How to Value AI Companies in 2026: OpenAI & Anthropic

Ask a roomful of investors what OpenAI or Anthropic is worth and you will get a roomful of different answers, all defensible, none provable. As of mid-2026 OpenAI has been priced at roughly $852 billion in its latest private round and Anthropic at around $965 billion, yet both are losing billions of dollars a year, neither has a stable stream of free cash flow, and neither looks like anything a traditional valuation model was built to handle. That is exactly why "how would you value a company like OpenAI?" has become one of the most revealing questions an interviewer can ask. There is no clean answer, so what you are really being tested on is judgment: which tools apply, where each one breaks, and how honest you are about the uncertainty.

This post walks through how investors and bankers actually approach frontier AI labs. We cover why the standard playbook fails, the handful of methods people fall back on, the numbers that get quoted carelessly and what they really measure, the strange corporate structures that change what "owning equity" even means, and the controversies (circular financing, GPU depreciation, and the math that says they cannot all win) that sit underneath every headline number. By the end you should be able to hold an intelligent conversation about it rather than reaching for a single multiple and hoping.

A frontier AI lab here means a company building large general-purpose models at the technology frontier, funded privately and burning enormous amounts of cash to train the next model. OpenAI and Anthropic are the cleanest examples, and because both have now started moving toward the public markets, the question of what they are genuinely worth has stopped being academic.

The instinct of any trained analyst is to reach for a discounted cash flow analysis or a set of trading multiples. For a frontier lab, both strain badly, and understanding why is the first half of a good answer.

There is no free cash flow to discount

A DCF values a business off the cash it throws off after reinvestment. Frontier labs do the opposite: they consume cash at a staggering rate to train ever-larger models and to rent the compute that runs them. OpenAI is reported to be on track to lose around $14 billion in 2026, with cash burn estimated near $27 billion for the year, and its own internal projections do not show cash-flow breakeven before roughly 2029. You cannot discount free cash flow that is deeply negative for years and entirely dependent on a revenue ramp nobody can confirm. Every input to the model (the growth rate, the eventual margin, the discount rate, the terminal value) becomes a guess, and a DCF built on stacked guesses tells you more about the analyst's assumptions than about the company.

There are no clean comparables

Multiples work when you can find similar companies trading on observable metrics. Frontier labs have no true peers. They are not software companies in the ordinary sense, because their cost of goods sold scales with usage in a way that pure software does not. They are not hardware companies, though they live and die by access to chips. And the two obvious comparables, OpenAI and Anthropic, are themselves private and unprofitable, so comparing one to the other just moves the uncertainty around rather than resolving it.

The numbers themselves are slippery

Even the basic figures are treacherous. Revenue is usually quoted as an annualized run-rate, not realized annual revenue. Valuations come from private rounds with terms attached, not from a liquid market price. And the gap between a company's headline valuation and what a share of common stock is actually worth can be enormous once you account for the rights senior investors negotiate. We will pull each of these apart later, because mislabeling them is the fastest way to look amateur.

With the textbook tools weakened, valuation of AI labs falls back on a handful of approaches, each capturing something real and each incomplete. The table below is the mental map; the sections after it explain how to use each one and where it fails.

Revenue and run-rate multiples

The most common shorthand is a multiple of revenue, because there is no positive EBITDA or net income to multiply. Across AI startups, revenue multiples have run anywhere from roughly 10x to 50x, with a median often quoted around 20x to 30x, against something closer to 6x for traditional software. Frontier labs sit at the rich end of that range because investors are paying for the growth rate and the possibility of dominance, not for current economics.

The trap is the denominator. When a lab is reported at a given multiple, that figure is almost always built on annualized run-rate revenue (the most recent month's revenue multiplied by twelve), not on what the company actually booked over the trailing year. For a company doubling every few months, run-rate can be two or three times realized revenue, which quietly makes the multiple look far more reasonable than it is.

Annualized Run-Rate Revenue

A forward projection that takes a recent period's revenue, often a single month, and multiplies it out to a full year as if that pace held steady. It is useful for fast-growing companies but flattering: it assumes momentum continues and can sit far above the revenue a company has genuinely earned over the past twelve months. Always check whether a quoted figure is run-rate or realized.

