Bloomberg reported that Cognition, the AI coding startup behind Devin, is on track to hit $1 billion in annualized revenue โ and the gap between that headline and what the company actually said is a precise illustration of how run-rate figures travel.
According to Bloomberg’s September 25 report, Cognition โ the AI coding company behind the autonomous software-engineering agent Devin โ is on track to generate annualized revenue of $1 billion. That framing is the body of the story. The headline goes one step further, stating the company “hits $1 billion in annualized revenue.” The company itself, separately, said in September that its run-rate revenue “topped $900 million.” Those are three distinct claims stacked on the same underlying unit, and separating them is worth the effort.
The trajectory is real and significant regardless of which number anchors it. A run rate of $492 million in May, rising above $900 million by Cognition’s own September statement, and tracking toward $1 billion by Bloomberg’s read of this month’s performance, represents roughly a doubling in approximately four months. That is the operative fact. Annualized run rate โ one month multiplied by twelve โ is a standard, widely used software metric, and reporting it is ordinary practice across the industry.
On the valuation side, three numbers have circulated in close succession. The $26 billion figure from May is a completed transaction. The roughly $40 billion figure reported in August by TechCrunch describes talks, not a closed deal. And the September number is also a completed transaction: Cognition’s own newsroom announced a Series E on 8 September, raising over $2 billion at a $48 billion valuation, led by Andreessen Horowitz and Accel. So two of the three are prices somebody actually paid and one is a conversation โ which is exactly why each needs its status attached. Earlier reporting had the September round at around $1 billion at roughly $47 billion; the closed round was larger on both figures, and the pre-close number is the one still circulating.
The key insight: Annualized run rate is one month multiplied by twelve โ not money the company has received. For a business growing slowly, run rate and trailing revenue converge. For one doubling every four months, they diverge sharply, and the gap widens with the growth rate. The more impressive the growth, the less the headline figure resembles cash that has actually arrived. That is arithmetic built into the definition of the unit, not a criticism of anyone who uses it.
Three bars, one unit, and none of them is money that has arrived.
The Structural Read
Run rate’s accuracy problem is asymmetric. For a flat business, the metric is a reasonable proxy for what the year will produce, and the distinction between annualized run rate and trailing annual revenue is essentially editorial. For a hypergrowth business, the two figures can differ by a factor that reflects the entire growth trajectory โ because every earlier month on the curve was smaller than the month being annualized. A company that moved from a $492 million run rate in May to above $900 million in September collected considerably less than either annualized figure across those months. The months at the start of the curve were smaller than the month at the end, and annualizing only the end-point does not capture that.
This is not a design flaw in the metric. Run rate is intended to convey momentum and forward direction, not to summarize what has already been collected. The problem surfaces in the retelling. Bloomberg carried the qualifier in its body text โ “on track to generate… based on its performance this month” โ which is exactly where a qualifier belongs. The compression happens downstream, in newsletters, slides, and group chats, where qualifiers cost characters and certainty travels better than hedging. By the time a figure reaches most readers, it is usually the least qualified version in the chain, and there is no visible record of which layer they are holding.
“The useful habit is not scepticism about figures. It is a single question asked reflexively: what unit is this in, and what does that unit exclude? Run rate excludes time. That is the whole of it.”
Cognition’s position in the AI stack โ an application-layer agent built on top of foundation models, competing for the engineering workflow โ makes its growth numbers particularly sensitive to this dynamic. Agentic coding tools are still in a land-grab phase where usage can compress into short windows, meaning a strong month can look very different from an annual average. None of that changes the underlying trajectory, which is substantial. It does mean that the unit doing the most work in the coverage deserves its definition written next to it.
The same compression pattern appeared across this week’s AI coverage in adjacent forms: a pull-request share that was real but did not measure the share of engineering work; a published token rate that held flat while consumption per task climbed; a device weight that described only part of the system. Every figure was accurate. Every figure measured something narrower than the sentence it appeared in. The Cognition run-rate story is the week’s clearest version of that recurring structure, which is why it earns the analysis rather than the others.
Three Implications
FOR READERS OF AI COVERAGE
Every run-rate figure arrives with an implicit question attached: what was the base month, and how fast was the business growing into it? A doubling company’s run rate in month four of a strong growth curve is a very different object from the same number reported by a company growing at 10% annually. The number without the growth context is half a sentence.
FOR VALUATION COMPARISONS
Mixing a completed transaction with a reported conversation describes sentiment rather than a ladder of achieved prices. May’s $26 billion and September’s $48 billion are prices somebody paid; August’s reported ~$40 billion was a conversation. Treating all three as equivalent entries in a series misstates the evidential weight of each — and a pre-close figure that circulates after the round has closed is the same failure in miniature.
FOR THE AI APPLICATION LAYER
Cognition’s trajectory โ whatever the precise figure โ signals that autonomous coding agents have crossed from demonstration to commercial adoption at speed. The structural question that run rate cannot answer is whether the revenue base is durable: its composition across contract types, usage patterns, and customer concentration determines whether the month being annualized is a ceiling, a floor, or an outlier. Those are the figures that would complete the picture, and they are not yet established in the public record.
Business Engineer Framework
The Map of AI โ Where Application-Layer Agents Sit in the Stack
Cognition operates at the application layer of the AI stack โ the layer closest to the end-user workflow and furthest from the infrastructure bets that define the layers beneath it. The Map of AI plots 200+ companies across nine layers, showing where margin pools form, where commoditization pressure travels, and which layer a given revenue figure actually belongs to. Understanding Cognition’s run rate requires knowing what layer is generating it and how defensible that position is as foundation models improve.
Cognition’s growth from a $492 million run rate in May to above $900 million by its own September count is a genuine and substantial commercial signal for agentic coding as a category โ the trajectory is not in dispute, and the metric reporting it is standard practice. What the annualized run rate cannot carry, by definition, is the time it took to get there or the durability of the month being multiplied by twelve; those questions require different units, and those units are not yet in the public record. The single most useful habit when any run-rate figure appears is not to distrust it but to ask what it excludes โ and in this case, as in every case, the answer is time.
Every figure above is an annualized run rate — one month extrapolated across twelve — not revenue received. Nothing above says Cognition made, earned or booked $1 billion. Three versions of the claim exist and all three are given: Bloomberg’s headline says the company hits $1 billion; its body says the company is on track to generate that based on this month’s performance; and the company itself said its run rate topped $900 million. No inflation, overstatement or impropriety is alleged against anyone. Annualized run rate is a standard, widely used software metric and reporting it is ordinary practice; the point above concerns what the unit measures, not anyone’s honesty. No outlet is faulted — Bloomberg carried the qualifier in its own body text. The valuations carry different statuses: $26 billion in May was a completed raise; roughly $40 billion in August was reported talks; and the September round closed — Cognition’s own newsroom announced a Series E on 8 September raising over $2 billion at a $48 billion valuation, led by Andreessen Horowitz and Accel. Earlier reporting had put that round at around $1 billion at roughly $47 billion; the pre-close figures were lower on both counts and are still in circulation. Trailing and GAAP revenue, margins, burn, profitability, customer counts, churn, contract terms, net revenue retention and headcount are not established and do not appear above. Nothing above is investment advice or predicts anything about revenue, growth, the raise or the company’s prospects.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur, HackerNoon and Search Engine People. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
