Can OpenAI ever become profitable?

What would profitable actually mean for OpenAI?

OpenAI would count as truly profitable only when its normal business can fund model operation, research, staff, sales and infrastructure without another giant financing round.

A profitable quarter would tell us very little if it came from delaying a model launch, receiving cloud credits or recording an accounting gain. We are looking for recurring operating profit and positive cash flow after the company pays the real cost of serving customers and developing future products.

This stricter definition matters more in 2026 because "adjusted" profit has entered the debate. In October 2026, Anthropic told investors it was profitable on an adjusted basis, according to reports, while its draft IPO filing reportedly showed an $80.7 billion net loss for the first half. Both statements can be true at once. They measure different things.

OpenAI's accounts need the same care. Microsoft has historically provided computing support. Employees receive large amounts of stock compensation. The investor rights created under OpenAI's former structure still produced large non-cash charges in early 2026.

Three simple tests get us closer to the truth. Does each customer generate more revenue than the direct cost of serving that customer? Can the gross profit cover research, staff and sales? And can the remaining cash finance OpenAI's share of the infrastructure it has promised to use?

The company reaches meaningful profitability only when the answer to all three stays yes for several years, rather than for one carefully managed reporting period.

If you want more recent data on this point, see the latest figures in our Frontier AI Labs report ↓.

Why is OpenAI still losing so much money in 2026?

OpenAI is still losing billions because it runs a mass-market app, an enterprise software company and a frontier research lab at the same time, and all three grew more expensive in 2026.

The 2025 accounts set the baseline. Audited figures reviewed by the Financial Times showed about $13.1 billion of revenue against roughly $34 billion of costs and expenses. Research and development took about $19.2 billion of that.

The 2026 pattern is similar, only larger. In the first quarter, OpenAI reported about $5.7 billion of revenue and a $9.3 billion operating loss to shareholders, according to The Information. Research and development cost about $8.6 billion in that quarter alone, or about $1.51 for every dollar of revenue. In the second quarter, revenue reached $6.7 billion and the operating loss widened to $12.3 billion, according to the Wall Street Journal. Both loss figures include stock-based compensation.

OpenAI keeps funding several costly activities together. It trained and launched the GPT-6 family between September and early October 2026, serves more than 1.2 billion weekly ChatGPT users and builds an enterprise sales operation. It is commercializing today's model while financing tomorrow's model and reserving the hardware for the model after that.

Management could reduce the loss by limiting free use and slowing model development. But Anthropic, Google and Meta all released strong new models in September and October 2026. For now, OpenAI still puts leadership ahead of profit.

Are OpenAI's losses really as bad as the headlines say?

The headline net losses overstate the cash damage, but the operating loss is real and still growing.

Most of OpenAI's 2025 net loss of about $38.5 billion came from a non-cash revaluation tied to its former investor structure, according to the Financial Times. The same effect appeared again in 2026. The first-quarter net loss topped $21 billion, but about $12.4 billion of it came from revaluing investor rights, according to The Information.

Cash tells a calmer story. OpenAI burned about $3.7 billion in the first quarter of 2026, roughly three times the rate a year earlier, and held about $73 billion of cash and securities at the end of March, according to the same report. That is plenty of runway for the next few years.

The operating loss is the number that should worry investors. It counts real expenses, including stock compensation, and it rose from $9.3 billion in the first quarter to $12.3 billion in the second. Revenue rose by $1 billion between those quarters. The operating loss rose by $3 billion.

Each measure answers a different question. The net loss is the official accounting result. The operating loss shows how far ordinary expenses exceed revenue. Cash burn shows how quickly the bank balance falls, helped by non-cash costs and by partners who pay for data centers upfront.

Read the figures plainly: the headlines are inflated, cash burn is manageable for now, and the gap between revenue and costs is widening, not closing.

