Loading...
In January 2025, DeepSeek released R1 — a reasoning model that matched GPT-4-class benchmarks while reporting training costs an order of magnitude below US frontier-lab norms — and triggered a roughly $600 billion single-day drop in Nvidia’s market capitalization. The popular framing names a Chinese startup outsmarting Silicon Valley with a small training budget; the structural framing is that frontier capabilities leak through distillation, algorithmic efficiency gains, and open-weight releases, and the GPU-compute moat narrative had absorbed assumptions about the durability of capability advantages that the open-weights ecosystem was already eroding. The Jevons paradox cuts the other way too: cheaper inference does not mean less demand for compute — it means more deployment of more mo...
Popular framing: A Chinese startup outsmarted Silicon Valley with a $5M training budget.
Structural analysis: Frontier capabilities leak — distillation, algorithmic gains, and open weights mean the compute moat is shorter-lived than infrastructure narratives suggest. Cheaper AI does not mean less AI demand; it means more deployment.
Naming the startup protects the moat narrative. The structural framing — capability overhang, Jevons paradox on inference, and substitutable inputs — points to interventions at the seams of export-control design, model-release norms, and value-chain repricing. The same shape will recur with the next open-weights leap.
