Four months can be an eternity in the artificial intelligence market. In May, Cognition raised over one billion dollars at a 26 billion valuation and reported 492 million dollars in annualized revenue. On September 8, the startup behind Devin announced a new round exceeding 2 billion at a 48 billion valuation, while its run-rate revenue reportedly surged to nearly 900 million.
The figures come directly from the company, and annualized revenue does not necessarily equal revenue actually recognized over a full fiscal year. Yet even with that caveat, the pace signals something significant: software agents are becoming a genuine enterprise spending category.
The round was led by Andreessen Horowitz and Accel alongside existing investors such as Founders Fund, General Catalyst, and Avenir. The roster of participants is massive and includes, among others, Benchmark, Bessemer, Kleiner Perkins, Greylock, Lightspeed, T. Rowe Price, and Nvidia. The market signal is unmistakable: investors are betting that software development automation could become one of the largest commercial applications of AI.
The point is not that Devin “writes code”
Assistants that suggest code have been around for years. Cognition's thesis is more ambitious: transforming the engineer from an executor into a coordinator of agents.
Devin is designed to receive an objective, analyze a repository, plan the work, write and test code, fix bugs, and complete tasks with an increasing degree of autonomy. With its latest features, Cognition is pushing delegation even further. Auto-Triage can perform an initial triage on incident analysis; Security Swarm can scan for and classify vulnerabilities; Automations can trigger work based on events in Slack, GitHub, Linear, and other systems.
This is the difference between a copilot and an agent. The copilot waits for the user to sit in front of the editor. The agent can take an assignment and work in the background, perhaps springing into action when a ticket status changes or a new issue arrives.
Growth suggests enterprises are buying engineering time
Cognition highlights clients including Nvidia for chip design, GE Aerospace for aviation, Citi for financial services, Mercedes-Benz for automotive, and Modal for AI infrastructure. These are company-curated examples, but they illustrate the market it is targeting: large organizations where engineering costs are high and software backlogs are massive.
The economic value of an agent does not necessarily hinge on replacing a human. It can also stem from reducing the time required for repetitive tasks: updating dependencies, writing tests, migrating code, investigating bugs, documenting legacy systems, or drafting an initial version of a feature.
This matters because software demand is not constrained by budgets alone. It is bottlenecked by the number of available engineering hours. Many companies have projects that never get built, maintenance that gets postponed, and code that nobody wants to touch because there are more urgent priorities.
If agents can make tackling part of that workload economically viable, they can expand the market rather than merely cannibalizing existing tools.
Nearly 900 million annualized is not yet 900 million in revenue
The metric chosen by Cognition warrants scrutiny. Run-rate revenue typically means taking the current revenue run-rate and extrapolating it over twelve months. It is useful for describing a fast-growing company, but it is not the same as saying Cognition has already booked 900 million dollars in revenue over the past year.
The distinction matters especially when a valuation escalates this rapidly. The run-rate can shift if customers scale back usage, if key contracts fail to renew, or if a portion of demand is tied to temporary experimentation.
That is why the decisive metric over the coming quarters will be the quality of that revenue: renewals, margins, customer concentration, and usage expansion following initial deployments.
Agents carry an invisible cost: compute
Traditional software enjoys extraordinary margins because, once built, it can be replicated at relatively low cost. AI products feature a different cost structure. Every session consumes inference, and an autonomous agent can generate vastly more calls than a chatbot because it plans, tests, verifies, and iterates on tasks.
This means high revenue does not automatically translate into classic SaaS margins. An agent running for hours may be exceptionally useful, but it can also be very expensive to execute. The ability to route different models based on the task at hand therefore becomes part of the competitive advantage.
Cognition emphasizes its ambition to remain an “independent agent lab,” capable of combining proprietary and third-party models rather than locking itself into a single provider. It is a strategic choice: if inference prices fluctuate rapidly, remaining independent allows workloads to be shifted to whichever model offers the best balance of cost and quality.
Competition won't just be about having the best model
OpenAI, Anthropic, Google, Microsoft, and a long list of startups are baking coding agents into their products. In such a crowded market, access to a top-tier model is not enough.
The differentiator may well reside in the operational layer: how deeply the system understands a repository, how reliable it is over the long run, how seamlessly it connects with enterprise tooling, how it manages permissions and secrets, and how transparent it is when it makes a mistake.
Companies don't buy benchmarks. They buy a system that has to work inside real-world development workflows, with thousands of dependencies and security constraints. This is where Cognition is trying to build its moat.
The risk is selling autonomy before it is reliable
The more work is delegated, the higher the cost of error becomes. An incorrect suggestion in an editor is caught by the engineer before it is accepted. An agent modifying dozens of files, opening pull requests, or reacting automatically to an event can cause far broader damage.
That is why the rise of coding agents will make logging, sandboxing, human review, and clear permission boundaries essential. The goal cannot simply be “how much autonomy can we give,” but “what autonomy can we control.”
In sectors such as finance, automotive, and aerospace, cited among Cognition's clients, change traceability is particularly critical. Code is not just an output: it can have regulatory and physical consequences.
The real test will be what happens to the organization
If agents like Devin become truly commonplace, the most interesting transformation could be organizational. A single engineer could manage multiple tasks in parallel. Teams could reduce time spent on maintenance and increase time dedicated to architecture, product, and review.
Yet the way junior engineers are trained could also change. Today, many skills are built precisely through the repetitive tasks that agents excel at automating. If those tasks disappear, companies will have to find new ways to develop engineers.
It is a dilemma reminiscent of other professional automations: eliminating low-value work increases productivity, but it can also erase the pathway through which high-value work is learned.
A valuation betting on a new unit of work
The 48 billion attributed to Cognition is not proof that Devin will win the market. It is a bet that the new unit of work in software will no longer merely be an engineer's billable hour or a tool license, but the compute allocated to an agent.
If this concept takes hold, the developer tools market could expand massively: every company could manage fleets of agents continuously maintaining, testing, migrating, and building software.
If, on the other hand, autonomy proves fragile, expensive, or difficult to govern, current valuations will look far more aggressive.
The metric to watch, then, will not be the next funding round. It will be how much real work is entrusted to Devin, how often it is accepted without major corrections, and whether customers expand their usage after moving past the pilot phase.
The race for coding agents has moved from demos to balance sheets. Now it must prove it can also move from balance sheets to productivity.
