OpenAI and Anthropic IPOs: the day artificial intelligence has to open its books

OpenAI and Anthropic have announced a combined $187 billion in funding within the space of a few months, while simultaneously preparing to go public. Their IPO prospectuses will reveal what private valuations and annualised revenue still allow them to conceal: the true cost of compute, the quality of enterprise revenue, the commitments made to suppliers and the margin generated by each unit of intelligence sold. Between ChatGPT’s mass distribution and Claude’s professional specialisation, Wall Street will not merely be choosing between two laboratories. It will have to settle a much broader question: does frontier AI belong to the economics of software, to those of industrial infrastructure, or to a model condemned to raise ever more capital simply to keep growing?

In late August, during an all-hands meeting, Sarah Friar summed up OpenAI’s new relationship with funding in a single sentence. According to CNBC, the chief financial officer said the company could go public as early as 2027, or even sooner if the business continues to accelerate, and that an IPO would not be a finish line but another step: “another fundraise.”

In March, OpenAI completed a funding round involving $122 billion in committed capital, taking its post-money valuation to $852 billion. Two months later, Anthropic raised $65 billion and reached a valuation of $965 billion. In just a few weeks, the two laboratories announced more capital between them than the entire US venture capital industry deployed in some full years.

Yet these sums do not close the funding cycle; they change its scale. If OpenAI and Anthropic are preparing to go public, it is because their cash requirements may soon exceed what private markets can sustainably absorb.

The models are leaving the benchmarks and entering the accounts

Since the launch of ChatGPT, the rivalry between AI laboratories has largely been told through technical performance. Every new model promises to write, code, reason, use tools or interpret images better than the one released a few weeks earlier. The gaps are measured on benchmarks whose subtleties fascinate engineers and whose rankings provide everyone else with a steady supply of instantly shareable screenshots.

Going public imposes a less flattering exercise. OpenAI and Anthropic will have to disclose recognised revenue, expenses, cash flows, stock-based compensation, compute commitments, customer concentration and their financial relationships with their principal investors and suppliers.

Both companies confidentially filed draft prospectuses with the SEC in June. Anthropic could disclose its filing first, with a listing reportedly contemplated as early as September, while OpenAI says it could wait until 2027, even as it retains the option of moving sooner.

As long as the documents remain confidential, the laboratories can still choose which metrics to release—and naturally select those that best reflect their current velocity: user numbers, monthly growth, enterprise customers and, above all, revenue run rate.

A run rate takes revenue from a short period, generally the latest month, and assumes that it will repeat twelve times in succession. OpenAI is thus reporting an annualised pace of more than $40 billion; Anthropic reportedly exceeded $65 billion at the end of July, up from $47 billion in May and around $9 billion at the end of 2025.

If accurate, these figures demonstrate indisputable commercial acceleration. They do not, however, represent revenue already earned over a full year—a distinction made clear by the second-quarter results.

OpenAI reportedly generated $6.7 billion between April and June, equivalent to $26.8 billion if the quarter is simply multiplied by four. Anthropic reportedly recorded around $11.6 billion, for an annualised pace of $46.4 billion. The respective $40 billion and $65 billion run rates therefore assume that the growth observed since June continues uninterrupted. That is an assumption, not an accomplished fact.

At the valuation of its latest round, OpenAI trades at around 21 times its current run rate, but almost 32 times its annualised second-quarter revenue. Anthropic, meanwhile, is valued at roughly 15 times its July run rate and 21 times its annualised quarterly revenue. The markets will therefore not merely be asked to pay for current revenue: they will have to buy in advance some of the revenue that both laboratories still hope to generate.

Anthropic is reportedly already projecting between $190 billion and $200 billion in revenue for 2028. Some banks are said to be using this forecast to suggest a valuation of up to $2 trillion—a method that effectively asks investors to cross two years of exceptional growth before they even begin applying their multiple. The practice is not unknown on Wall Street. What is more unusual is the number of years that must be treated here as almost guaranteed.

Anthropic overtook OpenAI where it seemed least likely

The reversal in the second quarter gives this confrontation significance far beyond the timing of the two IPOs.

According to data shared with investors and reported by The Wall Street Journal, OpenAI’s revenue rose by 18%, from $5.7 billion in the first quarter to $6.7 billion in the second. Over the same period, Anthropic reportedly more than doubled its revenue to around $11.6 billion. The laboratory founded in 2021 by former OpenAI employees would thus have overtaken, for the first time, the company from which it emerged.

The divergence is even sharper on earnings. OpenAI’s operating loss, including stock-based compensation, reportedly increased from $9.3 billion to $12.3 billion: revenue rose by 18%, while losses grew by around 32%. In other words, OpenAI lost almost $1.84 for every dollar of revenue recognised during the quarter. Anthropic, meanwhile, reportedly posted a small adjusted profit, attributed to more efficient use of its compute capacity.

