Cognition CEO Scott Wu’s previously undisclosed data shows that revenue at Cognition and several other AI application providers is surging rapidly. This provides some confidence to investors and hardware suppliers: even when facing competition from Anthropic and OpenAI, AI startups still have opportunities to thrive. Cognition is the developer of the AI programming assistant Devin. According to people familiar with the matter, the company’s current annualized revenue is approximately $900 million, or $75 million per month, having more than doubled since the beginning of this year. The company counts New York Bank, Santander, and Mercedes-Benz among its major clients and is currently advancing a new funding round with a post-investment valuation expected to reach $45 billion.

Multiple sources familiar with the company’s situation say executives forecast that Cognition’s annualized revenue will exceed $1.5 billion by year-end, with next year’s annualized revenue potentially reaching $4-5 billion. Other AI startups serving different customer segments are also experiencing rapid growth. AI image and video generation platform Higgsfield disclosed this month that its annualized revenue has surpassed $700 million, more than doubling since the start of the year. AI search standout Perplexity, benefiting from numerous small and medium-sized enterprises and independent entrepreneurs using its AI agents for white-collar work, is now generating at least $750 million in annualized revenue this year. Similar product maker Manus has surpassed $400 million in revenue since late last year. OpenEvidence, a ChatGPT-like product for doctors, has broken through $300 million in annualized revenue through search result advertising.

However, for most of these companies, the cost of rapid growth is extremely high: developing and running AI requires vast quantities of specialized servers. In Cognition’s case, cash burn this year could reach as high as $800 million. Furthermore, the revenue of Anthropic and OpenAI far exceeds the combined revenue of Cognition, Higgsfield, and the other 33 leading native AI startups. According to calculations by The Information, the revenue growth rates of Anthropic and OpenAI’s application and model businesses are accelerating further this year, with the two companies combined holding an 89% share of this market, up 4.5 percentage points from nine months ago. (This tally includes Cognition’s competitor Cursor, which was recently acquired by SpaceX.)

Some investors at firms such as Sequoia Capital have held a private viewpoint: in the current AI era, the vast majority of software value will be created by top large model developers rather than startups simply building AI applications. The above market trends further validate this judgment. Nearly all AI application startups covered in this report are highly dependent on Anthropic and OpenAI models. At the same time, the two model giants are also developing products for specific industries and job functions, competing directly with their own startup customers. This could ultimately constrain the continued growth of these startups.

Beyond the coding track, many investors still believe the sector will produce numerous winners. They argue that AI companies can tap into the multi-trillion-dollar global white-collar labor market while creating entirely new demand, further expanding the market pie. Similar to Anthropic and OpenAI’s approach, Cognition views programming as the natural foundation for achieving full digital automation of worker tasks. Anthropic states that its AI programming product is the underlying cornerstone of its Cowork product, which targets non-technical personnel for tasks such as data analysis. Cognition has told potential investors it will likely launch automation products for data science, security, product R&D and design, and finance.

Rapidly growing AI startups, in addition to contending with Anthropic and OpenAI, face other challenges: they have yet to achieve profitability, and their financial statements show continuously rising costs for purchasing NVIDIA specialized servers and supporting AI services. Supply constraints on servers are forcing numerous startups, including Cognition, to sign long-term contracts with cloud vendors to secure computing resources. Cognition has told some potential investors that it pre-pays a portion of cloud service fees to obtain computing capacity it does not fully need at present, with related costs amortized over time—a structure resembling capital expenditures like building proprietary data centers.

According to sources, Cognition maintains its own server clusters, both for training new models of its AI programming agent and for serving customers. The company trains new models by fine-tuning (post-training) existing open-source models. This server cluster leased from NVIDIA costs several hundred million dollars annually, which is the primary reason the company may burn through $800 million in cash this year. Another source noted that the company’s cash burn in the second quarter was approximately $200 million. One person familiar with the matter said that excluding the development costs of its proprietary programming models, Cognition’s free cash flow is already close to breakeven.

Gross margin pressure is evident. Sources reveal that Cognition’s enterprise business currently carries a gross margin of nearly 50%, and this calculation does not include model training costs. Anthropic and OpenAI similarly exclude hefty model training expenses when calculating gross margins. Ordinary software companies typically pursue gross margins above 70%, but AI startups must shoulder substantial computing costs and cannot achieve the profitability levels of traditional software. Cognition ran advertisements in San Francisco last month featuring a poster of CEO Scott Wu. Weighed down by unexpectedly high server costs, Anthropic and OpenAI are also grappling with gross margin issues. OpenAI’s gross margin actually declined last year to 33%, well below the prior expectation of 46%, compared to 40% in 2024. Anthropic’s 2025 gross margin stands at 40%, 10 percentage points below its optimistic target, because server inference costs from Google and Amazon ran 23% higher than planned. Both companies project steady gross margin improvements in the coming years.

On the positive side, sources say Cognition has achieved gains in model operational efficiency, allowing the same server cluster to generate more revenue, and management believes this trend will continue. Cognition has invested in building its own clusters, hosting the AI model inference compute for certain operations locally rather than outsourcing inference to professional cloud providers like other startups. This approach carries higher upfront investment but can save costs over the long term by eliminating the margin paid to intermediaries. In terms of revenue, Cognition trails competitor Cursor, which was acquired by SpaceX for $60 billion. Cursor’s latest annualized revenue stands at $4 billion, up from roughly $1 billion at the end of last year. However, Anthropic and OpenAI’s programming-related revenue still far exceeds both companies. Both coding businesses rely heavily on Anthropic and OpenAI base models, but both are developing proprietary large models in-house, attempting to migrate customer workloads to lower-cost proprietary models. "Now it actually works" is the new tagline.

Cognition primarily targets large enterprise clients. The company has told potential investors that approximately 50 corporate clients pay it over $1 million annually, with one undisclosed client contributing about 10% of Cognition’s business revenue. CEO Scott Wu founded Cognition three years ago, with investors including Founders Fund, Lux Capital, and 8VC, and cumulative funding has exceeded $2 billion. Bloomberg previously reported that Cognition is in talks for a new funding round with a valuation target of $40 billion. Cognition’s product Devin initially sought to automate complex programming projects—a highly ambitious concept—but early versions underperformed expectations. This explains why Cognition recently launched advertising in San Francisco with the slogan: “Now it actually works.” Scott Wu, a mathematical prodigy, is highly confident in the product. Cognition promises enterprise clients: if the engineering output delivered by Devin falls short of the value corresponding to the client’s payment, clients can receive up to $10 million worth of service credits. Wu himself did not comment on this report.