Training Data

The episode is genuinely dense with non-obvious technical and economic claims - the 100x co-design thesis, the throughput-interactivity curve as the master curve, power density breakthroughs beyond 1W/mm², and Jensen's multipolar world strategy are all substantive. The long origin story and some filler in the opening drag the average down.

“you take what could have been a 2x here, 2x here and instead of being multiplicative to 8x it's actually 100x because you've optimized it across all three layers”

“model costs drop for equivalent quality by like 60x a year”

Several genuinely fresh frameworks: the reframing of the CUDA moat as an ecosystem co-optimization problem rather than a developer tooling moat, the local-vs-global minima framing for ASIC bets, and the Jensen multipolar world thesis are all counterintuitive and non-recycled takes. These are not the standard semiconductor talking points.

“what people call the CUDA moat is not actually anything to do with Cuda, but it's like the fact that Deepseek, Kimi and Zippuai and Alibaba and Tencent...their models are a co design for GPUs and therefore if I want to run them on GPUs actually in some cases they don't run really well on tpus”

“A world where open, anthropic and Google models are the only models, is one in which he's screwed”

Dylan Patel is the genuine article - he built SemiAnalysis to 90 people and reportedly $100M revenue by attending 40+ conferences a year, cultivating primary supply-chain sources, and producing original research that moves industry. He has proprietary data, named contacts, and is clearly plugged into confidential deal structures. This is an actual practitioner-researcher, not a thought leader.

“we have 90 people and like a big chunk of them are technologists, engineers across the whole supply chain”

“I go to 40 plus conferences a year”

Extremely specific throughout: named rental rates per gigawatt for Trainium vs GPU vs SpaceX deals, Anthropic's Q2 profitability status and per-token margins, data center pricing trajectories in $/kW/month, and chip-level architectural details like NV Link connecting 72 GPUs vs Google ICI connecting 8,000. The level of named data points is rare for a podcast.

“Trainium uh sells at sub $10 billion per gigawatt rental rate uh to anthropic and to OpenAI GPUs at least before the craziness of the last 6 months usually went around 12 to $13 billion per gigawatt”

“Anthropic in Q2 is profitable, their net income profitable, um, excluding stock based compensation...their per token margin is so high”

Sean's best moment is a genuinely setup-for-disagreement question on where efficiency gains originate, which Dylan immediately contests and uses to deliver the co-design thesis - that's skilled hosting. The oil/Saudi Arabia analogy to probe data center quality differentiation is also creative. However, the origin story runs very long without redirection, and several questions are framed as agree/disagree softballs.

“Sean, I completely disagree with you by the way”

“To me it seems like in the last three years most of the games have come from hardware level and some from the model level. Like do you think that that uh, is what, do you agree with that?”