Reporting says OpenAI's gross margin was about 33% in 2025, down from 40% in 2024 and below an internal forecast of 46%, with inference costs of about $8.4 billion. A 2026 recovery to around 39% has been reported by one tracker without a named source, so treat it as unconfirmed.
Gross margin and inference are two of the most quoted and least understood numbers in AI. This guide explains both in plain terms, lays out the figures with their sources and dates, and shows why a single company can honestly be said to have very different margins depending on the definition. It is not investment advice. OpenAI is private, has not published audited accounts, and has not commented publicly on the leaked documents. Everything is as of 19 October 2026.
Key numbers at a glance
- 2024 gross margin: 40% (The Information, via a summary published 25 February 2026).
- 2025 gross margin: 33%, versus an internal forecast of 46% (same source).
- Long term target: above 60% by 2030, with a projection near 67% in the same summary (The Information, via that summary).
- 2025 inference cost: about $8.4 billion, roughly four times 2024 and above a summer 2025 forecast of $6.6 billion (same source).
- 2025 inference split: about $3.9 billion serving free users and about $4.5 billion serving paying users (same source).
- 2026 inference cost expected: $14.1 billion, about two thirds for paying users (same source).
- 2030 inference cost projected: about $85 billion, about 94% for paying users (same source).
- Q1 2026 gross margin: about 39% (a third party research page, no named source, so unconfirmed).
- 2025 cost of revenue line: $7.5 billion (leaked 2025 accounts as relayed by a secondary outlet, no attribution given for this line).
- 2025 payments to Microsoft: about $17.2 billion, of which $10.59 billion was for research and development (leaked accounts, via secondary coverage).
- Computing capacity: 0.6 GW in 2024 and 1.9 GW in 2025 (OpenAI CFO blog, 18 January 2026, via Reuters).
- Anthropic 2025 inference cost: about $2.7 billion (The Information, via the same summary), for comparison.
- Anthropic 2024 gross margin: about negative 94%, and a 2025 projection of 40% (same summary).
What gross margin means
Gross margin is the share of revenue left after subtracting the direct cost of delivering the product. A 33% gross margin means a third of each revenue dollar is left after paying to serve the customer. It is not profit. It ignores research, sales, salaries and most overhead. Its job is to tell you whether each additional customer makes money before the company pays for everything else.
For ordinary software, the direct cost is small, so margins of 70% to 80% are normal. AI is different, because every answer uses expensive computer chips. That is the heart of the story.
The catch is that "direct cost" is a judgment call. Is the cost of training the next model part of serving this month's customers? Is a revenue share paid to a cloud partner a direct cost or a reduction in revenue? Do you count the cost of serving people who pay nothing? Different answers produce different margins for the same company in the same year, and that is exactly what shows up in the reporting.
What inference is, and why it dominates
Training is the one time, very large expense of building a model. Inference is the ongoing expense of running that model each time someone asks it something. A simple analogy is a factory: training is building the factory, and inference is the electricity and labor to run it for each product.
The summary of The Information's reporting says OpenAI's inference cost in 2025 was about $8.4 billion, roughly four times the year before. It also says that was above a forecast of $6.6 billion made in summer 2025, and attributes the overrun to higher demand than expected and higher on demand server prices. On demand pricing is the rate you pay when you rent computing power by the hour with no long contract, which is usually higher than committed rates.
Two details in that reporting help a non specialist. First, nearly half of 2025 inference spending, about $3.9 billion, went to serve users who do not pay. Free users are a marketing and growth cost that also shows up in cost of goods. Second, the same summary says the share of inference going to paying users is expected to rise from about 66% in 2026 to about 94% by 2030, which is part of how the company hopes margins improve.
One research page also says inference cost per query is down about 95% since the launch of GPT-4 in early 2023, but it names no source, so the figure is unverified. Even if it is right, total spending can rise sharply when usage grows faster than unit costs fall.
