Anthropic and OpenAI Test the Rules for Valuing AI Revenue

The claim that Anthropic and OpenAI already generate more than $1 trillion in annual recurring revenue does not hold up against reported figures. The two companies appear to be running at roughly $78 billion to just over $100 billion in annualized revenue in mid-2026, with combined private-market valuations a little above $1 trillion.

That distinction matters. It separates a market repricing software around a new revenue model from a market that has detached from revenue reality. The trillion-dollar figure belongs to valuation, not ARR.

The real story is not trillion-dollar ARR today. It is that two foundation-model companies have reached a revenue scale that would have looked implausible for private software businesses a few years ago, and have done so on economics that do not map neatly onto classic SaaS.

Valuation has outrun revenue, but revenue has outrun precedent

Reported figures for Anthropic's run rate range from $30 billion in early 2026 to about $45 billion to $47 billion by May. OpenAI's run rate has been reported in the $25 billion to $33 billion range around the same period. A June industry revenue table put the two at roughly $78 billion combined. A separate securities-research summary citing SemiAnalysis projected Anthropic at $62 billion and OpenAI at $43.5 billion by June 2026, which would put the pair just above $100 billion combined.

Even allowing for differences in methodology, current revenue is large but nowhere near $1 trillion.

That changes the analytical frame. Investors are not valuing these businesses on present revenue alone. They are betting that usage-based AI can compound faster than earlier software categories, and that the leaders can hold enough pricing power and workflow control to make today's run rate an early marker.

Anthropic's implied private-market valuation has been reported around $965 billion to roughly $1 trillion. OpenAI has been cited around $850 billion to $880 billion. Those numbers imply rich multiples on current run rate, before adjusting for the fact that run rate is not contracted recurring revenue.

AI revenue scales like infrastructure consumption, not seat software

Investors tolerate those multiples because these companies do not sell software the usual way. Much of the revenue comes from API calls and inference usage, priced by tokens processed, model quality and workload intensity. Some broker research has estimated that 75% to 85% of Anthropic's revenue comes from API usage.

This has two effects.

First, the top line can scale faster than seat-based SaaS. If a customer moves from testing a model to embedding it in code generation, search, analytics, customer support and agent workflows, spend can rise sharply without the friction of headcount growth. Consumption expands with product depth, automation breadth and query volume.

Second, annualized revenue can look more dramatic than it would in traditional enterprise software. A run-rate figure often annualizes the most recent month or quarter. When usage is accelerating, that produces steep jumps in reported annualized revenue. Anthropic's move from roughly $30 billion to $47 billion in a matter of weeks shows why investors need to separate momentum from durability.

The term ARR can mislead in AI. In SaaS, ARR usually points to contracted or highly predictable subscription revenue. In foundation models, the number behaves more like annualized cloud consumption. It can be sticky, but it can also swing with workload mix, pricing, customer optimization and channel structure.

There is also a gross-versus-net question. Some analysts have warned that comparisons between labs can be distorted if one company reports on a gross basis while another nets out certain pass-through costs or partner economics. In a sector with heavy cloud and compute dependencies, that affects how investors judge quality of revenue and eventual margins.

The bull case rests on stickiness, not just speed

The strongest argument for these valuations is not that AI revenue is growing quickly. It is that enterprise AI, once embedded, may prove unusually hard to unwind.

A model provider that becomes part of software development, internal search, workflow automation and customer-facing applications can gain something more durable than consumer app traffic. It becomes part of operating infrastructure. That can support expansion revenue, lower churn and a widening moat built on data, developer tooling and workflow ownership.

Anthropic and OpenAI differ in mix but not in strategic logic. Anthropic appears more API- and enterprise-heavy. OpenAI has a larger consumer funnel but is also building enterprise products and developer distribution. In both cases, investors are watching whether usage grows because customers are experimenting or because they are standardizing around these systems.

That is a familiar software question with a new economic profile. In SaaS, stickiness often comes from user adoption and switching costs. In AI, it may come from workflow dependence, fine-tuned integrations, safety and governance tooling, and the cost of retraining teams and systems on another provider.

If that holds, run-rate revenue could deserve a higher multiple than a commodity API business would. The market is betting that at least part of this revenue behaves less like bursty cloud spend and more like mission-critical software infrastructure.

The bear case is still about costs

The counter-case is straightforward. Revenue may be scaling fast, but inference costs and capital intensity remain huge. If model performance improves only through ever-larger compute budgets, margin expansion could lag revenue growth for longer than private-market valuations assume.

Some academic and policy critiques argue that planned data-center spending is running well ahead of what realistic AI revenue can support. Others question whether enterprise customers will keep paying premium token prices once model quality converges and open alternatives improve.

Competition feeds that margin debate. If foundation models become more substitutable, pricing power weakens. If customers adopt multi-model strategies, vendor concentration falls, and so can the incumbent's hold on expansion revenue. And if hyperscalers absorb more of the value through infrastructure, distribution and bundled services, model providers may not capture as much of the profit pool as current valuations imply.

That makes the valuation debate less about whether AI is real than about where the economics settle. The market has already answered the first question.

What this means for tech valuation norms

AI is forcing a rework of software valuation frameworks. Traditional ARR multiples assume a relatively stable relationship between sales efficiency, gross margin, retention and long-term growth decay. Foundation-model companies break that template. They can add revenue at an unprecedented pace, but with more variability, more capex dependency and less clarity on steady-state margins.

That does not make the valuations irrational by definition. It means public and private investors are pricing in a different kind of software company, one closer to a hybrid of platform, infrastructure and application layer.

If Anthropic and OpenAI grow into the low hundreds of billions in annualized revenue over the next several years, the rest of tech will be repriced around that possibility. Cloud vendors, enterprise software companies, chipmakers and data-center operators would all have stronger grounds to claim AI-driven profit pools justify current spending and multiples.

If they do not, the correction will not stay with the labs. It would hit the broader chain of companies priced on the assumption that AI demand can absorb an enormous wave of capital.

Investors should start with the right number. Anthropic and OpenAI are not at $1 trillion in ARR. They are at a fraction of that in revenue and roughly that level in combined valuation. Even so, they are already large enough to challenge the old rulebook for how software businesses scale and how markets price them.