Cognition closed a $2 billion Series E at a $48 billion post-money valuation on 8 September — up from $26 billion in May 2026, an 85 per cent increase in four months. Andreessen Horowitz, Accel, Founders Fund, General Catalyst, and Avenir participated. Founded in 2024 by Scott Wu, the company makes Devin — an autonomous AI software engineer deployed in enterprise workflows. Current customers include Mercedes-Benz, NASA, Goldman Sachs, and Citi. Annualised run-rate revenue grew from $492 million in May to $900 million in September; management is projecting $4–5 billion by year-end. Annual compute costs run in the hundreds of millions; projected cash burn for 2026 is approximately $800 million. Cognition has begun training a proprietary model on open-source alternatives to reduce its dependency on OpenAI and Anthropic APIs.

TechCrunch's framing of the round — investors believe AI coding is "far from a winner-take-all market" — captures the explicit counter-thesis the investment represents. The April 2026 Cursor acquisition by SpaceX, partly motivated by the view that compute constraints would force consolidation in AI developer tooling, set one set of expectations. Cognition's $2 billion raise at $48 billion says that the addressable enterprise budget for AI coding is large enough to support multiple dominant platforms simultaneously, without the network effects or switching costs that produce winner-take-all outcomes in consumer software.

The decision to train a proprietary model deserves more attention than the valuation number. At $900 million ARR, Cognition's product is a workflow agent built on inference from third-party models. Every change in API pricing, capability, availability, or terms from OpenAI or Anthropic propagates directly into Cognition's cost structure, product performance, and customer commitments. The company has no contractual protection against the model providers it depends on, and those providers are also its potential competitors — OpenAI's Codex and Anthropic's artifact-generation capabilities are adjacent to Devin's core function. Training a proprietary model eliminates that dependency and converts a structural vulnerability into a defensible advantage. It is also the playbook that every successful AI application company of sufficient scale has ultimately followed: Midjourney, Character.ai, and Perplexity have all moved in this direction. The open question for Cognition is whether it can achieve competitive model capability against companies that have been scaling training infrastructure for years — but the strategic logic is straightforward.