The comparison should nevertheless be treated with caution. Anthropic’s precise accounting method is not public, and its previous disclosures excluded certain stock-based compensation expenses: an adjusted profit on one side cannot be directly compared with an operating loss that includes them on the other. Publication of the prospectuses should finally clear some of this accounting fog.

For now, the signal is strong enough: the laboratory with the largest audience, the best-known brand and the broadest product range no longer necessarily has the strongest commercial momentum.

OpenAI is turning its audience into an enterprise distribution channel

OpenAI nevertheless retains an advantage that its competitor cannot quickly replicate: ChatGPT’s distribution. According to the company’s own figures, it is approaching one billion weekly users—an audience that serves simultaneously as a consumer product, an acquisition channel and a global training programme for its interfaces.

When a company rolls out ChatGPT, some of its employees already know how to use it. OpenAI can therefore enter organisations through individual usage before extending its presence to teams, developers, business units and information systems.

The strategy is beginning to pay off. On 14 August, Sarah Friar said that enterprise revenue had surpassed consumer revenue several months ahead of internal forecasts. OpenAI had started the year with a 60/40 split between consumer and enterprise revenue; its professional business is now reportedly growing faster than individual subscriptions.

The company is no longer merely selling access to a model. It is seeking to become a general layer for digital work: ChatGPT as the interface, ChatGPT Work for professional tasks, Codex for software development, APIs for applications, agents for workflow execution, and advertising and commerce to monetise its consumer audience.

As early as November 2025, OpenAI said it had more than one million enterprise customers, a category encompassing both organisations purchasing ChatGPT seats and those consuming models directly through the API. Its “enterprise” revenue does not therefore necessarily have the characteristics of a conventional SaaS subscription: some of it depends on usage volumes, token prices and customers’ ability to switch to a competing model.

The 20 million weekly active users now claimed for its coding and work products follow the same logic: it is an adoption metric that says nothing about the number of paying users, the amounts spent or customer loyalty once pricing changes.

If the size of OpenAI’s audience is an asset, it is also a bill. The company must serve hundreds of millions of free or lightly monetised users, maintain several product families and simultaneously fund research, coding, voice, image, video, agents and future devices. It still has to demonstrate that ChatGPT is an efficient marketing engine, rather than merely the most expensive loss leader ever built.

Anthropic chose code before pursuing the mass market

Anthropic, for its part, is following a more focused trajectory, built around APIs, professional organisations, sectors where reliability is particularly important and, above all, Claude Code, which allows developers to delegate a growing share of software production, review and testing.

Code has the advantage of offering readily perceptible economic value: a company can compare the cost of Claude Code with an engineer’s time, a project’s duration, the number of features produced or the errors avoided.

This specialisation has allowed Anthropic to convert its technical capabilities into professional spending more quickly, while avoiding the cost of acquiring a consumer audience it does not possess. Yet Anthropic remains heavily dependent on growth in coding usage, a small number of cloud channels and customers who continually compare its performance with that of Codex, Gemini, GitHub Copilot, Cursor and open models.

Its enterprise position also relies heavily on Amazon and Google. Amazon has invested $8 billion in the company, while AWS is its primary cloud and training partner; Anthropic also helps optimise the Trainium accelerators developed by the group. This relationship spares it from having to build its entire infrastructure alone, but it also places part of the distribution, compute and potentially the margin with its main partner.

On one side, OpenAI is building a horizontal platform and trying to bring work into it. On the other, Anthropic starts with work and gradually expands its platform.

Every new user still arrives with their own bill

The software industry was built on a simple promise: initial development is expensive, but reproducing the product costs almost nothing. Generative AI disrupts this equation. A new query requires compute; a longer conversation requires more; an agent that consults files, uses several applications, writes code, tests the result and then starts again can consume resources for several minutes, or even several hours. Every additional user therefore increases revenue and costs at the same time.

Technical efficiency gains should improve the equation: more powerful chips at constant energy consumption, smaller models, caching, quantisation, distillation and routing all make it possible to reserve the most expensive systems for genuinely difficult tasks. Yet there is no guarantee that model producers will retain those gains.

Competition continually pushes prices down. Companies are becoming more attentive to the cost of each task and moving less demanding workloads to cheaper models; Chinese and open-weight systems introduce an additional price benchmark. If the technical cost of a query is halved but its commercial price falls by two-thirds, the customer captures the benefit first.

The true competitive advantage will therefore lie not only in having the most intelligent model, but in being able to complete an entire task at a sufficiently low cost while preventing the customer from immediately switching the workload to another provider. This is precisely where the difference between software economics and commodity economics lies: in the former, scale gradually improves margins; in the latter, every efficiency gain is immediately returned to the market through lower prices.

Funding rounds now finance digital factories

The scale of the capital raised by OpenAI and Anthropic can no longer be explained solely by the cost of research. The laboratories must secure chips, networks, electricity and data centres several years before they know precisely how much demand will use them.

