Did GPT-5 earn back its research and development costs?

Epoch AI's Jaime Sevilla, Hannah Petrovic, and Anson Ho built what they call "an educated guess" reconstruction of OpenAI's finances to ask whether GPT-5 was profitable to run and whether that was enough to recoup its R&D costs. On revenue of $6 billion against inference compute costs of $4 billion, they put the gross margin — revenue minus the cost of compute — at "about 30%... lower than the norm for software companies (where 60-80% is typical) but still higher than many industries." Once staff compensation ($1 billion), sales and marketing ($0.5 billion), and legal, office, and administrative costs ($0.2 billion) are added — bringing total costs to "close to $6 billion" — they find "OpenAI just barely broke even, even without including R&D," and conclude the company "likely ended up at a loss" once a revenue-sharing deal that hands Microsoft "about 20%" of the $6 billion is factored in.

The authors separate gross margin, which covers only the direct cost of running a model, from operating margin, which also includes marketing and staffing, and argue gross margin is "arguably a better proxy for long-run profitability" because S&M costs "typically grow sublinearly with revenue." Even using the more favorable gross-profit figure of about $2 billion, they find GPT-5's roughly four-month tenure — cut short because "Gemini 3 Pro had arguably surpassed the GPT-5 base model within three months" — was too short to earn back the R&D spent developing it. The mechanism they describe is a shrinking repayment window: a model's R&D costs must be recouped before the next model, from the same lab or a competitor, makes it obsolete.

This supply-side accounting of whether a lab's own product earns back its build cost sits opposite a cluster of demand-side surveys on whether AI's customers get value from it. Why do most enterprise AI pilots fail to deliver returns? measures whether buyers recoup value from deployment, while Are companies actually generating real value from AI investments? and Why do individual AI gains not translate to business results? track the same buyer-side question through survey data. Epoch AI's case study suggests the profitability question has a seller side as well as a buyer side, and the two can move independently: a model can find paying customers, as GPT-5 did, without the lab that built it covering its own development costs.

The authors are explicit that the case study is "an educated guess" built from public reporting rather than OpenAI's own books, and that it rests on one model generation at one company; they note "we don't have as much direct evidence on other models from other labs" beyond a passing comparison to Anthropic's "similar gross margins." The excerpt does not establish that AI model development is unprofitable as a rule, only that the GPT-5 case, as reconstructed, did not clear its own R&D bar in the time available — and the authors frame this as unsurprising for a fast-growing sector, comparing OpenAI's trajectory to Uber's "$32.5 billion deficit over 14 years" before its first profitable year.