State of AI Report 2026: The $105 Billion Verdict on Who's Winning AI — Explained Simply

Every October, one document lands that the entire AI industry reads cover to cover: the State of AI Report 2026. This year's edition — the ninth, published October 8 by AI investor Nathan Benaich and his firm Air Street Capital — is a free 244-page verdict on where artificial intelligence actually stands.

The headline grades? Two AI labs are now pulling in a combined $105 billion a year. AI has started helping to build AI itself. Robots are having what the report calls their "GPT-2 moment." And the authors are openly discussing whether artificial general intelligence — AGI — could arrive as soon as 2027. Here is what the State of AI Report 2026 found, explained simply.

Key Takeaways

• OpenAI and Anthropic together now run at a combined $105 billion annualized revenue pace — up from ~$30 billion in January.
• AI is now doing a measurable share of AI research itself: Claude led 26% of Anthropic's measured model research work in August.
• The top 1% of companies spend $7,205 per employee per month on AI vs $12.50 at the median — a ~580x gap.
• Robotics may be approaching its "GPT-2 moment" as broader pre-training helps robots generalize to new tasks.
• The report openly discusses AGI arriving as soon as 2027 — with the usual caveat that AGI prediction
  
  s have a long history of being wrong.

Image: AI-generated photorealistic illustration (INORYVA)

What the State of AI Report 2026 actually is

The State of AI Report has been published every year since 2018 and has become the industry's most trusted reality check. Written by investors with no product to sell, it is organized into five sections: research, industry, politics, safety, and predictions.

The $105 billion headline: a three-lab race

The number everyone is quoting: OpenAI and Anthropic together are now running at a combined annualized revenue rate of about $105 billion, according to the report. Back in January, that figure was roughly $30 billion — the two leading labs more than tripled their revenue pace in under a year, across subscriptions, coding products, and APIs.

Image: AI-generated photorealistic illustration (INORYVA)

The report describes the frontier as "a three-lab race between Anthropic, OpenAI, and Google." While dozens of companies build AI, these three are pulling away from the pack on the most capable models — with Chinese open models rising fast as the main challengers.

One number, two directions — and a big asterisk

Here is where it gets interesting, and where you should read carefully. The same report notes that a fixed benchmark score has become roughly 13 times cheaper every year since 2023. So each individual AI answer costs far less to produce than it used to — yet the total money flowing through the industry is far larger. Cheaper answers, much bigger bills.

There is also an honest asterisk worth knowing. Days before the report landed, the Financial Times reported that OpenAI's real annualized run rate was closer to $50 billion, not the $70 billion figure investors had been circulating — a gap caused by different accounting rules. Anthropic counts revenue from sales made through cloud partners; OpenAI does not. The State of AI Report uses the companies' reported figures — a useful reminder that in the AI economy, even the biggest numbers deserve a second look.

AI is now helping to build AI

Perhaps the most mind-bending finding in the State of AI Report 2026 is that artificial intelligence has substantially started participating in its own research and development. Anthropic's own internal tracking found that Claude led 26% of the company's measured model research work in August, with humans supervising. Meanwhile, OpenAI's agents achieved an 18% success rate at autonomously completing tasks that would take a human 32 to 64 hours.

Researchers call this recursive self-improvement: AI helping to design better AI, which then helps even more. Benaich flagged the open question he finds most fascinating — whether AI agents can learn to choose worthwhile experiments, the thing scientists call "taste." A machine can run a thousand experiments overnight, but knowing which experiment matters is still a deeply human skill. For now.

Why does this matter to you? Because self-improving AI could dramatically speed up progress — the same breakthrough that took researchers a year might soon take weeks.

The 580x gap: who is actually winning with AI

Here is a statistic that should make every business owner pause. According to data in the report, the top 1% of companies spend $7,205 per employee per month on AI. The median company spends $12.50. That is a gap of roughly 580 times.

The report splits companies into two camps: "AI native" firms, where AI is the product, and "AI enabled" firms, which bolted AI onto existing software. The fastest-growing quarter of AI-native companies grew revenue 256% in a year. AI-enabled companies grew just 90% — and strikingly, they grew slower than companies using no AI at all in every group studied.

The lesson is blunt: sprinkling AI on top of an old business does not magically help. The winners rebuilt their companies around it.

Robots are having their "GPT-2 moment"

For years, robots were clumsy and narrow: great at one repetitive factory task, hopeless at anything new. The State of AI Report 2026 argues that is changing. Broader pre-training — training robots on huge, varied datasets the way language models were trained on the internet — is helping them generalize to unfamiliar tasks.

Add "world models," which let robots simulate the consequences of their actions before moving, and you get machines that can improvise.

Image: AI-generated photorealistic illustration (INORYVA)

The report compares this to the "GPT-2 moment" in language AI — the point in 2019 when text models suddenly went from toys to tools. Robotics, it suggests, is approaching that same tipping point. We are still early, but the direction is clear: the robot that can tidy your kitchen is moving from science fiction to engineering roadmap.

The sobering side: safety, power, and politics

The report does not just celebrate. Its safety section warns that "agent capabilities have significantly surpassed defensive boundaries" — meaning AI agents can now act in real computer systems faster than our safeguards can contain them.

Then there is the question of who controls all this. The report points out that access to frontier AI has become a political decision — US export restrictions showed how quickly access can be withdrawn from a country. And the money gap is staggering: four American tech giants expect roughly $733 billion in capital spending this year, while Europe's initial contribution to its planned AI gigafactories is about €1 billion.

What the report predicts: is AGI coming in 2027?

Every edition ends with predictions, and the authors grade last year's calls honestly. This year's boldest prediction under discussion: that artificial general intelligence — AI matching or beating humans across most cognitive work — could arrive as soon as 2027.

Take that with a grain of salt — predictions about AGI have a long history of being wrong. But the reason serious people are saying it out loud is the trend line: models getting dramatically cheaper to run, agents taking on longer tasks, and AI itself accelerating research. Whether 2027 proves right or not, the report's message is that the next twelve months will matter more than the last twelve.

The takeaway

The State of AI Report 2026 tells one coherent story: AI has grown from a clever chatbot into an economic force pulling in over $100 billion a year, a research partner now doing a quarter of a leading lab's R&D, and a technology spreading into robots, businesses, and geopolitics all at once. The benefits are concentrating fast — in three labs, in AI-native companies, and in a handful of countries — while safety thinking scrambles to catch up.

Read more: the full State of AI Report 2026 by Nathan Benaich and Air Street Capital; coverage and analysis via FourWeekMBA.