Nathan Benaich on the State of AI Report: inference is now the business, and the build-out is limited only by capital

Nathan Benaich on the State of AI Report

Nathan Benaich, partner at Air Street Capital, publishes the annual State of AI Report to cut through what he sees as a media cycle that swings between "AI is dead" and "AI is back" every few weeks. The goal is a research-grounded, opinionated read for people who want something between the infinity take and the zero take.

His bottom line on the current moment: the build-out is limited by capital, not capability. Benaich says conversations with labs, hyperscalers, and neo-clouds consistently surface the same constraint — they know how to scale, the demand is there, but power, land, construction, and permitting are the binding problems. He predicted last year that NIMBYism would become a meaningful drag, and it has.

“Benaich: 'You either die getting to the frontier or you live long enough to serve inference.' On the infrastructure build-out: 'Progress is very limited by capital. We understand the recipe. The demand is there. Just need to deliver it in more efficient ways.' On AI-first vs. SaaS companies: 'The top quartile of AI-first companies grew revenue three times faster between $1M and $20M ARR, and 100% faster above $20M.'”

Inference is now the business

The sharpest observation is structural. Companies that tried to build frontier models and failed are pivoting into large data center businesses, often serving Chinese models. Benaich's line: "you either die getting to the frontier, or you live long enough to serve inference." The biggest inference companies today weren't born as inference companies — a sign that even the people building this technology didn't anticipate the demand scale.

Hyperscalers are spending roughly $700 billion a year on infrastructure and could double that year-on-year. Neo-clouds are reaching billions in contracted revenue and gigawatts of compute. Nvidia, in Benaich's framing, has become the central bank of AI.

The shift from pretraining to RL

On the widely circulated chart showing compute allocation shifting from pretraining toward reinforcement learning, with inference holding steady around 30%, Benaich reads it as a natural progression. The pretraining recipe is largely solved — data curation has improved to the point where some frontier models are getting better by removing low-quality data rather than adding more. RL spending is rising because it lets companies explore beyond the local maxima that human-generated data can reach, with the ceiling set only by how much you're willing to spend and how valuable the target problem is.

Revenue metrics

On ARR criticism, Benaich argues the substance is real but the signal is buried under gaming. The metric he finds more useful comes from one of his portfolio companies: 30% of new customers return within three months to add more token spend than their entire original annual contract. Separately, the State of AI Report found that AI-first companies in the $1–20M revenue band grew three times faster than SaaS companies that added AI features. In the $20M+ band, AI-first companies grew roughly 100% faster.

The financialization of AI is the right frame for where things stand. The recipe is understood, the demand exists, and the primary question is whether enough capital shows up to deliver it at scale.