Inside Cognition’s Booming Growth, High Cash Burn

Revenue is surging at Cognition and several other AI application providers, previously unreported data show, offering some reassurance to investors and hardware suppliers that startups can thrive despite competition from Anthropic and OpenAI.

Cognition, which makes the Devin coding assistant, is generating around $900 million in annualized revenue, or $75 million per month, according to a person with knowledge of the figure. That’s up more than three times since the start of the year. The company, which boasts big customers including BNY, Santander and Mercedes, is currently raising funds in a round that could value it at around $45 billion.

Executives have projected Cognition would end the year with more than $1.5 billion in annualized revenue, multiple people with knowledge of the company said, and reach between $4 billion and $5 billion in annualized revenue next year.

Other startups targeting different customers are posting strong growth too. Higgsfield, whose platform generates AI videos and images, said this month its annualized revenue surpassed $700 million, more than tripling since the start of the year. Perplexity, an early darling of the AI boom, has more than tripled annualized revenue so far this year to at least $750 million, thanks in part to smaller businesses and entrepreneurs that use its AI agent for automating white-collar tasks. Manus, which sells a similar product, has more than quadrupled its annualized revenue to more than $400 million since the end of last year.

And revenue at OpenEvidence, which is like a ChatGPT for doctors, has more than doubled since the start of the year to more than $300 million in annualized revenue from ads that appear in its search results.

For many of these firms, such growth is coming at a high cost due to the specialized servers they need for developing and running AI. Cognition’s cash burn could reach $800 million this year, for instance.

In addition, revenue at Anthropic and OpenAI dwarfs that of the next 33 biggest AI-native startups combined—including Cognition, Higgsfield and the others. And Anthropic and OpenAI’s revenues from app and model sales have actually grown faster this year, meaning they have captured 89% of that combined market, up 4.5 percentage points from nine months ago, according to The Information calculations. (Our analysis includes Cognition rival Cursor, which was just acquired by SpaceX.)

That trend could strengthen a private argument some investors at firms like Sequoia Capital have made that the vast majority of software value in the current AI era will be generated by the top developers of advanced AI models rather than by developers of pure AI apps. Almost all the AI app companies in our analysis depend substantially on models from Anthropic and OpenAI, which increasingly are competing with their startup customers by developing products targeting specific industries or professional roles. That could eventually make it harder for those startups to sustain their growth.

Beyond Coding

Many investors are nonetheless betting there will be room for plenty of winners. They see AI businesses capturing a portion of the global white-collar labor market, valued at tens of trillions of dollars a year, as well as creating new demand that enlarges the market itself. Startups like Cognition, similar to Anthropic and OpenAI, view coding as a natural foundation for automating all digital tasks for workers. Anthropic said its AI coding product effectively built the company’s Cowork product that nontechnical workers use to automate data analysis and other tasks.

Cognition, for its part, has told prospective investors it would likely release products to automate tasks related to data science, security, product development and design, and finance.

Fast-growing AI startups face challenges beyond competing against Anthropic and OpenAI. They don’t generate profits and their financials typically reflect the rising cost of obtaining specialized Nvidia servers to power AI services. Demand for such servers has exceeded supply, forcing startups including Cognition to sign long term contracts with cloud providers to gain access to them.

Cognition pays some of these cloud costs up front, gaining capacity that it may not need right away and which is amortized over time, making the costs similar to a capital expenditure like building a data center, it has told some prospective investors.

Cognition uses its server cluster to train new models for its AI coding agent and to power it for customers, according to a person with knowledge of it. Cognition trains new models by tweaking existing open-source models, a process known as post-training. The Nvidia-server cluster Cognition leased costs hundreds of millions of dollars a year, this person said, and is a key reason the company could burn $800 million in cash this year. (It burned around $200 million in the second quarter, said another person with knowledge of the company’s performance.)

Excluding the costs of Cognition’s development of its own coding models it would be close to breaking even, in terms of free cash flow, one of the people said.

Wu, a former math prodigy, has said he is so confident in the product that Cognition is promising enterprise customers up to $10 million in credits if Devin fails to deliver engineering results worth at least what customers pay for it.

Margin Squeeze

Cognition recently was generating a gross profit margin from its enterprise business of nearly 50%, this person said, because model-training costs aren’t included in its cost of goods sold. Anthropic and OpenAI also exclude their hefty training costs from their gross margin calculations.

Software businesses typically aim for gross margins above 70% but they don’t have the kind of computing expenses AI startups do.

Anthropic and OpenAI also have struggled with gross margins due to unexpected server costs. OpenAI’s margin actually went backwards last year, falling to 33%—well short of the 46% it had projected—compared to 40% in 2024. Anthropic, meanwhile, landed at 40% gross margins in 2025, 10 points below its optimistic target, because inference costs on Google and Amazon servers came in 23% above plan. Both Anthropic and OpenAI have projected such margins would steadily rise in the coming years.

On the bright side, Cognition has been able to run models more efficiently and squeeze more revenue from the same cluster of servers, a trend its leaders believe will continue, this person said.

Cognition also spends more money for the cluster so it can host some of the AI models that power its service, rather than outsourcing those computations, known as inference, to specialized cloud providers the way other startups do. Doing so requires more upfront costs but could save Cognition money over time because it isn’t paying intermediaries that are trying to generate profits from powering inference workloads.

In terms of revenue, Cognition lags rival Cursor, which SpaceX bought for $60 billion. Cursor was recently generating $4 billion in annualized revenue, up from about $1 billion at the end of last year, though Anthropic and OpenAI generate substantially more coding-related revenue.

While both companies rely heavily on Anthropic and OpenAI models to power their coding services, they’re developing their own bespoke AI models and trying to move their customers’ workloads to those models, which cost less to run.

‘Actually Good Now’

Cognition focuses on serving large enterprises, and it told potential investors that around 50 businesses are on pace to pay the startup more than $1 million a year. One unnamed customer makes up around 10% of Cognition’s business.

CEO Scott Wu founded Cognition three years ago, raised money from investors including Founders Fund, Lux Capital, and 8VC and by now has raised more than $2 billion.

Cognition’s product, Devin, was ahead of its time in trying to automate complex coding projects. The first versions of the product didn’t always live up to expectations, which explains why the company’s recent advertisements in San Francisco said that Devin was “actually good now.”