OpenAI

IWhat it does

OpenAI builds large AI models and sells access to them. The product most people know is ChatGPT, a chat assistant. The company also sells the same models to developers through an API (a way for other software to call the model), sells team and enterprise plans, and sells Codex, a coding assistant. ChatGPT reached about 1.2 billion weekly users in September 2026, as relayed by a data compilation³. In March 2026 the company was reported to have more than 50 million paying subscribers, and enterprise customers were said to account for more than 40% of revenue¹¹.

A second business sits underneath the first: buying and building computing power. OpenAI has signed very large deals for chips and data centres. Nvidia announced a letter of intent for at least 10 gigawatts of its systems, with Nvidia intending to invest up to $100 billion in OpenAI as each gigawatt is deployed⁷. AMD announced a deal for 6 gigawatts of its GPUs and gave OpenAI a warrant, a right to buy up to 160 million AMD shares, that vests in steps tied to deployment and to AMD’s share price⁸. A gigawatt is a measure of electrical power, used here as a rough unit of data-centre size.

OpenAI is therefore two things at once: a software and subscription company with fast-growing revenue, and a capital-heavy infrastructure buyer. Most of the debate about its value comes from the gap between those two.

IIHow it makes money

OpenAI earns money from consumer subscriptions, from business and enterprise seats, from API usage fees, and, since early 2026, from advertising. A data compilation reports that ads reached a $1 billion annualised run rate within about 200 days of launching in February 2026, attributed to OpenAI via MediaPost³. I could not open the original, so treat that as press-reported.

Revenue is growing fast, but costs are growing faster in dollar terms. Leaked audited 2025 figures reported by Fortune and the Financial Times show revenue of $13.07 billion, total costs and expenses of $34 billion, and an operating loss of $20.92 billion¹⁰. The year before, revenue was $3.7 billion and the operating loss was $8.78 billion¹⁰. These are leaked documents, not a company filing. OpenAI declined to comment, according to the coverage I saw, which I could not open directly, so I treat the numbers as press-reported.

In 2026 the Wall Street Journal reported quarterly revenue of $5.7 billion in the first quarter and $6.7 billion in the second, with operating losses of $9.3 billion and $12.3 billion; the second-quarter loss includes stock-based pay³. The CFO’s slides, as reported, showed the revenue run rate up 35% quarter to date and the enterprise run rate up 50%². Run rate means the latest month or quarter multiplied up to a full year. It is not booked revenue.

One more cost is paid to Microsoft. Under the April 2026 amendment, OpenAI keeps paying Microsoft a share of revenue through 2030 at the same percentage, now with a cap on the total, and Microsoft stopped paying a share to OpenAI⁵. The percentage and the cap were not disclosed in the post. One outlet reported a $38 billion cap, which I could not confirm.

Reported revenue and losses (USD)

IIIThe market and the competition

OpenAI sells into three markets: consumer assistants, business software and coding tools, and raw model access for developers. In each, it faces rivals with deep pockets. Anthropic, the maker of Claude, confidentially filed for its own IPO about a week before OpenAI and closed a round at a $965 billion valuation, above OpenAI’s $852 billion¹. Reported Anthropic run rates differ: $65 billion annualised at the end of July in one report², and about $47 billion in another¹⁰. They may use different accounting, and a data compilation notes that a $70 billion figure for OpenAI used Anthropic’s accounting method, so it is not comparable³.

Accounting basis matters here. Some AI labs count revenue gross, before paying cloud partners their share, and some count it net. Two run rates that look different may describe the same business. I have used the figures as reported and flagged where the basis is unclear. When a filing arrives, the first thing to check is how revenue is defined and how much of it passes straight to Microsoft under the revenue share⁵.

The market is also unusually concentrated in what it buys. Training and running large models needs advanced chips, power and data-centre space, and a few suppliers control them. That gives the suppliers leverage over every lab, OpenAI included, and it is why chip makers are willing to tie their own equity and warrants to OpenAI’s orders⁷⁸.

Google, Meta and Microsoft all build and sell AI too, and they fund it from large existing profits. Alphabet’s shares fell about 11% after its late-October report because AI spending pushed free cash flow negative again¹⁶. Meta raised its capex outlook to a range of $135 billion to $145 billion¹⁶. Those firms can lose money on AI for a long time without a funding problem. OpenAI cannot rely on that, and depends on investors continuing to pay.

