Media reports citing people familiar with the matter indicate that, based on this month's business performance, the latest annualized revenue of Cognition AI, a leading AI coding agent company backed by Silicon Valley venture capital icon Peter Thiel, is expected to reach $1 billion — more than double the annualized revenue run rate recorded just four months ago for this artificial intelligence coding startup.
In early September, the company stated that its annualized revenue run rate had already surpassed $900 million, driven by market demand for its Devin AI coding software powered by AI agent technology. The increasingly popular annualized revenue run rate in Silicon Valley is a projected estimate of a company's full-year sales or annual revenue based on revenue performance over a shorter period. As of May, the company's annualized revenue was previously calculated at $492 million.
Cognition is part of a growing group of AI application companies targeting the lucrative enterprise AI services market, dedicated to simplifying AI agent-driven code writing and debugging processes. Media previously disclosed that earlier this year, after Elon Musk's SpaceX announced a potential $60 billion acquisition of its competitor Cursor, investor interest in Cognition began to heat up. That deal was completed in August. SpaceX had also previously approached Cognition regarding a potential acquisition. Cognition declined to comment on the latest financial figures. The person familiar with the matter requested anonymity because the information has not yet been made public.
Earlier this month, Cognition said it raised $2 billion in a new funding round, with its valuation jumping from approximately $26 billion about three months earlier to $48 billion. The startup's core customers include major global corporations such as Nvidia, Citigroup, and Mercedes-Benz Group. Founded in 2023, Cognition's flagship product is an AI agent called Devin, designed to automate coding processes for engineers, and it has received backing from top venture capital firms including General Catalyst and Peter Thiel's Founders Fund.
Some people familiar with the matter said SpaceX's interest in Cognition stems not only from its technology but also from its business progress. Fully agentic workflows and fully automated coding through AI agents may have become the hottest track in the AI application space, with Anthropic PBC, OpenAI, and SpaceX — the strongest leaders in AI applications — all devoting a significant portion of their businesses to project-level engineering products. Cognition AI itself is one of the purest plays in this "Agentic Software" investment theme.
Its core product, Devin, is positioned as an autonomous AI agent software engineer capable of autonomously planning, writing, testing, and delivering production code within real codebases and development tool environments. It is understood that after completing the acquisition of Windsurf, the company formed a complete AI application software engineering platform consisting of "Windsurf/Devin AI agent collaborative development + Devin handling cloud-based asynchronous autonomous execution + Devin for Terminal undertaking complex engineering tasks." Devin is the "cloud-based autonomous AI engineer," Devin for Terminal is the "local command-line entry point," and Devin Desktop is the desktop IDE/console for managing collaboration between humans and multiple agents; all three are designed for use by customers and engineering teams. For example, Mercedes-Benz has deployed its products across its global R&D and IT systems, and in a four-week pilot, Devin analyzed over 200,000 lines of COBOL code, shortening what was originally estimated to be eight months of modernization work to eight days.
Cognition's official website shows that the company has established partnerships with multiple large enterprises including Mercedes-Benz Group and GE Aerospace. Cognition's commercial partnerships could help SpaceXAI enhance its appeal to potential customers.
Key Signal That AI Applications Are Beginning to Generate Revenue: Cognition Crosses the $1 Billion Annualized Threshold
As Meta Muse and OpenAI Astra drive faster penetration of agent applications, the AI coding application ecosystem is becoming one of the tracks with the most pronounced commercialization progress. Cognition officially announced on September 25 that its annualized revenue run rate had surpassed $1 billion, more than doubling from $492 million in May over approximately four months; the company also completed a new funding round of over $2 billion at a $48 billion valuation in early September.
The AI startup's revenue expansion and fundraising progress together indicate that enterprises are continuing to pay for AI tools capable of participating in real engineering work. The "annualized revenue run rate" here is a metric that converts recent revenue levels into an annualized scale and cannot be equated with confirmed full-year revenue.