The research firm Sacra, which tracks private-company revenue, estimated Anthropic's run-rate climbing from roughly $9 billion at the end of 2025 toward something approaching $50 billion by mid-2026, while OpenAI's sat around $25 billion in early 2026 (about $2 billion a month). Treat those as run-rate snapshots of a moving target, not audited annual results. The direction is real and remarkable; the precision is not.

The venture capital method

When you cannot discount cash flows, you can work backwards from a future exit. This is the venture capital method, and it is how a lot of late-stage AI pricing is rationalized after the fact. You estimate what the company could be worth at a future exit, then discount that back at a very high required return to reflect the risk of getting there.

Here $r$ is the investor's required annual return, typically 30% to 60% for high-risk venture positions, and $t$ is the years to exit. The exit value itself usually comes from projecting a future revenue figure and applying an exit multiple, which ties this method straight back to the multiples problem above.

The honest use of this method is not to produce a number but to reverse it: take the price investors are actually paying, and solve for what they must believe about future revenue and exit multiples. When you do that for the largest labs, the implied beliefs are aggressive, which is the whole controversy in one calculation.

Master the technical interview: Practice 1,000+ valuation, modeling, and behavioral questions with worked answers, download our iOS app and walk into your superday ready for the curveballs.

What a funding round actually tells you

It is tempting to treat the latest round as the answer: OpenAI raised about $122 billion at a roughly $852 billion post-money valuation, and Anthropic raised about $65 billion at a roughly $965 billion valuation. Those are real arm's-length prices set by sophisticated investors, which is genuinely informative. But a private round price is not a market price, and the post-money headline is not what a share of common stock is worth, for reasons we get to in the next section. A round tells you what one set of investors paid for a specific class of shares with specific protections on a specific date. It is a data point, not a verdict.

Strategic-stake comparables

A distinctive feature of this market is that the biggest investors are also the biggest suppliers and customers. Microsoft's stake in OpenAI and the large investments Amazon and Google have made in Anthropic put a price on these companies, but a strategic one. A cloud provider investing in a lab that will spend the money on that provider's compute is buying more than financial upside, so the implied valuation embeds strategic value you cannot extract cleanly. It is a comparable, but a distorted one.

Real options and sum-of-the-parts

The most generous framing treats a frontier lab as a portfolio of bets: a current product business, plus an option on transformative future products, plus a platform other companies build on. This real-options lens explains why investors pay prices that current revenue cannot justify, because they are partly buying the right to participate if one of these bets becomes enormous. The danger is that optionality can be used to justify almost any number. Used carefully it is a real insight; used loosely it is a rationalization.

Cost and unit economics

Underneath all of it sits a simple question: can the core service make money? For an AI lab the key metric is the gross margin on inference, the cost of actually running the model to answer a query.

Reported figures suggest Anthropic has run gross margins around 40% against OpenAI's roughly 33%, with inference margins improving sharply as models get cheaper to run. The cost of achieving a given level of performance has collapsed, by some estimates over 98% in two years, which is the bull case for margins. The bear case, and reporting by The Information has flagged that inference costs can run well above plan, is that heavy enterprise users consume far more compute, so serving the most valuable customers is also the most expensive, keeping margins structurally below the software businesses investors like to compare them to.

Half of looking competent on this topic is using the right number for the right thing. Three distinctions matter most, and they map directly onto the gap between enterprise value and equity value that every banker is expected to know cold.

Run-rate versus realized revenue

We covered this above, but it bears repeating as a discipline: whenever a revenue figure appears, ask whether it is annualized run-rate or trailing realized revenue. For a company growing this fast the two can differ by a factor of two or more, and every multiple and every margin calculation depends on which one you used.

Post-money valuation versus common equity

This is the one most people miss. A "$965 billion valuation" is a post-money figure: the price per preferred share in the latest round, multiplied across all shares as if every share were identical. They are not. Late-stage investors typically buy preferred stock carrying a liquidation preference, the right to get their money back (sometimes a multiple of it) before common shareholders see anything in a sale or wind-down.