Measure	Approximate result	What it tells us
Revenue, Q2 2026 (reported)	$6.7B	Up 18% from $5.7B in Q1
Operating loss, Q2 2026 (reported)	$12.3B	Up from $9.3B in Q1; costs are rising faster than sales
Reported net loss, Q1 2026	More than $21B	About $12.4B came from non-cash investor-rights accounting
Cash burn, Q1 2026 (reported)	About $3.7B	Real cash strain is much smaller than the net loss
Cash and securities, March 31, 2026	About $73B	Enough to fund several years of losses at this pace

Is OpenAI's revenue growing faster than its costs?

Not in 2026. Revenue is still growing quickly, but costs grew faster between the first and second quarters.

In 2025, the picture looked better. Revenue more than tripled, while total costs rose by about 174%. Each new dollar of sales carried slightly less expense than before.

That trend reversed in the first half of 2026. Quarterly revenue rose 18%, from $5.7 billion to $6.7 billion, while the operating loss rose about 32%. The company spent heavily on training GPT-6, on compute for a fast-growing user base and on staff.

The run-rate has since jumped. OpenAI told investors its annualized revenue was nearing $50 billion at the end of September 2026, according to the Financial Times and Reuters. Earlier reports of nearly $70 billion used a different basis that counted revenue shared with cloud partners. OpenAI is now pitching $70 billion as a year-end 2026 target, not an achieved figure. Its recognized revenue for full-year 2026 was forecast at about $36 billion in a company presentation reported by the Financial Times.

Growth is no longer the only test. Anthropic's revenue rose about 45% in the third quarter of 2026, to $16.8 billion, while it reported an adjusted operating profit. OpenAI's sales are growing fast, but its spending plans assume even faster growth.

The revenue engine can support a path to profit. It does not yet support every compute and research commitment OpenAI has made.

Does ChatGPT have enough paying users?

Yes, ChatGPT has enough paying users to support a large profitable business, while most of its audience still pays nothing directly.

OpenAI said ChatGPT passed 1.2 billion weekly users on October 7, 2026, up from about 900 million in March. Its latest official count of paying consumer subscribers is still more than 50 million, from March 31, 2026; no newer figure has been published. Comparing the two gives a rough paid share of about 4%. The ratio mixes different dates, so treat it as a direction, not a precise figure.

Fifty million subscriptions can generate serious money. At an illustrative average of $15 per month, they would produce $9 billion a year. The real average will differ because OpenAI sells several plans at different prices.

Advertising is no longer a small experiment. The US pilot started on February 9, 2026. OpenAI said on August 31, 2026 that ChatGPT ads had passed a $1 billion annualized run-rate, with tens of thousands of advertisers in more than 40 countries. That is ten times the $100 million rate reported in March, though still only about 2% of OpenAI's total run-rate.

Competition for attention is growing. Similarweb estimates that ChatGPT's share of visits to major AI chatbot websites fell from 78% to 57% in the twelve months to August 2026, while Claude's rose from 2% to 9.6%. ChatGPT is still dominant, but each free user is now harder to keep.

The danger has not changed. Too many ads could weaken the trust that made ChatGPT valuable. OpenAI must earn more from free users while keeping the answers good enough that they come back.

Current measure	Approximate figure	What it means
Weekly ChatGPT users (Oct 2026)	More than 1.2B	OpenAI has one of the largest consumer audiences in technology
Consumer subscribers (Mar 2026, latest official)	More than 50M	Direct paid demand is already substantial
Rough subscriber-to-user ratio	About 4%	Most weekly users remain outside the paid base
Illustrative revenue at $15 per month	$9B yearly	The subscriber base alone could support a major software company
Advertising run-rate (Aug 2026)	More than $1B	Ten times the March pilot, but still a small share of revenue

Can enterprise customers become OpenAI's profit engine?

Enterprise is still OpenAI's best route to profit, because businesses pay more for completed work, security and integration than for a chat window.

OpenAI said enterprise activities produced more than 40% of its revenue at the end of March 2026, and it aims for parity with consumer revenue by the end of the year. It reported more than 2.5 million business customers in September 2026, up from about one million in November 2025, and more than 9 million paying business seats in March.

