin’s own agent harness, running at roughly 1,000 tokens/second on Cerebras hardware. Cognition’s own reported numbers put it a few points behind Opus 4.8 on FrontierCode 1.1 and Terminal-Bench 2.1, ahead of GPT-5.5 on SWE-bench Multilingual, and explicitly positioned around cost-efficiency rather than chasing the frontier outright. That’s a coherent strategy — not every workload needs the most expensive model available — but it also means the thing investors are pricing at $47 billion is not “the best coding model,” it’s the orchestration layer, the workflow, and the brand sitting on top of a model that’s openly built from someone else’s base weights.
The pattern this blog keeps flagging#
That’s the same structural fact this blog raised about Cursor after the OpenAI cutoff: being a wrapper around other labs’ models is a business built on continued access to those models. Cognition mitigates that risk better than Cursor did — SWE-1.7 gives it a real in-house fallback, and ACP means it isn’t betting the product on any single external model relationship the way Cursor’s OpenAI dependency briefly threatened to. But the core distinction this blog keeps returning to still holds: Devin’s autonomous, delegate-to-the-cloud-and-review-the-PR workflow is philosophically much closer to how Claude Code actually works than Cursor’s IDE-anchored, human-drives-every-keystroke model ever was. Cognition, to its credit, seems to have absorbed that lesson before the market forced it to — the Agent Command Center and Devin Cloud framing are explicitly about directing autonomous work rather than pair-programming with an autocomplete engine.
Where the comparison still favors Anthropic is underneath the workflow layer. Claude Code doesn’t need an open protocol to swap in a better model when its own falls behind, because Anthropic builds both the agent and the frontier model it runs on — there’s no RL-post-trained-on-a-third-party-base-model step in between, and no scenario where a model-access dispute forces an architecture change. Cognition’s $47 billion bet is a bet that orchestration, workflow design, and enterprise distribution are worth pricing independently of who owns the underlying weights. That might be right — plenty of durable software businesses are built on someone else’s infrastructure. But it’s worth naming clearly: the number getting the headlines this week values a very good agent harness, not a new frontier model, and the two are not the same kind of moat.
What to watch#
The round hasn’t closed. If it does at anything close to $47 billion, expect two follow-on threads worth tracking: whether Cognition uses the capital to push SWE-1.7’s next revision closer to genuine frontier parity rather than staying in the “near-frontier at a discount” lane, and whether Kimi K2.7’s owner, Moonshot AI, has anything to say about a $47 billion valuation resting partly on RL post-training over its base model. Neither is confirmed yet. Both are more interesting than the valuation number itself.
Sources: AI Startup Cognition Set to Raise Around $1 Billion at a $47 Billion Value — Bloomberg (primary, via Yahoo Finance mirror), Investing.com syndication of the same report, SWE-1.7: Frontier Intelligence at a Fraction of the Cost — Cognition, Windsurf Is Now Devin Desktop — Devin; this blog’s own prior coverage of the SpaceX-Cursor acquisition and OpenAI model cutoff and Windsurf after the Cognition acquisition.