Stanford CS153 recap: Google depreciates compute over six years, NVIDIA ships a new arch every year
le_james94 · x · 2026-09-29
Engineer James Le worked through all 13 sessions of Stanford's newly published CS153 "Frontier Systems" course (taught by Anjney Midha and Michael Abbott of AMP PBC) and published long-form notes framing frontier AI as a stack whose binding constraint keeps moving down.
- Opening map: seven layers from capital, land/power/shell, compute infra, models & agents, apps & interfaces, deployment, to safety/policy/governance
- General Matter CEO Scott Nolan: uranium enrichment is the bottleneck to nuclear, which bottlenecks electricity, which bottlenecks AI; the US once held 86% of global enrichment capacity, its last plant closed in 2013
- On Nadella's vision of every company running its own RL environments and private evals: the "easy button" puts that loop on Microsoft's tooling — rent the frontier model, own the loop
- The number that sticks: Google depreciates compute over six years while NVIDIA ships a new architecture every year — either demand keeps old chips busy, or it's a writedown nobody has taken yet.
Related event: Stanford CS153 Course Draws Attention; Nadella's RL Vision Debated(2 posts)→
More from Companies & People
- Factory CEO: 'For AGI, we've already crossed the event horizon' — matanSF · 2026-09-30
- LangChain's LangSmith Roadshow Boston draws packed house as tour continues — LangChain · 2026-09-30
- Every to cover OpenAI DevDay live with vibe checks and a Sam Altman interview — danshipper · 2026-09-30
- ETH Zurich CS enrollment grew 3x as much as mechanical engineering since 2015 — randal_olson · 2026-09-30
- Anthropic inks up to $84.5B compute deal with SpaceX through 2029 — XFreeze · 2026-09-30
- How Google Is Quietly Turning Millions of People Into AI Users — timemagazine · 2026-09-30