Some links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.
OpenAI may earn more from serving users than it spends on inference, but that does not mean its flagship models have paid back the cost of building them. Epoch AI estimates that OpenAI’s GPT-5-era product bundle generated about $6 billion in revenue over roughly four months, with an estimated 30% gross margin. Once other operating costs are included, the bundle was close to break-even before research and development (R&D); after development costs, it likely had not recovered its investment.
That is a more precise conclusion than the headline “Not OpenAI.” Epoch’s analysis is an estimate based on public information and assumptions—not audited financial statements or proof that OpenAI as a company is unprofitable.
Three different answers to “Is OpenAI making money?”
| Measure | Epoch AI estimate | What it means |
|---|---|---|
| Gross margin | About 30% | Revenue exceeded the estimated cost of running the models for users. |
| Operating margin before R&D | Median estimate: -5%; 90% confidence interval: -30% to 10% | After other estimated operating costs, the bundle was roughly around break-even. The estimate excludes R&D and Microsoft revenue sharing. |
| Full model economics | Likely a loss over the period studied | The model’s estimated gross profit was not enough to cover the development spending attributed to it. |
These measures answer different questions. A positive gross margin can show that serving another user brings in more than the direct cost of inference. It does not establish that the business can pay for research, training, sales, infrastructure, administration, and the next generation of models.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
What Epoch AI studied
In an analysis updated March 6, 2026, Epoch AI examined what it calls the “GPT-5 bundle”: OpenAI products available while GPT-5 was the flagship, including GPT-5, GPT-5.1, GPT-4o, ChatGPT, the API, and other offerings. The analysis uses August 7, 2025, GPT-5’s release, through December 11, 2025, when GPT-5.2 was released, as the model’s economic period. That is an analytical boundary, not a claim that GPT-5 stopped earning revenue on that date. Older models can remain in use after a successor arrives.
#1 Best Overall
Epoch estimates the bundle brought in approximately $6 billion in revenue during those roughly four months. It estimates costs of about $4 billion for inference compute, $1 billion for staff compensation, $500 million for sales and marketing, and $200 million for legal, office, and administrative expenses. The figures are rounded estimates, not a published OpenAI income statement. Epoch AI explains its estimates and methodology.
Subtracting the estimated $4 billion in inference compute from $6 billion in revenue leaves about $2 billion of gross profit, or roughly a 30% gross margin. But the other listed operating costs total about $1.7 billion. That helps explain the reported median operating-margin estimate of -5% before R&D and Microsoft’s revenue-sharing arrangement. The uncertainty is substantial: Epoch gives a 90% confidence interval from -30% to 10%, so the estimate is consistent with results ranging from a meaningful operating loss to a modest positive margin.
Why development costs can turn a positive margin into a loss
Inference is the compute used to generate answers after a model is deployed. R&D is the much broader expense of creating and improving it: training runs, data preparation, research and engineering salaries, evaluation, safety work, failed experiments, and infrastructure used in development. These costs do not disappear because each user query earns more than it costs to serve.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problemsEpoch estimates OpenAI’s 2025 R&D spending at roughly $15 billion. Assigning that company-wide spending to one model is inherently difficult. As an illustrative, relatively conservative calculation, the researchers estimate about $5 billion of R&D in the four months before GPT-5 launched. That is more than the roughly $2 billion of gross profit they estimate for the GPT-5 bundle during its four-month flagship period. The comparison does not prove that GPT-5 alone cost exactly $5 billion to develop: it illustrates how the conclusion changes when development spending is considered.
A useful analogy is rapidly depreciating infrastructure. A model has to generate value while customers still consider it worth using—before a successor, competitor, or cheaper alternative pushes down its price or draws away usage. GPT-5 may continue to earn revenue after GPT-5.2’s release, but the relevant question is how much revenue remains attributable to the older bundle, and for how long.
The estimate depends on choices, not just arithmetic
OpenAI does not publish a model-by-model income statement with the information needed to independently reproduce this calculation. Epoch’s result depends on judgments about how to allocate shared costs and revenues. Its analysis draws on public reporting and company claims, staffing estimates, and inferred compute allocation; it does not have audited figures for GPT-5’s revenue or costs.
- Revenue: The approximately $6 billion estimate is for a bundle of products during the period, not GPT-5 alone. Revenue is inferred from reported or media-reported company figures.
- Inference: Compute costs are estimated and allocated to the period rather than taken from a model-specific bill.
