Author: Nita Farahany Source: https://nitafarahany.substack.com/p/why-china-quit-us-chips-inside-my Published: September 16, 2025
A class essay by Nita Farahany, the seventh in a 27-class introductory AI law and policy course, published September 16, 2025. It covers compute fundamentals — what "compute" means, how AI chips work, the chip supply chain, DeepSeek's architectural workarounds, and the energy demands of training and inference. Farahany frames the class around the January 27, 2025 Nvidia stock crash that followed the release of DeepSeek-R1, asking how DeepSeek built a competitive model for roughly $6 million when she states GPT-4 cost more than $100 million.
Summary
The class uses the January 27, 2025 Nvidia stock crash as its anchor. Farahany notes that Nvidia lost about $600 billion in market value that day, which she compares to the GDP of Sweden, after DeepSeek-R1's release raised questions about whether competitive models could be trained with far less compute than US labs assumed necessary. From there the essay builds an explanation of why compute matters, how the hardware works, why the supply chain is concentrated, how DeepSeek economized, and why the energy bill for AI keeps growing.
Key claims
Three meanings of "compute"
Farahany distinguishes three senses of the word: the math (FLOPs, the calculations themselves); the physical hardware (chips); and the full infrastructure stack (data centers, cooling, power). She argues DeepSeek needed all three but lacked access to chips, and so economized on the math.
FLOPs and chip throughput
She illustrates a FLOP as a single calculator key-press equaling one floating-point operation. An Nvidia H100, she states, performs roughly 67 trillion FLOPs per second. At one calculation per second without sleeping, a person would need more than 2 million years to do what an H100 does in one second.
CPUs, GPUs, and TPUs
Farahany frames a CPU as a single senior partner working alone and a GPU as 10,000 first-year associates working in parallel, and argues that AI's problem of selecting among 170,000 possible next words favors the parallel approach of GPUs. She notes that Google's TPUs (7th-generation Ironwood, April 2025) perform only AI math, and that Nvidia controls CUDA, producing software lock-in that extends even to non-Nvidia hardware.
The five-layer chip supply chain
The essay lays out a supply chain in five layers:
- Chip design — Nvidia, which she states spends more than $80 million per chip blueprint.
- CUDA software — described as a Nvidia monopoly.
- Fabrication — TSMC, which she puts at roughly 90% market share for advanced nodes, operating three fabs in three Taiwan science parks about 100 miles from China.
- EUV lithography machines — ASML of the Netherlands, described as a 100% global monopoly; the machines cost about $200 million each, are the size of a school bus, and the Dutch government has banned their export to China.
- Data center infrastructure — facilities that she says take 2–3 years and billions of dollars each to build.
DeepSeek's Mixture of Experts architecture
Farahany describes DeepSeek's Mixture of Experts (MoE) design through an analogy in which only the relevant specialist among six lawyers researches a given question rather than all of them. She states that DeepSeek-V3 uses 256 routed experts per layer, with only 8 activated per token plus 1–2 shared experts, yielding roughly 99% compute savings on routing.
Project Stargate
The class covers Project Stargate, announced January 21, 2025, a $500 billion buildout over four years. Farahany states the Abilene, Texas facility is planned for 2 million chips and 4.5 GW of power — more than the Hoover Dam, and enough for 3.5 million homes. She notes that the US grid needs 5–7 years to expand its connections while the data centers need power by 2027, and points to responses such as Microsoft restarting Three Mile Island and Oracle building small modular nuclear reactors.
The inference trap
Farahany distinguishes training, which is a one-time cost, from inference, which she describes as a recurring one. She notes that Sam Altman said OpenAI loses money on its $200/month Pro subscriptions, and cites a "$1,000 query" example of 10 quadrillion calculations, equal to 300 kWh, or an average home's use for 10 days. She adds that DeepSeek's MoE design does not save inference energy: citing MIT Technology Review, she states that chain-of-thought reasoning generates longer responses and uses more energy per query than comparable non-MoE models for the same task.
The export-controls paradox
The essay closes on what Farahany frames as a paradox: US export controls intended to slow China instead pushed it to innovate, so that DeepSeek exists because of, not despite, the restrictions. She notes that China had instructed its top technology firms to stop buying Nvidia chips entirely.
Provenance
Class 7 of a 27-class introductory course on AI law and policy by Nita Farahany, published on her Substack September 16, 2025. As a class essay it presents Farahany's framings and analogies; its quantitative claims (chip throughput, supply-chain shares, Stargate specifications, energy figures) are presented as the source states them.
Relationships
- part-of: Nita Farahany intro course series (Class 7 of 27)
- related: Nvidia & TSMC — AI Compute Infrastructure, ASML — EUV Lithography Monopoly, DeepSeek, Compute Governance, DeepSeek-R1
- previous: Inside My AI Law & Policy Class 6: Training Data, Discovery Wars, and Who Gets Paid (Farahany, September 2025) next: Inside My AI Law & Policy Class 8: Your Electricity Bill / Compute Governance (Farahany, September 2025)