Enterprise AI hit a US$177bn run rate, Menlo estimates; Anthropic and OpenAI take 86%

01The line: US$1.7bn to US$177bn in three years

Menlo splits AI spending in two. Deployment, using AI, was US$176.8bn: US$99.5bn of infrastructure (model inference sold by the labs, inference platforms and agent infrastructure) and US$77.3bn of applications. Development, building models, was a further US$102.5bn: US$87.2bn of training compute and US$15.3bn of training data and services. The headline number is deployment only.¹

Menlo’s explanation for the jump is agents: software that carries out multi-step work rather than answering one question. Spending averaged 4.7 times the previous year’s against a planned 1.9 times; the median company spent 2.2 times against a planned 1.8 times. The share of enterprises using AI rose from 55% to 65%, so most of the growth came from more spending per company, not more companies.¹

02Where it goes: two labs, and coding

Menlo’s shares are “estimated dollars spent based on proportion of production API usage”: respondents reported which models ran their AI workloads, weighted by their AI spending.¹ On that measure Anthropic went from 40% to 55%, which Menlo attributes to its Opus models in long-running tasks and coding; OpenAI from 27% to 31%; Google from 21% to 9%, its spending now skewed to cheaper Flash models.¹²

Applications. Coding was US$34.0bn, up 7.7 times, and the labs’ own tools, Claude Code and Codex, took US$24.0bn of it (70%). Enterprise chatbots were US$25.0bn. The two make up 76% of the US$77.3bn application market; healthcare, third, was US$3.7bn.¹

Open models. Open-weight models handled 19% of tokens but took about 4% of model spending, because their average price fell 3.3 times to US$0.13 per million tokens, against a blended US$0.70 for closed models, down 24%.¹

03What the numbers cannot tell you, and what to watch

The limits. The survey is of US companies. The spending totals are modelled estimates, and annualised run rates are not a year’s revenue. Menlo restates its own past figures, so editions compare only on the newest basis. Some comparisons mix groups: agents are used by 74% of AI builders this year, against 27% of startup deployments in the 2025 survey. And the publisher holds stakes in the leading lab and in several infrastructure firms the report names.¹²

What to watch. Anthropic’s run-rate figure for 31 October is left blank in its 2 November amendment, so the number the company itself uses for its current pace is not yet public.³ Menlo’s next edition will show whether its 86% holds after the discounts on flagship models it says began “in recent weeks”.¹ On the return side, our August analysis of McKinsey’s survey found 37% of firms reporting any profit from AI.⁴