“Everything will work”: David George on a16z’s growth thesis in the AI buildout

Key points

Key takeaways from the Uncapped with Jack Altman interview with David George, General Partner and Head of Growth at Andreessen Horowitz (October ​2026):

“Everything will work.” George endorses Eric Vishria’s framing that AI’s big debates are false binaries: open source will win and frontier labs will win, Nvidia and new chipmakers will both prosper — the answer is “and”, not “or”.

Infrastructure is the new railroad. Data-centre construction has just overtaken railways as a share of GDP,and George expects capacity spending to grow roughly tenfold over the coming years — multi-trillion-dollar annual outlays. Today’s 4GW of AI capacity is already supply-constrained,and an 8GW world would, he believes, be fully consumed.

A tiny base of paying customers is driving unprecedented revenue. Frontier labs and products like Cursor now add revenue at a combined run-rate of roughly $120bn — more incremental monthly revenue than any hyperscaler except Amazon —with ~30M programmers comprising most of today’s paying demand, against 1.5bn knowledge workers globally and 1bn+ consumers;enterprise penetration remains below 5%.

Inference, not training, is the mega-market. Inference demand will dwarf training; reasoning at test time keeps improving output quality—the more you compute,the better the result—and George struggles to construct a scenario where new capacity goes unused.

Frontier models have pricing power. Users willingly pay roughly 10x more for frontier tokens, especially in coding,and frontier compute should be reserved for the highest-value tasks such as cancer research; as N-1 models approach frontier quality, abstraction layers will emerge, contested by apps, labs’ first-party products and data platforms like Databricks, Palantir and Snowflake.

Labs’ app ambitions have limits. Labs will prioritise first-party products in coding and general knowledge-work productivity — Microsoft Office and Google Apps territory—but platform trust dynamics restrain them:a platform is only a platform if everything built on it earns more than it does. Vertical depth keeps apps like Harvey viable—legal adoption lags coding by roughly 12 months,but end-clients now mandate AI usage,and Kirkland’s $500m build-vs-buy decision shows the tension. Meanwhile Replit is thriving.

Consumer AI: mass but shallow. More than a billion people now use ChatGPT—the first true mass consumer AI moment—but most usage remains simple and often search-like, while younger users go considerably deeper.

Read more: Watch on YouTube