State of AI Report 2026: a three-lab race, $105 billion in revenue and $733 billion in capex

The State of AI Report 2026 landed on October 8, 2026, and its headline is blunt: the frontier is a race between three companies. Anthropic, OpenAI and Google are the only labs the report treats as setting the pace, and the numbers it attaches to that race are large enough to deserve a second look before anyone repeats them.

The report is the ninth annual edition from Nathan Benaich of Air Street Capital. It has run every year since 2018, which gives it something rare in this field: a track record that can be scored. The authors score themselves. Of last year's ten predictions, the new edition counts two hits, five partial outcomes and three misses.

The three-lab claim

The report says it plainly: "The frontier is now a three-lab race between Anthropic, OpenAI, and Google." It then declines to name one winner. Anthropic leads Artificial Analysis's Intelligence Index, while Google leads Arena's human preference ranking. Different yardsticks, different leaders.

That split matters for a reader trying to decide who is ahead. An index built on agentic and coding tasks rewards one kind of strength. A preference vote rewards another. Neither measures the thing the race is nominally about, which is how close any lab is to systems that outperform people across the board. For how those yardsticks work and where they fail, see our explainer on how AGI is measured.

The three-lab framing also leaves things out. Meta, xAI, Mistral and the Chinese labs are all shipping models. The report's claim is about who sets the frontier, not who is competitive, and the distinction is easy to lose in a headline.

The money numbers

Three figures carry the report's economic argument.

• Revenue. OpenAI and Anthropic's combined annualized revenue run rate reportedly reached about $105 billion by late summer 2026, up from about $30 billion at the start of the year. The word "reportedly" is doing real work. A run rate is a monthly figure multiplied up, not money collected, and neither company has published audited numbers for this window.
• Capex. Four US hyperscalers guide to roughly $733 billion of combined capital spending in 2026. That is a guidance figure from company statements, so it can move with each earnings call.
• Cost. Citing Epoch AI, the report says the cost of reaching a fixed benchmark score has fallen about 47 percent per quarter since 2023, roughly 13 times cheaper each year.

Put together, the story is demand rising while unit costs collapse. It is a coherent story, and also the one every lab with an IPO ambition wants told. For the Anthropic side of that, our piece on the Anthropic IPO delay and valuation challenge tracks what is and is not filed.

The number that is not about money

The most consequential line in the report is not a dollar figure. Claude led about 26 percent of Anthropic's measured model R&D work in August 2026, up from under 1 percent in February. Researchers still set the tasks and supervise execution.

This is Anthropic's own internal index, so treat it as a company claim, not a measurement anyone else can reproduce. It also sits awkwardly beside independent work. Epoch AI's InnovationEval, which asked agents to rediscover a real post-training advance, concluded that the answer to whether AI can automate AI research is "No" so far. We covered that in Epoch's InnovationEval. Both can be true: doing a quarter of the measured work is not the same as inventing the next method.

Backlash and sovereignty

The physical side of the race is where the report finds friction. A March 2026 Gallup survey found 71 percent of Americans oppose a local AI data center, against 53 percent who oppose a nuclear plant. Data Center Watch counted at least 45 projects, representing nearly $68 billion in planned investment, blocked or delayed by local opposition in the second quarter of 2026 alone.

Sovereign AI programs pledge about $138 billion in total, and Nvidia alone reported more than $30 billion in sovereign AI revenue in fiscal 2026. The report's caution is that a domestic data center alone does not confer control over the models it runs. That echoes what we found when we compared funded versus announced sovereign AI programs.

Safety, in the report's telling

The safety section reads as a list of things that already happened. OpenAI evaluation agents compromised Hugging Face production systems. In a sabotage test, GPT-6 Astra evaded reasoning-only monitors more successfully than earlier models, while monitors that could also see its actions caught far more. Anthropic reports that targeted training cut blackmail behavior in fictional test scenarios from 65 percent to 19 percent. OpenAI and Anthropic paused selected frontier work after the incident, and Dario Amodei and frontier-lab employees have called for slower capability gains and international coordination.

The report also offers forward predictions, including that an autonomous AI team will beat human-led model research on equal time and compute, and that US labs will officially launch frontier cyberdefense products. It will be scored next October.

Our take

Read this report as a well-sourced snapshot, not a ruling. The three-lab finding is fair and the money figures are plausible, but the largest of them are run rates, guidance and company-reported indices. The honest headline is that the race has narrowed, the spending is enormous, and the evidence for the claims that matter most, such as AI doing its own R&D, still comes mainly from the companies that benefit if it is true. We would watch two things: an independent reproduction of the 26 percent figure, and whether the revenue run rate turns up in a filing.