The Zhitong Finance App learned that some media quoted information revealed by people familiar with the matter as reporting that according to this month's business performance estimates, the latest annualized revenue data for Cognition AI, the AI programming leader supported by Silicon Valley's venture capital godfather Peter Thiel, is expected to reach 1 billion US dollars, which is more than double the annualized revenue run rate (run rate) statistics of this artificial intelligence programming startup four months ago.
The company previously stated in early September that its annualized revenue operating rate had exceeded 900 million US dollars, driven by market demand for Devin AI programming software driven by its AI smart device technology. According to reports, the increasingly popular annualized revenue run rate (run-rate revenue) in Silicon Valley is an estimation of the company's annual sales revenue or annual revenue based on revenue performance over a relatively short period of time. As of May, the company's annualized revenue was previously estimated at US$492 million.
Cognition is a member of a growing group of AI application-type enterprises. These companies target the lucrative AI enterprise-level service market and are committed to simplifying code writing and debugging processes dominated by AI agents. The media previously revealed that investor interest in Cognition began to heat up earlier this year after Elon Musk's SpaceX announced that it might buy its rival Cursor for $60 billion. The deal was completed in August. SpaceX has also previously approached Cognition about a potential acquisition.
Cognition declined to comment on the latest financial data. Since the relevant information has not been made public, the person familiar with the matter requested anonymity.
Earlier this month, Cognition said the company raised $2 billion in a new round of financing, and the valuation jumped from $26 billion about three months ago to $48 billion.
The startup's core customers include large global companies such as Nvidia, Citigroup, and Mercedes-Benz Group.
Cognition was founded in 2023. Its flagship product is an artificial intelligence agent called Devin, which aims to automate the programming process for engineers. It has received support from top venture capital institutions such as General Catalyst and Founders Fund under Peter Thiel, the godfather of Silicon Valley venture capital. Some people familiar with the matter said SpaceX's interest in Cognition stemmed not only from its technology, but also from the progress of its business.
Using AI agents for full-process proxy workflows and fully automated programming has probably become the most popular racetrack in the field of artificial intelligence applications. The strongest leaders in AI applications such as Anthropic PBC, OpenAI, and SpaceX have invested a large part of their business into their own project-level engineering products.
Cognition AI itself is one of the purest targets in this “Agentic Software (AI agent-led software)” investment line. Its core product, Devin, is positioned as an autonomous AI agent software engineer who can independently plan, write, test, and deliver production code in a real code base and development tool environment.
According to information, after completing the acquisition of Windsurf, the company also formed a complete AI application software engineering platform “Windsurf/Devin AI agent collaborative development+Devin responsible for asynchronous autonomous execution in the cloud+Devin for Terminal undertaking complex engineering tasks”; Devin is a “cloud autonomous AI engineer”, Devin for Terminal is a “local command line entry”, and Devin Desktop is a desktop IDE where managers collaborate with multiple agents/ Console; the three are used jointly by customers and engineering teams. For example, Mercedes-Benz (Mercedes-Benz) has deployed its products to global R&D and IT systems. In a four-week pilot, Devin analyzed more than 200,000 lines of COBOL code, shortening the modernization work originally anticipated in eight months to eight days.
According to Cognition's official website, the company has established partnerships with a number of large companies, including the Mercedes-Benz Group and General Electric Aerospace. Cognition's commercial partnership may help SpaceXAI increase its appeal to potential customers.
An important sign that AI apps are starting to redeem revenue: Cognition crosses the $1 billion annualized mark
At a time when Meta Muse and OpenAI Astra are accelerating the penetration of smart applications, AI programming application systems are becoming one of the most prominent commercialization tracks. Cognition officially announced on September 25 that the company's annualized revenue operating rate exceeded 1 billion US dollars, more than double in about four months from 492 million US dollars in May; the company also completed a new round of financing of more than 2 billion US dollars at a valuation of 48 billion US dollars in early September.