The forward plans point the same way. OpenAI told investors in February 2026 that compute spending through 2030 would be around $600 billion, and a July 2026 presentation reported by the Financial Times put computing power and infrastructure at about $856 billion for 2026 to 2030, according to TheNextWeb on 19 September 2026. TheNextWeb notes the two figures may not cover the same categories, and neither splits training from inference, so they show the scale of the bill rather than the cost of serving users.
What OpenAI itself says about compute and revenue
OpenAI has not published a margin, but its finance chief has described how revenue and computing capacity move together. In a blog post on 18 January 2026, reported by Reuters, chief financial officer Sarah Friar said annualized revenue passed $20 billion in 2025, up from $6 billion in 2024, and that growth was "closely tracking an expansion in computing capacity". She gave that capacity as 1.9 gigawatts in 2025, up from 0.6 gigawatts in 2024. The post gave no cost or margin figure.
Read alongside the leaked numbers, that statement points to the central tension. If revenue rises roughly in step with capacity, then the cost of that capacity rises too, and margin depends on whether each gigawatt earns more over time. Friar also said OpenAI was keeping a "light" balance sheet by partnering rather than owning, with contracts flexible across providers and hardware. Renting keeps data centre construction off OpenAI's own books, but the capacity is then paid for as an ongoing cost rather than built once.
The user base explains the free user share of that cost. When it closed its $122 billion round on 31 March 2026, OpenAI said it had more than 900 million weekly active users and more than 50 million subscribers, and that enterprise sales had reached 40% of revenue, according to SiliconANGLE. A small paying share of a very large audience is exactly the pattern that makes free usage such a large line in the reported inference bill.
The margin story across 2024 to 2026
The reported path is a fall and then a hoped-for recovery.
In 2024, gross margin was 40%. In 2025, it fell to 33%, well short of a 46% internal forecast, according to The Information as summarized in February 2026. The cause given is inference costs running above plan.
The same summary reports that paying user "compute margin," meaning the margin on paying customers after the cost of running the models for them, improved from about 35% in January 2024 to about 52% by the end of the prior period and about 70% in October. The summary's wording on the dates is loose, so read these as a trend line, not as exact quarter by quarter results. The important point is that the average margin can be pulled down by free users even while the paying business improves.
For 2026, a third party research page says margin recovered to about 39% in Q1 and that it is aiming higher. It does not say who reported that, and no Q2 margin has been reported publicly. The 39% should not be quoted without the original source.
The long term target in the reporting is above 60% by 2030, with a projection near 67%. That is a plan.
Why the numbers do not line up neatly
If you try to cross check, the figures do not obviously agree. The leaked 2025 accounts reportedly show a cost of revenue line of $7.5 billion, relayed by one outlet without attribution, while The Information's reporting puts inference alone at about $8.4 billion and the margin at 33%. A cost line smaller than the inference bill cannot both be the full direct cost behind a 33% margin. There are plausible explanations: the margin might be calculated on a different revenue basis, such as net of partner payments, some costs might sit in other lines, or the documents may define the lines differently. Public reporting does not settle it. That is why each figure here is listed with its own source rather than blended with others.
The leaked accounts also show about $17.2 billion paid to Microsoft in 2025, of which $10.59 billion was for research and development. That means a large share of compute spend sits in research and development rather than in cost of revenue. In plain terms, money spent to train models sits below the gross margin line, while money spent to run them for customers sits above it. That split decides what the margin looks like, and it is a reporting choice, not a fact of nature.
Comparing with Anthropic
The same summary reports Anthropic's 2024 gross margin at about negative 94% and a 2025 projection of 40%, with 2025 inference costs near $2.7 billion. A separate tracker says Anthropic's margin in 2026 might be anywhere from 44% to 60% depending on whether training cost is netted against compute spend, and that the Financial Times reported a figure above 80% in September 2026 before partner revenue share and training costs. That narrower definition is a good example of how much the label matters. See Anthropic's revenue, losses and margins for those numbers in full.