Liquidation Preference

A right attached to preferred stock that pays those investors back first, ahead of common shareholders, if the company is sold or liquidated. A 1x preference returns the original investment before anyone else is paid; higher multiples return more. Because of these rights, the headline post-money valuation can substantially overstate what a share of common stock, including employee equity, is actually worth.

The practical consequence: founders and employees holding common stock are worth less per share than the headline implies, and the steeper the preferences, the bigger the gap. Investors and auditors handle this with allocation models such as the option pricing model rather than assuming every share is equal. So when someone says a lab is "worth a trillion dollars," the honest follow-up is: worth that to whom, and which share class?

Private-round valuation versus IPO valuation

A private round valuation and an IPO valuation are different animals, and conflating them is a classic error. Both OpenAI and Anthropic have moved toward public markets: Anthropic confidentially filed for an IPO in mid-2026, and OpenAI confirmed it had confidentially submitted a draft registration statement to the SEC, while cautioning that an actual listing "may be a while." A confidential filing carries no offering price, no deal size, and no public valuation. Until a company prices a deal, the only valuation you can legitimately cite is its last private round, and you must call it exactly that. As Fortune reported on the Anthropic filing, the $965 billion figure is a private-round valuation, not an offering price, and treating it as an IPO valuation would be wrong.

For the contrast, look at the other mega-listing of 2026. SpaceX is not an AI lab, but it became the cleanest live test of how the public market prices AI exposure. Having acquired xAI, Elon Musk's AI company, in an all-stock deal earlier in the year, SpaceX went public in June at roughly a $1.75 trillion valuation, raising about $75 billion in the largest IPO on record. That single price bundles Starlink's real operating profits with xAI's heavy losses (reported on the order of $2.5 billion a quarter), so investors buying the stock were forced to put one number on a profitable business and a cash-burning AI business at once. The crucial difference from the labs: SpaceX filed a public S-1 with the SEC and printed an actual offering price near $135 a share. That is precisely what a confidential filing does not give you, and it is why an IPO price and a private-round mark should never be quoted as if they were the same kind of number.

Two of the labs people most want to value have governance structures that complicate the very idea of owning a piece of them. This is not trivia; it directly affects who controls the company and what economic rights a share carries.

OpenAI: a nonprofit foundation over a public benefit corporation

OpenAI began as a nonprofit, then created a "capped-profit" subsidiary in 2019 that let investors earn up to a fixed multiple on their money. In late 2025 it restructured again: the nonprofit became the OpenAI Foundation, and the for-profit became OpenAI Group, a public benefit corporation. According to OpenAI's own description of its structure, the Foundation controls the PBC, including the power to appoint and remove directors, while holding a reported 26% economic stake; Microsoft holds around 27%, with the remainder spread across other investors and employees. The move off the capped-profit model toward ordinary stock made the company easier to value, but the nonprofit's control means shareholders do not have the unfettered authority they would in a normal corporation.

Public Benefit Corporation

A for-profit company legally required to pursue a stated public mission alongside shareholder returns, with directors obligated to weigh both. Both OpenAI's operating entity and Anthropic are public benefit corporations, which means a buyer of the equity is buying into a company that is bound by charter to balance profit against a broader purpose.

Anthropic: a PBC governed by a Long-Term Benefit Trust

Anthropic is also a public benefit corporation, with an additional layer: a Long-Term Benefit Trust made up of independent members who hold a special class of shares and, over time, gain the right to elect a growing share of the board, eventually a majority. Anthropic also uses a multi-class share structure that concentrates voting power with founders and the Trust even as outside investors hold most of the economics. For a valuation, this means control and economics are deliberately separated: you can own a large fraction of the company's value and still have little say over its direction.

You cannot value these companies honestly without engaging the debate about whether the whole sector is priced sensibly. Three threads matter most.