The product line has moved toward agents. In July 2026, OpenAI folded Codex into the ChatGPT desktop app and launched ChatGPT Work, an agent mode that returns finished documents, spreadsheets and web apps. In October 2026, Atlassian expanded its partnership so that GPT-6 models power its enterprise agents, although the contract value was not disclosed.

Earlier deals still matter. Accenture is rolling out ChatGPT Enterprise across tens of thousands of employees. Snowflake signed a $200 million agreement to bring OpenAI models into customer data environments. OpenAI models and Codex are available through AWS.

These contracts are hard to replace once they work. An AI system connected to customer records, software repositories and approval rules cannot be swapped out casually.

The cost is real, too. Deployments need engineers, consultants and support. Enterprise can carry OpenAI toward profit only if each new contract needs less human help than the one before. Anthropic's results suggest the model can work: enterprise and coding customers drive most of its revenue, and it reported an adjusted operating profit in the third quarter of 2026.

If you want more recent data on this point, see the latest figures in our Frontier AI Labs report ↓.

Has Anthropic overtaken OpenAI in enterprise AI?

Yes, on most available measures, although OpenAI's own workplace business keeps growing.

Revenue is the clearest sign. Anthropic reported $11.6 billion of second-quarter 2026 revenue, against $6.7 billion for OpenAI, according to Reuters. Most of Anthropic's revenue comes from businesses and developers. The two companies count cloud-partner revenue differently, so the gap is not an exact like-for-like comparison, but the direction is clear.

Spending data points the same way. Menlo Ventures estimated that OpenAI's share of enterprise spending on large language model APIs fell from about 50% in 2023 to 27% in 2025, while Anthropic reached about 40%, and 54% in coding. No newer independent survey of the same kind was found for 2026.

Coding shows the gap most clearly. A JetBrains survey of professional developers from May to July 2026 found that 39% used Claude Code at work, against 16% for Codex. Codex use still grew about fivefold from January.

Companies also mix models more than before. A September 2026 Stacklok survey found that about 92% of sampled North American enterprises use more than one model provider. That weakens the loyalty OpenAI enjoyed when GPT was the obvious default.

OpenAI is answering by selling the whole workplace system rather than the model alone. Codex, ChatGPT Work, ChatGPT Enterprise and Frontier combine models with permissions, data access and interfaces. That is harder to swap than an API call. OpenAI can still win the broader workplace market, but it is now the challenger in the most profitable part of it.

Why can Anthropic report a profit when OpenAI cannot?

Because Anthropic earns more revenue from each dollar of compute and spends less on consumer users, at least on the adjusted figures it shared with investors.

On October 14, 2026, Anthropic told prospective investors that third-quarter revenue reached $16.8 billion, with an 11% adjusted operating margin and positive cash flow, according to the Financial Times. It also reported about $1.70 of revenue for each dollar spent on compute. These figures are not audited, and the company has not said which costs the adjusted measure leaves out.

The business mix explains much of the gap. Anthropic sells mostly to companies and developers, who pay for heavy use. OpenAI serves more than a billion weekly users, most of whom pay nothing. Free users bring reach and data, but they also consume compute that subscriptions and ads have to cover.

Anthropic is not a clean model of profitability either. Its draft IPO filing reportedly showed an $80.7 billion net loss in the first half of 2026, mostly from financing-related accounting, and more than $600 billion of noncancelable infrastructure commitments, according to the Wall Street Journal in October 2026.

The lesson for OpenAI is mixed. Enterprise-heavy revenue can produce an operating profit even during a model race. But both companies have committed to compute bills that today's revenue cannot pay on its own.

Can OpenAI keep its margins while AI models get cheaper?

OpenAI can protect its margins only by selling outcomes and workflow control, because token prices keep falling and rivals now set some of them.

The price drop has been dramatic. The original GPT-4 API charged about $30 per million input tokens and $60 per million output tokens. OpenAI's flagship GPT-6 Astra, released in September 2026, lists at $10 for input and $50 for output. The mid-tier GPT-6 Sol is listed at $2 and $10 by third-party price trackers, more than 90% cheaper on input than GPT-4.

