OpenAI and Anthropic have both posted exceptional increases in their revenue run rates (annualized from recent revenue). Together they now earn at roughly a $100 billion per year run rate; industry estimates put the entire generative AI market approaching $200 billion per year (Exponential View). These figures indicate that Claude- and GPT-class models currently attract nearly ten billion dollars of monthly spending worldwide.
Key numbers and timeline
- OpenAI’s run rate was about $13 billion last August; over the past year it roughly tripled and now exceeds $40 billion.
- Anthropic’s revenue grew from $1 billion to $9 billion in 2025, and its run rate more than tripled in Q1 2026; reports indicate it reached about $65 billion by the end of July.
- Together, the two labs tripled in 2024 and then grew more than fourfold in 2025; in 2026 they have so far increased approximately 3.5×—from roughly $30 billion to $105 billion combined as of August.
It is important to note that the firms use different revenue accounting practices: when tokens are sold via cloud platforms, OpenAI records only the cut it receives from the cloud provider, while Anthropic books the full payment, which inflates Anthropic’s reported revenue relative to OpenAI’s.
What might explain the surge?
Explanations split broadly into two categories: technical progress and diffusion. One hypothesis sees the 2025–2026 revenue spike as at least partly temporary, triggered by one or more critical capability thresholds—such as coding agents reaching high usefulness around the Opus 4.5 timeframe. A precedent exists in the ChatGPT/GPT-4 era: GPT-3 was initially a niche product generating at most tens of millions per year, then ChatGPT and GPT-4 produced an explosion of usage and revenue that brought OpenAI to roughly $1 billion annualized in 2023.
By analogy, the recent boom in revenue from coding agents could subside quickly as products mature and market diffusion saturates. Alternatively, if each new capability level spawns its own diffusion curve, growth could be sustained through stacked S-curves of adoption.
Economic significance and longer-term outlook
The authors note that if the 2026 growth rates were to persist (for example, a 3× annual growth), the implied expansion would be enormous—naively, such a trajectory could approach the size of the current global economy in roughly six years. That forecast is not meant to be realistic; it is intended to illustrate how large the current growth rate is. More plausibly, if revenues and model capabilities continue to rise through 2027, AI could become a trillion-dollar industry over time.
In the short term, rapid revenue growth shows that AI has become substantially more useful than it was months earlier, and those revenues fuel further investment and compute capacity in a ‘‘compute feedback loop’’: better models generate more revenue, which funds more compute, which enables better models.
Risks and what to watch next
The central question is how robust and durable this revenue dynamic is: when and how quickly will growth curves start to bend, if at all? Each additional quarter or year of continued hypergrowth provides evidence about whether frontier AI follows a fundamentally different trajectory from prior technologies or simply undergoes a late but finite growth spurt.
Finally, while OpenAI and Anthropic are the dominant frontier model developers, most other dedicated model developers remained below $1 billion per year in run rate as of late 2025/2026. The further course of the industry will depend on a combination of technical progress, the shape of diffusion for each new capability, and differences in revenue accounting across firms.
Summary
OpenAI and Anthropic produced unprecedented revenue expansion between 2024 and 2026, and by August 2026 their combined run rate stood at roughly $100–105 billion per year. The next few years will show whether this pace is sustainable: if it is, the economic transformation could be dramatic; if not, growth may revert toward the slower rates seen in more mature technology sectors.
