🎙 We held the fifth Lecture Club call on Stanford MS&E 435: Economics of the AI Supercycle.
Lecture 5 — Sachin Katti (OpenAI). Topic — AI infrastructure using the frontier lab as an example: inference vs training, data centers, chips, memory, energy. The call was led by Boris. Overall, it felt like another basic lecture (like No. 4), but the discussion turned out to be one of the best.
Some thoughts from the discussion:
• Boris (lead) — 3 takeaways from the ex-CTO of Intel: inference will grow to ~70% of compute (and it's not just for responses, but for synthetic/post-training/RL); giant data centers are still more profitable (economies of scale + shortage of electricians); recursion — AI designs chips itself (cycle ~3 years).
• Nikita (me) — latency is crucial: 40 ms already affects engagement in ChatGPT. But I live in Claude Code, where delays are in minutes, running 6-8 tasks in parallel and would pay for speed. Memory around GPU is primitive → CPU cache evolution will repeat. Dig deeper into the stack; AI wrappers will be consumed by the next model update.
• Ira — counter: Web Accessibility is a wrapper over AI, and it grows due to law (USA ~5 years, EU since last year). Money flows where regulation is, not where the deep stack is.
• Viktor — chips = main bottleneck, everything is monopolized (NVIDIA + fab at TSMC). Crypto analogy: small companies grew on mining chips — will it repeat in AI? Value leaks down to infrastructure.
• Grigory — game theory: it's beneficial for independent labs (there are ~4) to tacitly slow down the race → fewer features, everyone "breathes out." Energy limits help with this.
• Stepan — provocation: energy is not the bottleneck, chips (USA bought everything). Jensen's pyramid top — not "applications," but agent-employees; the main thing in them is context (the very "John who knows everything"). You pay for the result, like for a finished house, not for building materials.
• Mikhail — lectures have become irrelevant to practice (gigawatts, billions of tokens). The pie won't flip, but will add a layer (agents for salary). By estimates: to replace all white-collar workers, revenue of ~$1T (~1% of US GDP) and only +3% to electricity output is needed. The real bottleneck is the decision of managers to lay off: the last mile effect.
• Mars — AI will force even conservative energy to reconsider; DCs are designed so that launching thousands of GPUs doesn't spike network load. The logical continuation — AI will start designing infrastructure for itself.
• Nikita (closing) — layoffs will be slow (Amazon's tactic: not hiring + automating + creating outflow). The overarching conclusion of the course: every lecturer hits the hardware/memory/energy → bottleneck in infrastructure. Developers have become managers of agents; in companies, FOMO from above + sabotage from below.
📺 Call recording:
https://youtu.be/ixXB3KGDZJU