Competitors push from every side. Google launched Gemini 4 Argon on October 13, 2026 at an introductory $2 per million input tokens and $10 per million output tokens. Anthropic's Claude Opus 5.5 lists at $4 and $20. DeepSeek's V4.1 Flash costs $0.15 for input at off-peak times. On Arena's leaderboards in October 2026, Google led overall text, while Anthropic led web development and agents. OpenAI's GPT-6 Astra ranked second in both of those categories.

OpenAI still charges a premium for its top model, which helps margins. But that premium only holds while customers believe the model is clearly better, and in October 2026 the independent rankings did not show that.

A model API alone will struggle to keep software-like margins. Price comparisons are easy, and most companies already use several providers. OpenAI needs revenue from products where the model is one piece of a larger service.

Subscriptions, enterprise agents, coding systems, advertising and commerce all move in that direction. The product becomes a resolved support case, a finished software task or a completed purchase. Customers care less about the number of tokens when the result saves real time or money.

If you want more recent data on this point, see the latest figures in our Frontier AI Labs report ↓.

Will AI agents make OpenAI richer or poorer?

AI agents can become OpenAI's most valuable product and its fastest way to waste compute.

A normal chatbot might answer after one model call. An agent can search files, browse the web, use software, check its work and try again. One request may trigger dozens of model calls before the user sees a result.

That is already how OpenAI's business customers use it. In June 2026, OpenAI said Codex generated 64% of the output tokens across Codex and ChatGPT for enterprise customers. OpenAI reported more than 5 million weekly Codex users on June 2, 2026, and reports put the figure at about 7 million by mid-July.

Here is the pricing problem. A fixed monthly subscription works well when usage stays predictable. It becomes dangerous when one active customer can consume more compute than the fee covers. In July 2026, Sam Altman reset weekly usage caps as Codex and ChatGPT Work passed user milestones, which shows how closely OpenAI has to manage that balance.

The upside is much larger than chat. A company may argue over the price of a million tokens, yet happily pay more for an agent that closes a support ticket or repairs a software bug. Coding has strong economics because useful software can be worth far more than the compute behind it.

The same logic fails for many casual requests. Spending $10 of compute to replace $100 of paid work creates a strong margin. Spending $20 to satisfy a user on a $20 monthly plan leaves nothing for research or infrastructure.

OpenAI needs strict usage limits, model routing and prices that rise with the work performed. Agent demand can transform the business, but only when revenue follows the compute bill.

Can inference get cheaper faster than usage grows?

OpenAI is improving its delivery economics, but more slowly than its usage and ambitions grow.

Gross margin was about 39% in the first quarter of 2026, up from about 33% a year earlier, according to The Information. Cost of revenue, mostly inference, was about $3.5 billion that quarter. A separate estimate based on the audited 2025 accounts put full-year 2025 gross margin near 43%. The two figures use different periods and cost definitions, so they do not show a decline; both point to margins stuck around 40%.

There are signs of real savings. In June 2026, OpenAI engineers found an optimization that cut inference costs by more than half on some workloads, according to The Information. The company has not confirmed how widely it applies.

The new model lineup helps too. GPT-6 Astra targets the hardest work, Sol balances capability and cost, and Luna handles high-volume tasks at about $0.10 per million input tokens, according to price trackers. Routing a simple question to Luna instead of Astra can cut the input price by about 99%.

The harder part is demand. Reasoning models think longer. Agents repeat steps. ChatGPT added about 300 million weekly users between March and October 2026, most of them on free plans. Every saving creates room for a heavier product.

Profit arrives only when OpenAI keeps part of each saving instead of spending it on longer answers, more reasoning and lower prices. So far, margins around 40% suggest it is keeping only a small share.

Will frontier research keep eating the profits?

Yes, for now. Research still costs more than OpenAI's entire revenue.

Research and development reached about $19.2 billion in 2025, about 147% of revenue. In the first quarter of 2026, it was about $8.6 billion against $5.7 billion of revenue, or roughly 151%. The ratio did not improve.