SpaceX, which merged with xAI earlier in 2026, listed in June and is named alongside OpenAI, Anthropic and Google as a key AI competitor¹. I did not find a clean xAI revenue figure, so I cannot compare it.

Executive turnover is part of the picture. Reports in August noted that the revenue chief, a longtime executive and the product chief had all left in recent months². I cannot judge how much that matters.

Demand looks strong at the moment. Microsoft reported Azure revenue of $33.2 billion, up 48%, and a commercial backlog of $677 billion, up 73%⁶. That backlog is not all OpenAI, and the release does not say how much is. I mention it only to show that the cloud market OpenAI buys from is still growing quickly, which supports the idea that AI demand is real while telling us little about whether OpenAI keeps its share of it.

IVThe moat and how it could be disrupted

OpenAI has real advantages. ChatGPT is the best-known AI assistant, with about 1.2 billion weekly users³. Habit, brand and distribution matter in consumer software. Enterprise adoption adds switching costs, since companies build workflows and approvals around a tool. Its scale also attracts chip and cloud partners who want to supply it, as the Nvidia and AMD deals show⁷⁸.

But the evidence for a durable moat is thin. Model quality has moved between labs, and a rival with a better model can win business quickly, because developers can swap one API for another. Anthropic’s reported valuation is higher than OpenAI’s¹, and Friar told staff not to worry if Anthropic lists first². Coding is the most contested area: the CNBC report describes Codex as competing directly with Anthropic’s Claude Code¹.

The cloud tie to Microsoft has also loosened. Since April 2026 OpenAI may sell all its products on any cloud, and Microsoft’s licence to OpenAI’s models through 2032 is no longer exclusive⁵. That is good for OpenAI’s reach, and it also means the old structural link is weaker.

My rating is contestable. The advantages are real and the user base is large, but none of them is hard to challenge within a few years. A credible challenger already exists.

VThe numbers we can see

There is no public prospectus, so nothing here is audited and filed. What we have is the company’s own funding announcements as relayed by the press, a Microsoft earnings release, two leaked sets of figures, and reporting from the Financial Times, Bloomberg, the Wall Street Journal and others, which I saw only through secondary pages³¹⁰.

At $852 billion and a run rate near $50 billion, the price is about 17 times run-rate revenue (my arithmetic). Against the $36 billion of revenue expected for 2026 in a company presentation, it is about 24 times³. For comparison, Microsoft trades at about 11 times trailing revenue, with $353 billion of revenue and a 15% free cash flow margin¹⁵. Alphabet and Meta trade at 9.5 and 8.0 times sales¹⁷.

The cash picture is the hard part. A Financial Times estimate, as relayed, puts negative free cash flow at $278 billion over 2026 to 2030, with about $856 billion of compute and infrastructure spending³. Fortune reported in November 2025 that OpenAI had signed up to $1.4 trillion of compute commitments over eight years and expected to turn cash-flow positive in 2029 or 2030⁹. OpenAI also contracted to buy an additional $250 billion of Azure services⁴.

Microsoft’s own results offer one small window. Its press release excludes OpenAI investment effects from non-GAAP income. Net losses from OpenAI investments reduced net income by $3.1 billion in the quarter ended September 2025, while net gains raised it by $207 million in the quarter ended September 2026⁶. The release does not explain the swing, and I would not read much into it.

Three cautions on all these figures. First, nearly every number above comes from leaks or anonymous sources, and the company has not published audited accounts. Second, the figures are on different bases: booked revenue, run rate and company targets are not the same thing, and I keep them apart in the tables. Third, the valuation is a price set by a small group of large investors in a negotiated round, not a price discovered by thousands of buyers and sellers. The August tender at the same $852 billion price³ shows employees could sell at that level, but it was a company-run event and not an open market.

Valuation history

VIWho runs it and who owns it

Structure first. A nonprofit, the OpenAI Foundation, controls OpenAI Group PBC, a for-profit public benefit corporation¹³. A public benefit corporation is a company that must weigh a stated public purpose alongside profit. The recapitalization completed on 28 October 2025. Microsoft held about 27% on an as-converted, diluted basis, worth about $135 billion at the time, down from 32.5% before recent funding rounds⁴. I did not open a source giving the Foundation’s exact stake, and the stakes have been diluted by the 2026 round, so I give no percentage. Press coverage put the Foundation near a quarter at the recapitalization, which is unverified here.