The AI startup's revenue expansion and financing progress together indicate that the company continues to pay for AI tools that can participate in real engineering work. The “annualized revenue operating rate” here is an indicator that converts recent revenue levels into full-year scale, and cannot be equivalent to confirmed annual revenue.
There is already a clear product connection between Cognition's growth and cutting-edge model upgrades. When Astra was released, OpenAI quoted Silas Alberti, senior vice president of Cognition Research, as saying that the company connected the model to Devin's intelligent operating framework on the day it was released, using improvements in computer operation, code base understanding, and writing capabilities to improve the quality of testing and delivery. At the same time, Meta Muse brings intelligence into everyday tasks such as email and travel arrangements through dedicated cloud virtual machines, browser operations, and back-office task mechanisms. These two paths each expand the scope of tasks that enterprises and individuals can entrust to AI, making the commercial value of AI applications increasingly dependent on whether the work can be completed reliably and the comprehensive cost of completing the work.
Judging from the enterprise IT budget and procurement logic, AI programming has a clear way to measure value: whether the code can pass testing, whether troubleshooting is accelerated, and whether engineer review and rework time is reduced can all be included in the project evaluation. The Devin products disclosed by Cognition have covered initial incident investigation, vulnerability discovery and classification, and workflows automatically triggered by events in Slack, GitHub, Linear, etc.; customers include companies such as Nvidia, Citi, and Mercedes-Benz.
What can be deduced from this is that when intelligent devices reduce the comprehensive cost of software development and maintenance, upgrades, automation, and new function projects that were previously postponed by enterprises due to insufficient engineering resources may be converted into additional payment requirements. This provides room for growth in the AI application market, beyond the simple replacement of existing development tool budgets. As AI applications take the world by storm, the stock price performance of popular AI application leaders in the global stock market has been strong, and the “AI application weather vane” Palantir's stock price has risen as high as 150% since 2025.
The commercialization of AI programming is accelerating! The more capable intelligent the future, the stronger the demand for computing power
The working method represented by Muse, Astra, and Devin is extending one user instruction to multiple rounds of “reasoning - execution - reading results - verification” calculation process.
In software engineering tasks, for example, agents may need to understand the code base, develop modification plans, generate code, compile tests, check for errors, and continue to fix them. Among them, accelerators such as GPUs undertake model inference, while CPUs perform tool calls, code execution, database queries, and sandbox tasks. According to Nvidia's technical data, CPU processing speed affects the waiting time between model calls, which in turn affects the throughput of the entire intelligent system. Therefore, with the increase in parallel tasks, the computing performance and memory bandwidth of server CPUs, HBM high-performance memory and NAND warm layer cold storage, and CPU-led intelligent task scheduling capabilities have also become important conditions for expanding the scale of AI services.
The expansion in demand for DRAM/NAND memory chips mainly comes from the three levels of model calculation, execution environment, and task status. Model inference requires high bandwidth memory to carry weight and related computational data; longer contexts and higher concurrency increase the capacity and access requirements of the key value cache (KV Cache); the browser, code execution environment, and a large number of parallel sandboxes consume server DRAM. The codebase, files, history, and task products also require durable storage. Nvidia Dynamo's technical solution has been discussed to layer the context cache in GPU HBM, CPU DRAM, local NVMe, and remote shared storage. It can be seen from this that the increase in storage brought about by the popularity of smart devices covers HBM, server memory, and enterprise-grade SSDs, and further increases data transmission requirements between computing nodes and storage systems.
Judging from the investment logic, Cognition's revenue expansion provides an important observation sample — that is, AI applications can continuously create customer value, and only then can the improved model capabilities be transformed into more durable computing power procurement requirements.
More powerful and cutting-edge models may complete the same task with fewer calls, lower costs, and higher efficiency, while better success rates and economics can also attract more users and more tasks to the AI system; as the expansion of adoption rates and task scale exceeds the decline in resource consumption per task, overall computing and storage requirements are expected to continue to blowout. This provides support for demand from AMD, Intel, and ARM architecture server CPUs, as well as AI infrastructure participants such as Micron and SK Hynix, and optical interconnect hardware vendors.