OpenAI chose a race with no natural finish line. It launched the GPT-6 family between September and October 2026. In the same weeks, Anthropic released Claude Opus 5.5, Google released Gemini 4 Argon and Meta released Muse Wonder. Each launch makes a long pause more dangerous.

Microsoft, Alphabet and Meta can fund frontier research with profits from cloud, advertising and mature software. Meta alone guided 2026 capital spending of $130 billion to $145 billion. OpenAI asks its AI products to fund the research needed to defend those same products.

The company has shown some discipline. In April 2026, the New York Times reported that Sam Altman had cut back side projects to focus spending on ChatGPT, coding and enterprise products. The push into Codex and ChatGPT Work points toward work with a clearer route to revenue.

The number to watch is simple: how much of each revenue dollar goes back into research. OpenAI could spend $35 billion a year on research and still be profitable if revenue reached $120 billion with healthy margins. Profit stays distant if research keeps climbing dollar for dollar with revenue.

Has OpenAI's infrastructure spending gone too far?

OpenAI's compute plan is now large enough to decide whether the company is ever profitable, and the plan's numbers do not all agree.

The Wall Street Journal reported in October 2026 that projected compute spending stands at about $750 billion through 2030, up from about $600 billion earlier in the year. A company presentation reported by the Financial Times in September put computing and infrastructure spending at about $856 billion through 2030. The two figures likely use different scopes; neither is an official total.

The same presentation projected about $278 billion of negative free cash flow from 2026 to 2030, even as revenue rises to about $350 billion in 2030. In plain terms, OpenAI's own plan does not pay for itself this decade.

New projects keep arriving. In August 2026, Nvidia announced credit support for an Ohio campus of up to 8 gigawatts, with OpenAI as the planned tenant under a 20-year lease. Earlier deals include Oracle, AWS, Broadcom, AMD and Cerebras. Many overlap, so they cannot be added together.

Partners carry part of the upfront cost. SB Energy will build, own and operate the Ohio site, for example. That lowers OpenAI's immediate cash needs, but it turns capital spending into long-term lease and purchase obligations.

Execution risk is rising. OpenAI's compute chief, Sachin Katti, is leaving in November 2026, the second change in compute leadership this year. The Financial Times counted at least 47 major US data-center projects stalled or canceled in 2026.

Utilization is the central test. A cluster running paid workloads around the clock is valuable. The same cluster becomes a trap if demand misses the forecast or customers move to cheaper models.

If you want more recent data on this point, see the latest figures in our Frontier AI Labs report ↓.

Does Microsoft make profitability easier or harder?

Microsoft made OpenAI's growth possible, but the current deal still sends money and future cloud spending back to Microsoft.

Microsoft supplied early capital, Azure supercomputers and global enterprise distribution. Without them, OpenAI would have found it much harder to train large models and sell them to major companies.

The April 27, 2026 amendment sets the current terms. OpenAI's revenue-share payments to Microsoft continue until 2030, subject to a total cap. The Information reported the share at about 20% and the cap at about $38 billion; neither figure is official. Microsoft stopped paying its own revenue share to OpenAI. OpenAI had also committed in October 2025 to buy an additional $250 billion of Azure services.

Microsoft keeps a license to OpenAI models and products through 2032, now non-exclusive, and OpenAI can serve its products across other clouds. That freedom has already led to larger deals with Amazon, Oracle and others.

The cap is good news for OpenAI. It limits how much of future growth goes to Microsoft. At about 20% of revenue, the share is still a heavy cost until the cap is reached.

Microsoft is both an accelerator and a toll collector. The partnership helped create OpenAI's revenue, while the revenue share and Azure purchases absorb part of the profit that revenue may produce.

How much revenue would OpenAI need to break even?

On its latest quarterly cost base, OpenAI would need roughly $100 billion to $150 billion of annual revenue before an operating profit becomes realistic.

The same test on the 2025 accounts gave a range of $40 billion to $60 billion. Costs rose sharply in 2026. In the second quarter of 2026, OpenAI reported $6.7 billion of revenue and a $12.3 billion operating loss, so total costs were about $19 billion for the quarter.