Microsoft’s rights were reset twice. In October 2025 its IP rights ran through 2032 and covered post-AGI models, and an independent panel must verify any declaration of AGI (artificial general intelligence, a term with no agreed definition)⁴. In April 2026 the licence became non-exclusive and revenue-share flows changed as described above⁵.

The 2026 round was co-led by SoftBank. Amazon committed $50 billion, of which $15 billion was upfront and $35 billion conditional; Nvidia and SoftBank each committed $30 billion; other investors included Andreessen Horowitz, MGX, TPG, D. E. Shaw and accounts advised by T. Rowe Price. About $3 billion came from individual investors through banks, a first for the company¹¹. Microsoft also took part, in an undisclosed amount¹¹. Note that these are commitments, and part of Amazon’s is conditional. The same firms that invest are also suppliers or cloud partners, so some of the money may return to them as chip and cloud spending. That circularity does not make the round fake, but it makes the headline figure a weaker signal of independent demand than a round led only by financial investors would be.

Control deserves a plain statement. Because the Foundation controls the PBC¹³, ordinary shareholders, including future public shareholders, do not decide strategy in the usual way. I have not read the governing documents, so I cannot say how far these limits go.

Sam Altman is chief executive and Sarah Friar is chief financial officer². Goldman Sachs and Morgan Stanley are working on the IPO filing¹.

On what an outside owner can take out: OpenAI has paid no dividends that I found, it loses money, and the nonprofit holds control. Returns to a new shareholder would come only from selling shares at a higher price, whether in a tender, a secondary market or an IPO. The August tender let employees sell at $852 billion³.

VIIValuation: is the last price supported?

Equity value in US$ billions. Not a share price: a private company has none. Latest mark: primary round, post-money (also used for the August 2026 employee tender).

What the last valuation assumes

With my other inputs held fixed (8x terminal multiple, 12% discount rate, 26% dilution), $852bn needs about $230bn of revenue in 2030, against $200bn in my base. It also assumes new investors keep funding the gap between revenue and the planned compute spending, so dilution stays near my base. The pending round at a reported $1.25tn pre-money target would need about $340bn of 2030 revenue on the same method, close to the company’s own $350bn target³.

Bear, base and bull

• Bear. 2030 revenue of $110bn, about 2.2 times the 2026 expectation of $36bn³. Terminal multiple of 5 times revenue, below Meta’s 8.0 times sales¹⁷, because growth has slowed and margins stay thin. Discount rate 12%, 4.15 years from November 2026 to end-2030. Dilution of 30% from further funding. Equity value about $241bn, shown as $240bn.
• Base. 2030 revenue of $200bn, in line with a 2025 company projection reported by Fortune⁹ and well below the $350bn target reported by the FT³. Terminal multiple of 8 times revenue, between Meta at 8.0 times and Alphabet at 9.5 times sales¹⁷, and below Microsoft at 11.0 times¹⁵. Discount rate 12%, a rough cost of equity for a risky, unlisted growth company; it is my choice, not a market figure. Dilution of 26% to fund the cash burn estimated at $278bn over 2026 to 2030³. Equity value about $740bn.
• Bull. 2030 revenue of $350bn, the company target reported by the FT³. Terminal multiple of 10 times, near Microsoft’s 11.0 times¹⁵, because margins reach those of a mature platform. Discount rate 12%. Dilution of 20%. Equity value about $1,749bn, shown as $1,750bn.

Comparables used

Where I come down

My base of about $740bn is below the last round of $852bn, by roughly 13%. I call the round above my base but not far from it, and it is seven months old, so it is stale: revenue has grown since, and the company is now seeking a higher price. To believe $852bn, you would have to believe revenue reaches about $230bn by 2030, that margins eventually look like Microsoft’s, and that dilution stays below a quarter. The market could be right even if this looks high. Revenue grew from $3.7bn to $13.07bn in a year¹⁰, the run rate is reported near $50bn³, and the largest investors, such as Amazon, Nvidia and SoftBank, are also suppliers or customers, which gives them reasons beyond the return on equity¹¹. The margin of safety figure of $555bn sits well below the last round. The 25% haircut is a rule of mine, not a market signal. All of this is research, not investment advice.