Circular financing

A growing share of AI funding flows in loops: chip and cloud vendors invest in the labs, and the labs spend that capital buying the vendors' chips and cloud capacity. Nvidia has said it would invest up to $100 billion in OpenAI, which plans to fill its data centers with Nvidia chips; AMD struck deals reportedly worth around $200 billion that handed customers equity warrants; and Oracle committed roughly $300 billion in cloud infrastructure. Bloomberg mapped the web of deals and put the total of such circular arrangements above $800 billion. Critics note the resemblance to the vendor-financing that flattered the late-1990s telecom bubble: revenue and investment that partly circle back to the same players can inflate apparent demand and magnify losses if real demand disappoints.

GPU depreciation and "useful lives"

How long a GPU lasts on the books quietly drives reported profits across the whole AI supply chain. Investor Michael Burry argued in late 2025 that hyperscalers depreciate AI chips over five or six years when their real economic life may be closer to two or three, which would understate depreciation and overstate profits by an estimated $176 billion across the industry between 2026 and 2028. Nvidia and large cloud operators pushed back, citing four-to-six-year usage and strong resale values for older chips. The dispute matters for valuation because it goes to whether the reported economics of the entire ecosystem, including the labs' largest suppliers and customers, are as healthy as they look.

The math that says they cannot all win

The sharpest critique is an aggregate one. NYU's Aswath Damodaran, often called the dean of valuation, has argued that if you add up the revenue every AI player would need to justify its current price, the total is implausibly large relative to any realistic size of the market. Investors are pricing the leading labs as though several of them will each become dominant and enormously profitable, yet by definition they cannot all capture the same revenue. Some of today's valuations are therefore pricing in outcomes that are mutually exclusive across the group. That does not mean any single lab is overvalued; it means the sector's combined valuation embeds a level of collective success the arithmetic struggles to support.

Sequoia Capital's David Cahn made the same point from the spending side in his widely cited "AI's $600 billion question." His method is deliberately simple: take the industry's chip and data-center spending, gross it up for the total cost of running that infrastructure and the margin operators need, and you arrive at the annual revenue AI must eventually generate to be rational. Through 2026 that implied gap widened rather than closed, with aggregate AI capex running ahead of both the revenue and, increasingly, the free cash flow available to fund it. Whether you come at it from Damodaran's revenue ceiling or Cahn's spending floor, the conclusion rhymes: the sector is priced for a market that does not yet exist.

If you are asked to value an AI lab, the interviewer does not expect a number. They expect a structured way of thinking, delivered calmly.

Start by naming why the standard tools fail: no positive free cash flow for a DCF, no clean comparables for multiples. Then lay out the menu: revenue and run-rate multiples for a market-based read, the venture capital method to back into implied returns, the latest funding round as a recent arm's-length data point, and unit economics to test whether the business can ever earn money. Flag the number-traps explicitly, run-rate versus realized revenue, post-money versus common equity, and private round versus IPO price, because showing you know what each figure measures is most of the battle. Close with the controversies and a view: is the price defensible given burn, circular financing, and the aggregate-revenue problem? Land somewhere, and justify it.

This is also why the topic connects to bigger questions about the industry, including whether AI changes the analyst job itself. The labs you are being asked to value are the same ones reshaping the work you are interviewing for.

Go deeper on the fundamentals: Get the complete frameworks behind every valuation method in one place, access the IB Interview Guide and learn how multiples, DCFs, and deal math fit together.

Valuing OpenAI and Anthropic is hard for an honest reason: the companies are growing faster than any framework was designed to measure, while burning the kind of cash that makes those frameworks unreliable. The right response is not to pick one number and defend it, but to triangulate across several methods, label every figure for what it actually measures, account for the structural quirks that change what a share is worth, and stay clear-eyed about the controversies that sit under the headlines. That is also what makes it such a good interview question. It rewards the candidate who can wield the full valuation toolkit and is honest about where each tool runs out.

If you can explain why a DCF fails here, why a run-rate multiple flatters, why a post-money headline overstates common equity, and why the sector's combined valuation strains the arithmetic, you are not just ready to talk about AI labs. You are demonstrating the exact judgment that separates a strong analyst from someone who has only memorized formulas. Build that habit of asking what a number really measures, and it will serve you on every deal you ever touch.