OpenAI once pushed to keep AI chips from China. Now it's staying quiet. - Politico

OpenAI's silence is more telling than its earlier lobbying. The commercial logic for restricting frontier chip exports has eroded: as inference replaces training as the dominant cost center, and China's AI stack pivots toward distillation and inference-optimized architectures, the "training moat" thesis that justified restrictions no longer holds. Technically, the CoWoS packaging bottleneck is being diluted by chiplet interconnect and 3D-stacking architectures — the single-point chokehold narrative is losing its engineering basis. On compliance, the lobbying vacuum opens a 6-to-12-month enforcement-flexibility window for NVIDIA and AMD, but it simultaneously strips away the scarcity premium that anchored their data-center pricing. The real competitive threat isn't Chinese silicon catching up; it's US firms losing the pricing power that artificial scarcity created. Huawei's Ascend line already serves China's inference workload adequately. Over the next 18 months, expect export controls to fragment from blanket bans into sector-specific, node-specific patchwork rules. The geopolitical "chip war" framing will yield to commercial repricing as the interests of US chipmakers, hyperscalers, and policymakers diverge further.