The trillion-dollar physics of the AI buildout — and where the payback comes from
Key points
Key takeaways from Source interview with SemiAnalysis founder Dylan Patel and technical staff member Jordan Nanos on Big Technology Podcast (October 2026):
The buildout is bigger than the charts say. Viral comparisons put the AI infrastructure buildout at a median 3.6% of GDP versus 2.2% for railroads, but Patel argues the real numbers are higher: US capex next year will approach $2 trillion including the full supply chain, close to 5–6% of GDP — larger than any physical infrastructure cycle ever, partly because US capacity serves inference demand from Europe and Asia too.
Growth must decelerate. Hyperscaler capex grew 116% year on year in Q3 2026, yet Wall Street already models it dropping below 100% and toward 70% next year — Patel agrees limits hit towards the end of next year, not from demand but from the sheer difficulty of marshalling another gigawatt and raising trillions in capital.
Definitional deceleration. Anthropic went from under $10B to north of $100B revenue — a 10x year — and Nvidia's growth is still 2x year on year, but doubling from an ever-larger base means the growth rates have to shrink even if the dollar numbers keep rising.
Revenue can still clear the bar. Bain's claim that hyperscalers need $6 trillion in annual revenue by 2031 to break even and a Columbia professor's $3.5 trillion AI-services figure both look too low to Patel: he sees OpenAI and Anthropic alone reaching $600–700B revenue by December next year and $1.5 trillion by the end of 2027, with Anthropic already profitable on compute and OpenAI two to three quarters behind.
Amazon's AI is profitable today. The biggest infrastructure builder is making strong gross margins on Bedrock — AI has actually lifted overall gross margin — and Patel dismisses claims of Meta-style debt risk: Meta can stop adding capacity whenever it likes and its cash flow covers committed spend.
Where the money comes from. Revenue arrives as R&D and COGS budgets shift to AI and products improve — but the deeper mechanism is a rebalancing of labour versus capital: AI tokens are capital disguised as labour, accelerating the decline in labour's share of GDP that has run since the 1970s.
Chip design as the template. US R&D chip-design headcount has been flat for 20 years while the value created exploded — AI-assisted design is the same lever applied across every industry, from phones and robots to sensors.
Beyond per-token pricing. Nanos argues AI revenue will be measured differently as frontier models automate knowledge work: returns from robotics, self-driving, drug discovery and life sciences could resemble the returns seen in coding, if they are the only players able to compete at the frontier.
Read more: Big Technology Podcast (YouTube)
