Cognition's $47 Billion Valuation Rests on a Margin It Doesn't Yet Have

Cognition AI is set to raise about $1 billion at a $47 billion valuation. That is roughly 52 times its annualized revenue of $900 million. The last time this happened, in May, the valuation was $26 billion. Three months is a long time for any company to more than double its price tag.

The numbers are dramatic enough that you might stop here and move on. But there is a more interesting number buried inside the headline, and it has nothing to do with valuation. It has to do with the fact that Devin does not just sell software. It does work. And work costs money to run.

Here is what Devin does. A developer hands it a task — fix a bug, migrate a legacy codebase, generate tests, write documentation. Devin takes the task. It opens a virtual machine, plans the work, writes code, runs tests, iterates on failures, and opens a pull request. The developer reviews and merges. End to end. That is the product. Not a suggestion in your editor. Not autocomplete. An agent that sits down and does the job.

Cognition charges for this in something called Agentic Computing Units. One ACU is roughly 15 minutes of autonomous work. On the Team plan, 250 ACUs come with a $500 monthly fee. On the entry plan, they cost $2.25 each. There are no seats. Unlimited engineers can use it. What you pay for is the time the machine spends thinking and acting.

This is where the economics diverge from every other software company you know.

Traditional SaaS — the kind that sells at 30x or 50x revenue — typically carries gross margins of 75 to 90 percent. You build the product once. Every additional customer costs almost nothing to serve. Devin does not work that way. Every task it completes burns inference tokens, virtual machine hours, and bandwidth. Those costs sit between the price charged and the revenue recognized. Industry estimates put Devin's gross margins somewhere between 30 and 60 percent now, with a potential path to 60 to 75 percent as inference costs fall over time.

A 52x revenue multiple on 40 percent gross margins is a very different claim than a 52x multiple on 85 percent margins. One is betting on a software company. The other is betting that a compute-heavy service will eventually become a software company.

This is not unique to Cognition. Nearly every company selling AI-powered tools today carries inference costs that traditional SaaS companies never had to think about. But Devin is an extreme case. It is not a feature inside a product. The product is the compute.

Some founders argue that inference costs are not a gross margin problem at all — they are just customer acquisition in disguise. If your product is good enough that it sells itself, the inference spend is doing the work that a sales team would otherwise do. That is plausible for products where the inference is a demo or a trial. Devin is not a trial. The inference is the product. The cost does not stop after the customer signs on. It continues for every unit of work the customer buys.

So what does a $47 billion valuation actually require?

At $900 million in annualized revenue, Cognition trades at roughly 52x. For that multiple to make sense even under generous assumptions, the company needs to reach something like $2 billion or more in revenue within the next year or two, and those margins need to improve toward the high end of the 60-to-75 percent range. Otherwise, you are not buying software. You are buying a service business with extraordinary growth and asking it to eventually become software.

There is another complication. The $900 million figure is annualized from recent monthly usage. It is not contracted annual recurring revenue locked into multi-year agreements. Usage-based revenue can accelerate quickly, as Cognition's has — from $1 million ARR in September 2024 to $492 million annualized by May 2026 to roughly $900 million now. But it can also decelerate if usage drops, if customers switch tools, or if the economics of running Devin make it less attractive than alternatives.

The competitive landscape has changed in a way that matters here. SpaceX completed a $60 billion acquisition of Cursor in August. That is a different kind of company — Cursor sits inside your editor and helps you code, rather than taking full tasks end-to-end. But the acquisition sends a signal about what capital believes autonomous coding is worth, and about who gets to play. Microsoft is shipping GitHub Copilot's coding agent. Anthropic's Claude Code has reportedly crossed $1 billion in annual revenue. OpenAI is pushing Codex. Every one of these companies has deeper pockets than Cognition for the inference arms race.

Cognition's response to this threat has been to build its own model stack rather than rely purely on third-party APIs. It acquired Windsurf's assets, including a model called SWE-1.6 that runs at up to 950 tokens per second. The idea is clear: if you own the models that do the work, you control your own margins. It's a good instinct. But building models that outperform the big labs on software engineering tasks is a different kind of company than selling a coding agent, and it requires sustained investment with no guarantee of payoff.

Let me put the pieces together.

Cognition was founded in November 2023 by three competitive programming prodigies. Scott Wu, the CEO, won gold medals at the International Olympiad in Informatics three times as a teenager. He walked away from Harvard to build a machine that writes code. The founders understand programming at the deepest possible level — because they are among the best programmers their generation has produced. That matters. The people building the product think like the people who will use it.

They shipped Devin in March 2024. It resolved 13.9 percent of tasks end-to-end on the SWE-bench benchmark at launch. That sounded small, but it was far ahead of GPT-4 and Claude 2 at the time. More importantly, they figured out which tasks to sell first: long-tail maintenance, bug fixes, legacy migrations, the kind of work every engineering team has plenty of and nobody wants to do. Nubank used Devin to migrate 6 million lines of code in weeks instead of the 18 months it would have taken human engineers. Eight to twelve times faster. Twenty times cheaper.

The company is dogfooding aggressively. By 2026, Devin was writing 89 percent of the code committed at Cognition itself. If you are selling an AI software engineer, and your own engineers aren't using it, there is no point talking to customers.

So the product works. The revenue is growing at a pace that is almost hard to take seriously. The question is not whether Cognition is doing something valuable. The question is whether the economics of what it's doing can sustain the valuation that capital is assigning to it.

A $47 billion company needs to produce $47 billion worth of value over its lifetime. At current revenue and margin assumptions, Cognition would need several years of sustained growth, followed by margin expansion, followed by scale that justifies the multiple. Any one of those steps could fail. Inference costs could plateau instead of declining. Enterprise customers could prefer tools bundled into their existing Microsoft or Google stacks. A major security failure — autonomous agents executing shell commands in production environments at Goldman Sachs or NASA — could freeze procurement across the enterprise segment.

None of this is to say the valuation is wrong. Private market valuations during manias are rarely right. They are prices, not forecasts. The question for anyone watching Cognition is what you believe about the trajectory of AI inference costs and whether Devin can become a marginally cheap form of engineering capacity rather than an expensive one.

I suspect the real test will not be in revenue or valuation. It will be in the gross margin line. Watch it. If Devin's margins move toward 60 to 75 percent over the next two years, the business model is working and the compute-heavy phase was a temporary cost of building demand. If they stay below 40 percent, no matter how fast revenue grows, Cognition is a service business with a beautiful growth curve, and service businesses do not trade at 52x revenue for long.