← All episodes

The Supply and Demand of AI Tokens | Dylan Patel Interview

| 16 products mentioned
Watch on YouTube ai economics token demand and supply frontier models execution vs. ideation ai adoption infrastructure bottlenecks competitive advantage

Dylan Patel of SemiAnalysis details an explosive shift in AI token economics: his firm's annual spending exploded from tens of thousands to $7M annually in just months as non-technical team members discovered Claude for coding and other high-value tasks. As implementation costs collapse and execution becomes trivially easy, the strategic advantage now lies entirely in identifying *which ideas are worth the capital spend*—and those with early access to frontier models like Anthropic's Opus and Mythos will capture outsized economic value before token demand outpaces supply infrastructure.

Key takeaways
  • Ideas, not execution, now drive competitive advantage: execution on AI has become so cheap that the constraint shifts to choosing which ideas justify token spend; this inverts the old startup adage and means founders must get comfortable with massive token bills on the critical path to growth.
  • Token demand is exploding far faster than supply can scale: a single analyst spent $6,000/day for weeks building a US power grid model that would have required 100 people a decade to create; CPUs, memory (DRAM), and logic chips are all sold out with lead times extending to 2027–28, creating a multi-year supply crunch that will drive margins up across the entire hardware stack.
  • Frontier model access is becoming a competitive moat: Anthropic's Mythos (unavailable to most users) represents a 2-month jump from L4 to L6 engineer capability; those with enterprise contracts and capital to burn will get earlier, unfettered access and will be able to "crush" competitors who rely on cheaper, older models or rate-limited tiers.
  • Phantom GDP masks real economic value creation: traditional metrics don't capture the massive deflationary benefit of AI-accelerated work; a $7M annual token spend is generating multiples more in economic value than it costs, but that value doesn't show up in conventional GDP measurement, making it hard for CFOs and executives to justify the spend upward.
  • Don't use slower or cheaper models once you've experienced frontier capability: users shown a 10% better model immediately abandon previous versions; this "Claude psychosis" behavior means cheaper alternatives (like GPT-4 class models) can't capture demand even if they hit the same capability tier as prior frontier models, because users don't want to work slower.
  • Robotics represents a second, massive demand curve still in early innings: few tokens are currently consumed by physical systems, but once vision language action models and reinforcement learning improve via easy implementation, robot adoption and token consumption will explode, extending the supply shortage years further.

Recommendations (6)

ChatGPT
ChatGPT uses

"Last year we thought we were heavy users of AI. Everyone's using ChatGPT."

Dylan Patel · ▶ 0:42

Claude
Claude uses

"Everyone's using Claude. Everyone's got you know I'm providing whatever subscriptions anyone wants"

Dylan Patel · ▶ 0:45

Opus
Opus uses

"it really started in late December with Opus that included Doug who's president"

Dylan Patel · ▶ 0:58

Anthropic
Anthropic uses

"we signed, you know, an enterprise contract with Anthropic and it's gone to the point where now, um, I think when I last talked to you it was 5 million spend rate. It's actually 7 million spend rig..."

Dylan Patel · ▶ 1:23

CoreWeave
CoreWeave uses

"it's able to create this application that is GPU accelerated runs on a server that we have at CoreWeave"

Dylan Patel · ▶ 3:02

AWS
AWS uses

"or some AWS instance or some bucket that we've spun up all of that requires CPU"

Dylan Patel · ▶ 37:43

Mentioned (10)

Jane Street
Jane Street "especially like the super like the Jane Streets of the world and the Citadels. They're they're re..." ▶ 10:14
Citadel
Citadel "especially like the super like the Jane Streets of the world and the Citadels. They're they're re..." ▶ 10:16
Mythos
Mythos "they have Mythos, they have Opus 4.7. So they clearly didn't decrease their research compute spend" ▶ 11:38
DeepSeek
DeepSeek "DeepSeek, for example, on GPT-4 was 1/600th the cost. And since then, the costs have fallen furth..." ▶ 15:33
GPT-4
GPT-4 "Of course, no one gives a crap about GPT-4 class models. They want the frontier because the front..." ▶ 15:41
ASML
ASML "You see ASML is completely sold out and they need Carl Zeiss to expand faster" ▶ 32:02
Carl Zeiss
Carl Zeiss "You see ASML is completely sold out and they need Carl Zeiss to expand faster" ▶ 32:04
TSMC
TSMC "TSMC just had their earnings. Uh, they keep upping capex. Ultimately, you know, it takes them qui..." ▶ 35:32
Lam Research
Lam Research "companies like Lam Research or Applied Materials or ASML or their further downstream supply chains" ▶ 36:46
Applied Materials
Applied Materials "companies like Lam Research or Applied Materials or ASML or their further downstream supply chains" ▶ 36:48