🎙 We held the second call of the Lecture Club on Stanford MS&E 435: Economics of the AI Supercycle.
Lecture 2 — fireside chat with Brad Gerstner (Altimeter Capital) and Sunny Madra (Groq → Nvidia). The topic — the economics of inference and compute as a bottleneck for AI. The waiting room and password on Zoom worked — we didn’t catch any "uninvited guests" this time.
A few thoughts from the discussion:
• Nikita (me) — The course is led by an Altimeter employee, the guests are from portfolio companies, so we filter out all the "we are geniuses" talk. The main practical signal — the rental price of H100: the 5-year-old chip still lacks supply. And — the models we are using today were trained on Hopper, not on Blackwell/Rubin. The most interesting is yet to come.
• Jean (futurist) — GDP as a metric is originally manipulative (invented in the 30s, distorted by military spending since WWII). The graph of "doubling GDP" from the lecture is a pretty picture, not a scientific argument. And the rhetoric of "we urgently need new hardware" is partly market lobbying for itself.
• Alexander — Positive gross margin of Anthropic = the bubble is unlikely to burst but will "deflate" by discarding the weak. A parallel conclusion for each of us: either you become a deep specialist, or AI will learn from your actions and replace you — just like robots replaced sorters in Chinese warehouses.
• Pasha — Inside Nvidia, Jensen demands "×100, no less" from each iteration — this reshapes the planning horizon. A historical parallel: the transition from foraging to agriculture = the transition from labor to services. IQ is being commoditized, EQ is becoming the new rarity.
• Artem — A breakthrough in inference may come not from hardware but from model architecture: fewer tokens per request = the same effect as new chips. Separately: the mentioned Madra's "deterministic AI" — potentially the next quantum-like breakthrough, because determinism is critical for business processes.
• Dmitry — Response speed is becoming a key UX metric for AI products. Big players will start making their own chips — the precedent of Apple M-series. Europe risks missing the wave due to slow regulation and turning into a "tourist economy".
• Ilina — Devil's advocate: technologies are changing faster than people can adapt. Between "AI is already here" and "people are not yet adapted" — there is a window for a crisis on the scale of the COVID lockdown. Enthusiasm exists, readiness does not.
• Victor — A parallel with 2000: those who invested in S&P 500 at the peak of the dot-com bubble broke even after 14 years. The main anti-bubble indicator is the size of real demand. And a counter question: who is actually using 5-year-old H100s and for what?
• Artem (second thought) — We are at a fork in scenarios. One of them is cyberpunk: high-tech + low life, corporations more powerful than states. A provocative alternative: build an economy on "agent slavery" — humane, and wealth distribution is solved through the agent owner's salary.
• Andrey — Perhaps the next leap will be analog chips for specific networks (just as GPUs once separated from CPUs). And separately: the conversation lacks a Marxist framework — not just the industry is changing, but the nature of labor is changing.
📺 Recording of the call: https://youtu.be/YFR1e1bKheg
📆 Next meeting — Monday, May 18, 10am ET / 5pm MSK
📍 Registration on luma: https://luma.com/13cl7bnz
To watch — Lecture 3: Chase Lochmiller (CEO, Crusoe)
🎥 https://www.youtube.com/watch?v=4zk-hJ50vmU
Additional materials:
• A Primer on AI Data Centers — https://www.generativevalue.com/p/a-primer-on-ai-data-centers
• The AI Infrastructure of the Future (McKinsey podcast) — https://www.mckinsey.com/capabilities/tech-and-ai/our-insights/the-ai-infrastructure-of-the-future
Discuss in chat: @rvnikita_blog_chat