OpenAI has not published a second-quarter gross margin. If we assume it stayed near the 39% reported for the first quarter, direct delivery costs were about $4.1 billion. That leaves about $15 billion a quarter, or about $60 billion a year, of research, sales, administration and stock compensation.

At a 39% gross margin, OpenAI would need about $153 billion of revenue to cover those expenses. A 60% margin lowers the requirement to about $100 billion. At 70%, it falls to about $85 billion.

This calculation holds operating expenses flat, which will not happen. It also includes stock compensation, which does not use cash. A cash break-even would come earlier than an operating one.

The range gives the right order of magnitude. OpenAI's latest run-rate is nearly $50 billion. Reaching $100 billion means doubling it; reaching $150 billion means tripling it. The company's plan expects about $350 billion of revenue in 2030, so the target is within its ambitions. The problem is that each step up in revenue has so far come with a similar step up in spending.

Assumed gross margin	Revenue needed to cover about $60B of other yearly expenses
39%	About $153B
50%	About $119B
60%	About $99B
70%	About $85B

When could OpenAI realistically become profitable?

Around 2030 at the earliest, and more likely later, unless spending slows.

OpenAI has not published a break-even year. Its own forecast, as reported by the Financial Times in September 2026, shows about $278 billion of cumulative negative free cash flow from 2026 to 2030. A plan that burns that much cash cannot include several profitable years before 2030.

The revenue side looks strong. From a run-rate of nearly $50 billion, three years of 40% growth would reach about $137 billion, within our break-even range. The company's plan is far more ambitious, at about $350 billion in 2030.

The company also has time and access to capital. In March 2026, OpenAI raised $122 billion of committed capital at an $852 billion post-money valuation, and Nvidia and SoftBank reportedly paid their final $10 billion installments by October 1, 2026. In late September, OpenAI was reported to be seeking at least $30 billion more at a valuation of about $1.4 trillion before the new money. That round has not closed. Sam Altman said in September that OpenAI would not go public in 2026, and Reuters reported that a listing is now expected in 2027, with no date guaranteed.

Capital does not fix weak economics by itself. The FT's report suggested the March money could run out by 2028 at the planned spending pace. OpenAI will likely need to raise again before it earns its way out.

A realistic route to profit needs several changes together. Enterprise products must become the largest revenue source. Gross margin must climb well above 40%. Ads must keep scaling. Agents need pricing tied to the work they do. And research and compute must grow more slowly than sales.

OpenAI could report a profitable quarter earlier by delaying projects. A sustained yearly profit looks more credible around 2030 or after. A longer model race or another expansion of the compute plan would push it further.

Can OpenAI ever become profitable?

Yes, OpenAI can become profitable, but its strategy keeps pushing the moment back on purpose.

OpenAI has the ingredients of a profitable company. ChatGPT reaches more than 1.2 billion people a week. Tens of millions pay. Advertising passed a $1 billion run-rate in its first year. Codex and ChatGPT Work give it a growing agent business, and more than 2.5 million companies are customers.

The comparison with Anthropic is now harder. Anthropic, with a smaller consumer audience, earned more revenue than OpenAI in the second quarter and reported an adjusted operating profit in the third. That weakens the argument that a frontier lab must lose money while the race continues.

OpenAI keeps reinvesting almost without a ceiling. Higher revenue becomes a reason to reserve more compute, train larger models and enter more markets. Its costs grew faster than its revenue in the first half of 2026, and its own plan burns about $278 billion of cash through 2030.

We still see a credible profitable business inside OpenAI. It would charge for work completed, route tasks to cheaper models, keep research below gross profit and build compute in stages as paid demand appears.

Our final judgment: OpenAI can become operationally profitable, but probably around 2030 or later, not before. The company has proved that customers will pay. It still has to prove that it can stop spending money faster than it arrives, and its closest rival has now shown that this is possible.

If you want more recent data on this point, see the latest figures in our Frontier AI Labs report ↓.