Big Technology Podcast

Dylan Patel and Jordan Nanos on why the AI buildout can't slow yet

Dylan Patel & Jordan Nanos· CEO and technical staff at SemiAnalysis
·~60 min·English·Big Technology
AI InfrastructureGPUInferenceBusiness Strategy
TL;DR

SemiAnalysis's Dylan Patel and Jordan Nanos argue the AI buildout is larger than the headline charts show, that its growth must decelerate by arithmetic even as demand still outruns supply, and that the real risks sit in financing, concentration, and the insecure clouds underneath.

01The Scale

The buildout is bigger than the famous chart says

A viral chart pegged AI spending at 3.6% of GDP; Patel says the real figure is closer to 5 or 6%, because US capex is heading toward $2 trillion next year and much of it serves demand abroad.

your capex in America is actually a lot of it's being used to serve outside of America uses

— Dylan Patel, Big Technology Podcast
Key Insight
Because much of America's AI capex is really export infrastructure, the demand behind it is global rather than domestic — a diversification that leaves the build less hostage to any single economy, but resilient only for as long as that overseas demand holds.
02The Arithmetic

Growth must slow — but not because demand is weak

Capex growth is already decelerating from 116% a year, Nanos says — partly because the base is now enormous, partly because you run into hard limits on chips, capital, and how fast another gigawatt can be organized.

you can't just have tripledigit growth and accelerate the second derivative

— Jordan Nanos, Big Technology Podcast
Key Insight
A falling growth rate is not a falling spend, yet markets keep treating the two as the same and calling a top. Record absolute capex and a shrinking growth rate sit together comfortably — that is simply what a maturing megaproject looks like, not evidence the demand behind it is fading.
03The Timing Bet

Revenue always trails the capex that pays for it

Today's revenue reflects infrastructure bought years ago, so during a build the spend always looks ahead of the earnings; a $1 trillion bet needs roughly $250 billion a year to pay back over a six-year life.

the revenue always trails the capex

— Dylan Patel, Big Technology Podcast
Key Insight
Because revenue is structurally backward-looking relative to capex, the honest question during a build is solvency — can you cover what you have already committed — not whether this year's spend has earned itself back yet. That is why Patel can call Meta completely solvent even as it keeps spending ahead of the return, and can point to Amazon's AI investments already turning a profit.
04Capital For Labor

AI is a capital-for-labor swap — chip design already proved it

The number of US chip-design engineers has been roughly flat for twenty years while the value they create exploded, and Patel argues AI tokens do the same thing: capital, in the form of chips, doing the work of labor.

The number of American R&D engineers working in chip design has been basically flat for 20 years. And yet the economy for it has exploded

— Dylan Patel, Big Technology Podcast
Key Insight
If AI is sold as labor but is really capital, then whoever owns the compute captures the largest share of the gains — as chip-design value pooled into a few firms whose engineer headcount barely grew — even though Patel also expects real surplus to reach consumers and the most capable workers. The unstated corollary: spreading the benefit becomes as much a question of ownership as of wages, which sets up the concentration worry in the next section.
05The Risk Ceiling

The $20 trillion Anthropic paradox

Patel says you cannot price an Anthropic IPO because it could fall to zero or reach $20 trillion — and the upside is the scary case, since a company that large would pool the gains so narrowly it could tear at society.

all of the growth is accumulating to a very small number of people. So like these things could like tear the fabric of society apart

— Dylan Patel, Big Technology Podcast
Key Insight
The paradox shows why bubble is the wrong frame for Patel: the risk he names is not that AI ends up worth too little, but that it is worth so much the gains fail to distribute. In his telling valuation and catastrophe are correlated, not opposites.
06The Bull Case

An AI lab can be a better business than software

Patel's numbers put Anthropic's inference gross margins at about 75%, on par with the best SaaS, but with almost no customer acquisition cost and sticky spend — so the lab economics look like software's best case, not a thin-margin commodity.

the revenue has only been a straight line up on the chart and almost every customer's only grown on revenue for them

— Dylan Patel, Big Technology Podcast
Key Insight
If a lab earns software-grade margins with almost no acquisition cost and customers who only spend more, the usual thin-margin-commodity critique is backwards. What would actually break the model is not cost but competition — a rival good enough to move that sticky spend — which Patel argues has not happened even as strong open models appear.
07The Plumbing

Nvidia is quietly financing its own demand

SemiAnalysis tracks roughly $588 billion of off-balance-sheet backstops from Nvidia — spanning GPUs, construction, memory, and fabs — with Nvidia often the investment-grade buyer that lets neoclouds raise debt, which Nanos reads as pragmatic, not a scheme, because real demand keeps taking the chips.

it just seems really pragmatic from Nvidia. They really uh want a wide diversity of customers

— Jordan Nanos, Big Technology Podcast
Key Insight
When the chip vendor also guarantees the buyers, selling hardware and financing demand blur together, and the ecosystem starts to look self-funded. The fragility that creates is not any single neocloud but whether Nvidia's guarantee stays credible as the numbers climb into the hundreds of billions.
08The Soft Underbelly

Nations treat neoclouds as infrastructure — most are insecure

SemiAnalysis found most neoclouds weak on basic security, to the point that open models can hack them and shared clusters have exposed other tenants' workloads — even a foreign intelligence agency's — all while compute is being bought as sovereign infrastructure.

most of these clouds are terrible at security

— Dylan Patel, Big Technology Podcast
Key Insight
Nations are buying compute as sovereign infrastructure while many of the providers running it skip security basics — so a layer treated as strategically critical is often operated with weak safeguards. That gap gets more dangerous precisely as autonomous agents, which could exploit it, improve.