What I cannot judge

I cannot judge future gross margins, because there is no audited disclosure of the cost of serving customers. I cannot judge how firm the compute commitments are, since some were letters of intent⁷. I have not seen the S-1, the cap table, or the terms of the newest preferred shares, so liquidation preferences and the real share count are unknown. The leaked financials and run-rate figures are press-reported and not checked by me against a filing. Different outlets disagree on the date the IPO was filed (May or June 2026). The model is sensitive: each 1x change in the terminal multiple moves the base by about 12%, so treat the range, not the point, as the finding.

VIIIThe IPO

IPO filed confidentially. OpenAI said on 8 June 2026 that it had confidentially submitted a draft registration statement (S-1) to the SEC. It said it expected the news to leak and had not decided on timing: ’it may be a while’¹. Goldman Sachs and Morgan Stanley are working on the filing; no exchange or ticker has been named¹. Earlier reporting suggested a debut as soon as the fourth quarter of 2026¹. In August the CFO told staff she expects OpenAI to be public in 2027, or sooner if the business keeps inflecting². A data compilation says Altman ruled out a 2026 listing, which I could not verify from a primary source³. In October the FT reported a person close to the company expects an IPO in the first half of 2027, with a new private round meant as a bridge¹². No public S-1 is on EDGAR as far as I could see. I could not confirm an IPO size, price range or share count.

What to watch

• A public S-1 filing, which must appear at least 15 days before a roadshow, with audited revenue, margins, compute obligations and the cap table
• Whether the pending private round closes, at what price, and how large it ends up (reported target $1.25tn pre-money)¹²
• How Anthropic prices its own IPO, which investors said they wanted to see first¹²
• Whether the Musk appeal is filed and heard in the Ninth Circuit¹⁴
• Microsoft’s disclosures on its OpenAI stake and any change to the capped revenue share⁵⁶
• Delivery of the first gigawatt from Nvidia and AMD in the second half of 2026⁷⁸

IXWhat could go wrong

The first risk is funding. The plan needs hundreds of billions of dollars more than revenue provides, and private investors may balk. In October 2026 the Financial Times reported OpenAI cut its target price for a new round to $1.25 trillion pre-money from $1.4 trillion after resistance from wealth clients. Demand was described as weaker than in past rounds, and some investors wanted to see how Anthropic prices first¹². A later headline on the same page mentioned a cut to about $20 billion; I could not confirm it. Tech shares also fell in October.

The second risk is that revenue growth slows while commitments stay fixed. If compute contracts are fixed in dollars and revenue underperforms, the company has to renegotiate, raise more or cut spending. Commitments reported in the press, such as $1.4 trillion over eight years, are not all firm; the Nvidia agreement was a letter of intent⁷⁹.

The third risk is legal. A jury found in May 2026 that Elon Musk’s claims against OpenAI, which sought up to $134 billion and the removal of Altman and Brockman, were filed too late; the verdict form addressed only timing. Musk said he would appeal to the Ninth Circuit¹⁴. I could not check whether an appeal has been filed. An appeal that succeeded might revive claims about the structure, though the judge’s comments suggested a hard path.

The fourth risk is competition. If rivals match OpenAI’s models at lower cost, prices fall and the revenue base is harder to defend. The fifth is dependence on partners. The Microsoft deal gives it a revenue share through 2030⁵, and chip suppliers hold side agreements such as AMD’s warrant⁸, which add complexity.

What most likely destroys the business is not a single event. It is a funding gap: revenue growing at a fast but not fast enough pace, while costs and commitments run ahead, and capital markets closing at the wrong time. To make the base case work, several things have to hold together: revenue reaching around $200 billion by 2030, funding staying available, margins improving toward large software peers, and no rival taking a large share. None of this means the business fails. It means the price leaves little room for ordinary setbacks, and the range of outcomes is very wide. Treating those as roughly 40%, 80%, 50% and 70% (my rough judgement, not data) gives about 11%, or around one in nine. The inputs overlap, so the true figure is likely higher, but it shows how much has to go right together.

For the same dollar, the next best alternative is a listed large technology company with positive free cash flow. Microsoft has a 15% free cash flow margin and trades near 11 times revenue¹⁵. Such shares carry their own risks, but their numbers are audited and filed. This is a comparison, not advice.