Cognition's Devin Triples to $900M ARR — and Burns $800M a Year Doing It

Cognition, the startup behind the Devin autonomous coding agent, is generating roughly $900 million in annualized revenue — more than triple where it stood at the start of 2026 — according to sources cited by The Information. Executives reportedly project the company will exit the year with more than $1.5 billion in annualized revenue. But the same report carries a stark counterweight: Cognition could burn through as much as $800 million in cash this year as it spends aggressively on Nvidia servers to power its agents.

The numbers, in short: revenue tripling, cash nearly all going back out the door. It is the cleanest illustration yet of the economics of the agentic AI era — where growth is spectacular, and so is the compute bill.

From experiment to $900M machine

Devin, launched in early 2024 as a fully autonomous “AI software engineer,” was initially met with skepticism. Critics called the demos overproduced and questioned whether enterprises would trust an agent to write, test, and ship code end-to-end. Two and a half years later, the market has answered. Third-party estimates tracked the company’s annualized revenue at around $492 million by May 2026 — up from an estimated $37 million a year earlier — before the latest jump to roughly $900 million, a figure that works out to about $75 million per month.

Cognition’s portfolio now spans the Devin agent platform, the Windsurf IDE it acquired in mid-2025, and enterprise deployments that treat agents as members of engineering teams. In May 2026, the company raised more than $1 billion in a Series D at a $26 billion valuation, led by Lux Capital, General Catalyst, and 8VC — making it one of the most valuable private AI application companies in the world.

At $900 million in annualized revenue, that $26 billion valuation implies roughly 29x current ARR — rich by traditional SaaS standards, but defensible if the company really does close the year above $1.5 billion. Forward multiples compress quickly when revenue triples every few quarters.

Why the burn is so large

The $800 million cash-burn figure deserves as much attention as the revenue. Unlike classic software companies, whose marginal cost per customer approaches zero, agent companies pay for every unit of work their products perform. Devin-style agents run long, multi-step sessions — reading codebases, running tests, debugging autonomously — and each session consumes GPU time that Cognition, which does not own hyperscale infrastructure, buys at market rates.

That cost structure turns the traditional software P&L inside out. Gross margins on agentic workloads are materially lower than the 80%+ margins investors expect from SaaS, and they only improve if model inference costs keep falling or if agents get more efficient per task. Cognition is effectively betting that compute prices decline faster than pricing pressure erodes what enterprises will pay per unit of completed engineering work.

It is also a bet on scale itself: the more Devin works, the more data Cognition collects about how agents succeed and fail, and the better its orchestration and reliability become — advantages that are hard for rivals to replicate without similar volume.

The coding-agent land grab

Cognition’s trajectory is unfolding inside the fiercest competitive race in applied AI. Anthropic’s Claude Code has become the default agent for many professional developers — the company just announced it will permanently raise weekly Claude Code usage limits by 25% on September 14. OpenAI’s Codex added persistent mode, letting agents work continuously in the background. Cursor, the SpaceX-owned IDE, is locked in a public feud with OpenAI, which says it will cut off Cursor’s model access from November 12.

Against that backdrop, Cognition’s $900M ARR is significant precisely because it is not a foundation-model company. It is an application-layer business proving that autonomous agents — not chatbots — are where enterprises are actually spending. Coding was always predicted to be AI’s first billion-dollar vertical; Cognition’s numbers are the hard evidence.

The comparison to infrastructure players is instructive, too. The same week the Cognition figures emerged, AI cloud provider Lambda raised about $1 billion in private short-dated debt to buy Nvidia GPUs that will be leased by Microsoft. Neoclouds are financing chips with debt against contracted revenue; application companies like Cognition are burning equity cash on the same chips. Both models hinge on the same assumption — that demand for agentic compute keeps outrunning supply.

What to watch

Three questions will determine whether Cognition’s math works out. First, can net revenue retention hold as enterprises get smarter about managing agent usage and costs? Second, do inference costs fall fast enough to expand margins before competition compresses prices — particularly as model providers themselves ship competing agent products? Third, does the $1.5 billion year-end projection materialize, validating the growth narrative that underwrites the $26 billion valuation?

What is no longer in question is demand. Enterprises are not experimenting with autonomous coding agents anymore; they are budgeting for them. Cognition’s tripled revenue and eye-watering burn are two sides of the same coin — a company racing to capture a market it helped create, paying the compute toll in real time, and betting that being the biggest spender on agents today makes it the most profitable agent company tomorrow.