Nathan Benaich: AI enters its financialisation stage

Key points

Key takeaways from the TBPN interview with Nathan Benaich, founder and solo general partner of Air Street Capital and author of the State of AI Report (October 2026):

Financialisation stage. AI has moved from recipe discovery to delivery: the buildout is now limited by capital, power, land and permitting rather than algorithmic progress, with hyperscalers spending around $700B a year on compute and potentially doubling annually.

Frontier or inference. “You either die getting to the frontier or you live long enough to serve inference” — companies that failed to build frontier models are pivoting into data-centre businesses serving customers such as Chinese model labs, and some inference providers run at thin or negative gross margins while raising heavily oversubscribed rounds.

Compute allocation. Pre-training's share of GPU utilisation has fallen from roughly 50% toward 10% while reinforcement learning grows and inference holds around 30% — a sign the recipe is maturing, with less exploratory pre-training spend, more RL hill-climbing and frontier models improving through data curation rather than ever more data.

ARR as demand proxy. Annualised run rate remains the best demand metric available and criticism of it is overblown; the substance is in renewal behaviour — at one Air Street portfolio company, roughly 30% of new customers return within three months to add more token spend than their entire annual contract.

AI-first growth. In a cohort comparison, the top quartile of AI-first companies grew revenue three times faster than AI-augmented SaaS between $1M and $20M, and about 100% faster above $20M.

Nvidia as central bank. The market narrative swings between “AI is dead” and “AI is so back” on a three-week cycle while users see continuous progress; Nvidia now functions as the central bank of AI, and the binding constraint is capital and build sites rather than chips.

Genius bar for AI. The biggest gap is adoption: 99% of people do not know how to use these tools, and Benaich suggests OpenAI or Anthropic should open consumer pop-ups — a “genius bar” — to teach the other 99%.

Political headwinds. Last year's State of AI Report predicted that “nimbyism” over data-centre buildout would become influential by the midterms; Benaich says power and permitting constraints are already getting worse.

Read more: TBPN on YouTube · TBPN channel · State of AI Report