
Sign up to save your podcasts
Or


In January 2025, DeepSeek R1 — a reasoning model from a Chinese quant hedge fund spin-off — topped global app store charts, rattled US tech stocks, and was immediately labelled China's "AI Sputnik moment." This episode interrogates that framing from both sides. The steel-man of the sceptical case is genuine: DeepSeek trained on American chips, the six-million-dollar cost figure is a single training run, and open-weight release is not a market position. But the counter-narrative — that export controls are holding and the gap is intact — is also insufficient. DeepSeek's Mixture-of-Experts architecture emerged in part from hardware constraint, and may be more durable under future compute restriction than the brute-force approach it competed against. Three markers to watch: whether the efficiency advantage holds across model generations, the hardware trajectory as Chinese labs respond to chip restrictions, and whether the hedge-fund-backed research-first organisational model gets replicated.
REFERENCES
By The China MemoIn January 2025, DeepSeek R1 — a reasoning model from a Chinese quant hedge fund spin-off — topped global app store charts, rattled US tech stocks, and was immediately labelled China's "AI Sputnik moment." This episode interrogates that framing from both sides. The steel-man of the sceptical case is genuine: DeepSeek trained on American chips, the six-million-dollar cost figure is a single training run, and open-weight release is not a market position. But the counter-narrative — that export controls are holding and the gap is intact — is also insufficient. DeepSeek's Mixture-of-Experts architecture emerged in part from hardware constraint, and may be more durable under future compute restriction than the brute-force approach it competed against. Three markers to watch: whether the efficiency advantage holds across model generations, the hardware trajectory as Chinese labs respond to chip restrictions, and whether the hedge-fund-backed research-first organisational model gets replicated.
REFERENCES