A clear product connection has already emerged between Cognition's growth and frontier model upgrades. When OpenAI released Astra, it cited Silas Alberti, Senior Vice President of Research at Cognition, who said the company integrated the model into Devin's agent operating framework on the day of its release, leveraging improvements in computer operation, codebase understanding, and writing capabilities to enhance testing and delivery quality. Meanwhile, Meta Muse brings agents into daily work such as email and travel arrangements through dedicated cloud virtual machines, browser operations, and background task mechanisms. These two paths respectively expand the range of tasks that enterprises and individuals can delegate to AI, making the commercial value of AI applications increasingly dependent on whether work can be completed reliably and the overall cost of completing that work.
From the perspective of enterprise IT budgets and procurement logic, AI coding has a relatively clear value measurement method: whether code passes tests, whether troubleshooting is accelerated, and whether engineer review and rework time is reduced can all be incorporated into project evaluations. Cognition's disclosed Devin product already covers initial incident investigation, vulnerability discovery and classification, and workflows automatically triggered by events in systems such as Slack, GitHub, and Linear; customers include Nvidia, Citigroup, and Mercedes-Benz, among others. It can be inferred that when agents reduce the overall cost of software development and maintenance, enterprise projects for upgrades, automation, and new features that were previously delayed due to insufficient engineering resources may translate into new paid demand. This provides growth space for the AI application market that goes beyond a simple replacement of existing development tool budgets.
As AI applications sweep across the globe, the stock prices of leading AI application companies in global stock markets have performed strongly, with "AI application bellwether" Palantir's stock price surging as much as 150% since 2025.
AI Coding Commercialization Accelerates! The More Capable Future Agents Become, the Stronger the Computing Power Demand
The way of working represented by Muse, Astra, and Devin is extending a single user instruction into multiple rounds of "reasoning — execution — reading results — verification" computational processes. Taking software engineering tasks as an example, an agent may need to understand the codebase, formulate modification plans, generate code, compile and test, check for errors, and continue to make corrections. Among these, GPUs and other accelerators handle model inference, while CPUs execute tool calls, code execution, database queries, and sandbox tasks. Nvidia's technical documentation points out that CPU processing speed affects wait times between model calls, which in turn affects the throughput of the entire agent system.
Therefore, as parallel tasks increase, server CPU computing performance, memory bandwidth, HBM high-performance memory and NAND warm-tier cold storage, as well as CPU-led agent task scheduling capabilities, have also become important conditions for scaling AI services. The expansion of DRAM/NAND memory chip demand mainly comes from three levels: model computation, execution environment, and task state. Model inference requires high-bandwidth memory to carry weights and related computational data; longer contexts and higher concurrency increase the capacity and access requirements of KV Cache; browsers, code execution environments, and large numbers of parallel sandboxes occupy server DRAM. Codebases, files, historical records, and task artifacts also require persistent storage. Nvidia's Dynamo technical solution has discussed placing context caches in tiers across GPU HBM, CPU DRAM, local NVMe, and remote shared storage.
It can be seen that the storage increment brought by the proliferation of agents covers HBM, server memory, and enterprise-grade SSDs, and further raises the data transmission requirements between computing nodes and storage systems. From an investment logic perspective, Cognition's revenue expansion provides an important observational sample — namely, that AI applications can continuously create customer value, which is the prerequisite for translating model capability improvements into more sustained computing power procurement demand. More powerful and more frontier models may complete the same task with fewer calls, lower costs, and higher efficiency, while better success rates and economics can attract more users and more tasks into AI systems; when the expansion of adoption rates and task scale exceeds the decline in resource consumption per task, overall computing and storage demand is expected to continue growing explosively. This provides demand support from application proliferation for AI infrastructure participants such as AMD, Intel, and Arm-architecture server CPUs, as well as memory suppliers like Micron and SK Hynix, and optical interconnect hardware suppliers.
Comments