Source: Tomasz Tunguz
DeepSeek's cost-efficient inference model and SpaceX's Starshield deployment are shifting value capture in AI compute away from training-heavy incumbents toward operators who can optimize inference at scale. Inference, not training, is where most AI workloads run and where sustainable margins sit; companies that cannot compete on cost per token will see their installed base become stranded. The result: enterprise AI adoption accelerates where inference is cheap, developer platforms consolidate around whoever wins the inference cost war, and GPU suppliers face pressure to diversify beyond NVIDIA's premium-positioned chips.