OpenAI now presents durable access to compute as one of its main strategic advantages. The Stargate project calls for up to $500 billion in investment over four years in US infrastructure—a sum that is not carried by OpenAI’s balance sheet alone, since the project also involves SoftBank, Oracle, MGX, Microsoft and Nvidia. The company nevertheless remains the commercial operator around which this capacity is being built: it must fill the data centres, honour its contracts and generate enough revenue for its partners to continue financing the next generation. The quality of its growth therefore matters not only to its shareholders, but also to suppliers whose own investments depend on this future demand.

Anthropic currently outsources more of this infrastructure to AWS and Google Cloud, a choice that reduces the investments recorded on its balance sheet but may also shift part of its margin to the cloud providers and create contractual commitments.

In both cases, the boundaries between investors, suppliers and distributors are becoming porous. Amazon, Nvidia, SoftBank and Microsoft participated in OpenAI’s latest round; Amazon and Google are simultaneously Anthropic shareholders, technology partners and distribution channels. Some of the capital invested in the laboratories can therefore flow back to the same groups through purchases of chips, cloud services or data-centre capacity.

This circulation is not enough to prove that demand is artificial, but it does require a more precise question: what proportion of revenue comes from independent customers, and how much of the growth depends on contracts concluded within the circle of strategic investors? The prospectuses will need to disclose capacity commitments, take-or-pay contracts, guarantees, minimum-purchase clauses and investments kept off the balance sheet. The decisive issue is not merely the cost of the infrastructure, but who will continue paying if demand falls short of forecasts.

Future shareholders will not necessarily receive power commensurate with their capital

OpenAI and Anthropic share another peculiarity: neither is governed like an ordinary technology company.

OpenAI Group is a public benefit corporation controlled by the OpenAI Foundation. Following the recapitalisation completed in October 2025, the foundation held 26% of the equity and Microsoft around 27%. The March 2026 funding round has probably changed those proportions, but no fully updated capitalisation table has been published. More importantly, the foundation retains the power to appoint and replace OpenAI Group’s directors: future public shareholders could therefore acquire an economic interest without gaining proportionate control over the company.

Anthropic is incorporated under the same legal form. Its Long-Term Benefit Trust can appoint some members of its board; the trust’s members hold no shares, do not participate in profits and are required to safeguard the laboratory’s long-term mission. Former Federal Reserve chair Ben Bernanke also joined the structure in July.

While these mechanisms may protect the companies from the pressure of quarterly results and allow them to delay a product deemed dangerous, they may also prompt investors to apply a discount because of limited control rights or decisions made in the name of interests broader than shareholder returns alone.

OpenAI must also demonstrate that it has the organisation required of a public company. The recent departures of Denise Dresser, Brad Lightcap and Fidji Simo do not by themselves prove internal chaos, but they come as the company must formalise its commercial leadership, financial controls and the allocation of responsibilities between Sam Altman, Greg Brockman and Sarah Friar. The promise of a 2027 IPO also serves an internal purpose: it gives employees holding shares a liquidity horizon—and one more reason to wait.

Anthropic could set the price of its former parent

When Sarah Friar says that OpenAI need not worry if Anthropic goes public first because the two companies are each running their own race, the line may resonate internally. It is likely to sound less convincing to investors, who will approach the same institutions, compare the companies with the same listed technology stocks and need to be persuaded that the falling price of intelligence will not prevent margins from improving.

If Anthropic lists in September, its prospectus will become the first public benchmark for the economics of a laboratory developing frontier models: the market will finally be able to examine the cost of compute, gross margin, customer concentration, cloud commitments and the true level of stock-based compensation.

A successful listing would create a new public-market category and probably make OpenAI’s own IPO easier. Anthropic would benefit from a scarcity premium, while its competitor could emphasise its broader distribution and the recent acceleration in enterprise revenue.

A difficult listing would have the opposite effect. If the accounts reveal weaker margins, excessive commercial dependence or heavier-than-expected compute commitments, OpenAI could postpone its transaction or be forced to accept a lower multiple.

The first IPO may earn the privilege of setting the criteria. It must also bear the cost of disclosure.

Three different economies could emerge.

The first would be that of software: revenue would grow faster than costs, inference would gradually become cheaper and margins would converge towards those of the largest technology platforms.

The second would be that of infrastructure: vast revenue, but also a relentless obligation to build new capacity, carry long-term contracts and share value with suppliers of chips, cloud services and energy.

The third would be that of permanent financing, in which every increase in revenue would require an even greater increase in committed resources, turning successive funding rounds into a condition of growth rather than a simple accelerator.

Sarah Friar may already have indicated which of the three OpenAI is preparing to embrace: the one in which an IPO is not the finish line, but another fundraise.

For four years, the laboratories have measured the intelligence of their models. This time, once the compute bill has been paid, they will have to measure what remains.