OpenAI’s revenue for the quarter ended September was $10 billion. Its operating loss was $10.9 billion.
The figures, reported by The Wall Street Journal on Nov. 10 citing documents it had reviewed, show revenue rising about 50% from the previous quarter and crossing the $10 billion mark for the first time. The operating loss widened from $10.1 billion three months earlier. On an adjusted basis, the operating margin improved to minus 109% from minus 150%.
That last number is the one the company wants read closely. A margin of minus 109% means that for every dollar of revenue booked, roughly $2.09 of cost was incurred. A margin of minus 150% meant $2.50. The direction is right and the level is still extraordinary, which is roughly the story of the business in a single ratio.
The numbers were presented on an investor call on Tuesday. Revenue growth came mainly from enterprise subscriptions and from API usage, according to the report. Compute spending continued to grow faster than revenue did, which is the sentence that explains why a company with a $10 billion quarter is still losing more than it earns.
OpenAI is not the only lab in this position, but it is the one whose numbers get compared most often. Anthropic, the developer of Claude, reported revenue of $17 billion over the same period and an operating loss that narrowed to $567 million. Anthropic is preparing to list its shares next week, which gives its figures a second audience and its bankers a reason to present them carefully.
The comparison is stark and should be handled with care. A loss of $567 million on $17 billion of revenue is a loss of about 3% of sales. OpenAI’s adjusted margin implies a loss larger than its entire top line. The two companies sell into the same market with different products and different contract structures, and the gap between them is either a difference in efficiency or a difference in what each company chooses to expense when it computes an adjusted figure.
Private company accounts are self-reported, and the definition of adjusted operating margin is not standardized across the industry. Nothing in the disclosed figures says whether both companies treat compute commitments the same way, whether they capitalize the same costs, or whether one is counting depreciation that the other defers. Until both are public, the comparison measures direction better than it measures discipline.
What can be compared is the shape of the growth. OpenAI’s revenue rising 50% in a quarter is a rate that few enterprise software companies have sustained at any size. It also means the base is compounding from a number that was itself small a year ago, so each quarter of growth adds a larger absolute amount of compute to pay for. Gross additions that big require capacity that has to be contracted in advance.
That is the structural problem. A company selling subscriptions and API calls can price them however it likes. The cost of the machines behind them is negotiated months ahead, on multi-year terms, with suppliers who have their own order books.
If demand doubles and capacity does not, the company either turns customers away or pays a premium. If capacity arrives and demand slows, the company owns the depreciation.
The competitive response has been visible in two moves. OpenAI named Dali Rajic as chief revenue officer, a hire that points at enterprise selling rather than consumer growth. It also cut the price of its latest model, the standard method for buying share in a market where buyers route requests by cost per token and switch when a cheaper option appears.
Price cuts trade margin today for volume tomorrow, and volume is what spreads fixed capacity costs. The logic works if the volume arrives and if the cheaper model reduces the compute required per request. It fails if the price cut mostly moves existing customers onto a lower rate.
The enterprise market is where this gets decided, because enterprise buyers sign annual contracts and consumer subscribers cancel. Anthropic’s revenue mix is weighted toward developers and coding tools, which are high-frequency, high-token workloads with customers who pay by usage. OpenAI’s mix includes a large consumer subscription base, which pays a flat fee and consumes support, app store fees and moderation along with compute.
Investors have funded the losses on the assumption that scale eventually bends the cost curve. Cheaper inference per token, better utilization and more efficient models all argue for that. The counterargument is that the frontier keeps moving, and each new model generation costs more to train than the last, so the savings get spent on the next run.
There is a timing question buried in the report. OpenAI’s loss widened in absolute terms while its margin improved in percentage terms. A company can improve its unit economics and still lose more money every quarter for several quarters. The market has tolerated that pattern in platform businesses before, and it has stopped tolerating it quickly when funding conditions change.
The next quarter will be read for one thing: whether the adjusted margin keeps moving toward zero at the same rate. A further improvement of forty points would put the company close to breakeven on an adjusted basis within a year, which would change the conversation entirely. A margin that stalls while revenue doubles would suggest the cost structure is fixed to growth rather than to scale.
The question left open is whether the loss is narrowing because the technology is getting cheaper to run or because the company is choosing to run less of it.