The point is not that one lab is better. The point is that a 33% figure and an 80% figure can both be true under different definitions, and nobody should compare them directly.
What this means for buyers of AI services
Most customers will never see a lab's margin, but the numbers matter to you in three ways.
Price stability. A provider with thin margins and heavy spending has an incentive to adjust prices, rate limits and free tiers. Public reporting does not show any announced price change tied to margins, and this post makes no prediction. Reading a pricing page carefully is still good practice.
Concentration. If your product depends on one model provider, its finances are part of your risk. A backup plan is cheap compared with a forced migration.
Capacity. Spending figures show that supply of computing power, not demand, is the constraint labs talk about. OpenAI's spokesperson told the Wall Street Journal that demand for AI exceeds available compute supply, as reported by Fortune in November 2025.
How to read the next margin headline
When a new margin figure appears, run it through four quick questions before you repeat it. Which year or quarter does it cover? Does it include the cost of serving free users? Is the cloud partner's share treated as a cost or taken off revenue? Is model training counted in cost of revenue or in research and development? If the report does not answer, the figure is a headline and not a measurement.
It also helps to ask whether the number comes from the company, from a leak, or from an analyst. A company statement is the only one that carries any accountability, and even then it may use a narrow definition. A leak shows what was in a document on a given day. An analyst estimate is a model. The guide to reading AI lab financial reports walks through all five checks, and the figures here are a good practice set.
What is firm and what is uncertain
Firmer: OpenAI's 2025 gross margin was well below its forecast, and inference was the main reason, according to reporting that traces to The Information. The direction of the cost is not in dispute. It is large and growing.
Uncertain:
- The exact definition behind the 33% figure and the revenue base used.
- The Q1 2026 margin of about 39%, which has no named source.
- The paying user margin trend, because dates in the summary are loose.
- Every projection for 2026 and 2030.
- Any claim about a provider's internal unit costs. No primary source has published them, and this post makes none.
What to watch next
- A public prospectus, which would have to define gross margin and show cost of revenue explicitly.
- Whether OpenAI or Anthropic ever reports margin on a stated, consistent basis.
- Whether inference spending in 2026 lands near the $14.1 billion expectation.
- Any change in what free users can access, since they account for a large share of reported inference spending.
- Q2 and Q3 2026 margin figures, none of which has been reported.
Quick check
What was OpenAI's gross margin in 2025?
Reporting traced to The Information says about 33%, down from 40% in 2024 and below an internal forecast of 46%. OpenAI has not published audited accounts, so this is a reported figure.
Why is inference so expensive?
Every answer a model gives runs on costly chips that must be paid for, whether you own them or rent them by the hour. Unlike traditional software, the cost grows with every user request.
Can I compare one lab's margin with another's?
Only if both use the same definition. Costs such as training, partner revenue share and free users can be included or excluded, which moves the result by tens of percentage points.
Sources
- OpenAI and Anthropic miss gross margin targets as inference costs rise, BigGo Finance, 25 February 2026, summarizing The Information.
- OpenAI revenue 2026: ARR reportedly nears $70B, Value Add VC, updated 15 October 2026.
- Is Anthropic profitable in 2026?, Value Add VC, updated 15 October 2026.
- OpenAI reports rapid revenue growth, larger losses, Let's Data Science, 16 June 2026.
- OpenAI revenue, valuation and funding, Sacra, data through about October 2026.
- OpenAI CFO says annualized revenue crosses $20 billion in 2025, Reuters via Yahoo Finance, 19 January 2026.
- OpenAI just closed a record breaking $122B funding round, SiliconANGLE, 31 March 2026.
- OpenAI's $856bn compute plan versus the $600bn reset, TheNextWeb, 19 September 2026.
- OpenAI says it plans to report stunning annual losses through 2028, Fortune, 12 November 2025.
Last reviewed: 19 October 2026.