- People and overhead: Compensation is divided between model operation and R&D using assumptions; sales, marketing, and administrative costs are also estimated.
- Free users: Users who do not directly pay may still create future conversion or distribution value, but their inference has a real cost. Epoch’s March update revised inference upward and sales-and-marketing estimates downward after excluding inference costs associated with free users from that category.
- R&D and lifespan: Development spending is not cleanly attributable to a single release, and ending the measured period at GPT-5.2 is a convenient boundary rather than an accounting rule.
Epoch reports commercial relationships with AI companies, including OpenAI and Anthropic. Its sensitivity analysis does not materially change the broad picture, but its confidence intervals reflect assumed ranges, not statistical measurements drawn from OpenAI’s books. The study therefore supports a reasoned estimate, not a definitive statement about OpenAI’s net income.
Microsoft makes “OpenAI profit” harder to interpret
Epoch estimates that Microsoft receives a share of relevant OpenAI revenue of about 20%, based on public reporting. It says the rate is not publicly confirmed and the arrangement is more complex than a simple percentage split. Epoch excludes this arrangement from its model-level operating-margin calculation and discusses it separately.
Rank #3
The distinction matters. A revenue share can reduce what OpenAI retains even when serving a model has positive unit economics. At the same time, Microsoft provides capital, infrastructure, distribution, and technology access. The arrangement is part of OpenAI’s particular corporate and financing structure; it should not be treated as a standard cost for every AI company.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.“Who profits from AI?” has a wider answer
A model provider is only one part of the AI business. Spending on AI can generate revenue for infrastructure suppliers before a model developer has recouped its research investment. GPU makers, cloud providers, data-center operators, networking and storage vendors, power suppliers, and construction and cooling firms may sell equipment, capacity, or services. That does not mean every supplier makes a profit on every AI project; it means their revenue can arrive on a different schedule from the model company’s eventual return.
Software distributors may also capture value. Microsoft, for example, can sell cloud usage and AI features through products businesses already use. Its U.S. Microsoft 365 Copilot Business pricing page, checked August 18, 2026, listed $25.20 per user per month on a monthly commitment and promotional pricing of $18 per user per month when paid yearly for eligible customers, with a qualifying Microsoft 365 license required. The stated promotion ran July 1 through September 30, 2026, subject to conditions. A listed price is not evidence of sales volume or profit; it illustrates how AI can be packaged into an existing software business.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →Application companies can try to earn money by selling coding assistants, customer-support automation, document processing, industry-specific tools, or workflow automation. They still face their own risks: dependence on a model provider’s API and pricing, competition, customer reluctance, and the possibility that a foundation-model company builds a competing product.
Customers may be the most important beneficiaries even if the AI vendor is not yet profitable. A business can come out ahead if time saved, improved output, or avoided costs are worth more than subscriptions, API usage, integration, human review, and the cost of correcting errors. The right test is measured business value—not a supplier’s gross margin or an AI label.
What could make AI companies profitable—or keep them from it?
The optimistic case is that inference gets cheaper, data centers run at higher utilization, and enterprise contracts create steadier, stickier revenue. Longer-lived models would have more time to repay development costs. Companies might also earn more from specialized services, advertising, higher usage, or AI agents tied to valuable business workflows. A fast-growing company can rationally accept current losses if investors expect future revenue to rise enough to cover investment.
The skeptical case is that costs keep pace with growth. Larger or more frequent training runs, staffing, data, and infrastructure could absorb revenue gains. Price competition can reduce what providers earn per query; free users can add serving costs without subscription income; and rapid releases can shorten the period a model commands premium prices. Weak customer adoption, heavy sales and support costs, and revenue-sharing obligations can further narrow returns. Revenue growth alone does not show that profit is approaching.
For a business deciding whether to buy an AI tool, the practical questions are whether it solves a specific problem, whether savings exceed the full cost of licenses and implementation, how sensitive data is handled, and whether the workflow can survive a price increase or provider change. Model-provider economics and a buyer’s return on investment are related, but they are not the same thing.
Bottom line
Epoch AI’s GPT-5 case study suggests that OpenAI’s flagship-era products could earn a positive margin on serving users while still failing to repay the cost of developing the model. The estimated bundle was roughly around break-even before R&D, not definitively unprofitable as a company. The unresolved business question is whether models can stay valuable long enough—and whether efficiency and revenue improve fast enough—to cover the cost of building their successors.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
