Nvidia faces a more hostile Washington - Semafor

Washington's escalating hostility toward Nvidia is not a trade dispute—it is the codification of compute sovereignty into legislative action. Bipartisan consensus on dismantling a single chip platform means Nvidia faces structural deconstruction, not marginal export tweaks. On the tech stack, CUDA's moat is being eroded by policy force. If Washington mandates interoperability standards or open-sourced inference layers, Nvidia's hardware-software lock-in dilutes within 18 months. Every percentage point of hyperscaler in-house silicon (Trainium, TPU) share compresses Nvidia's data-center gross margin by roughly 50bps. Upstream, CoWoS advanced packaging in 中国台湾 and HBM supply chains face fragmentation under friend-shoring, stretching delivery cycles beyond the current 52 weeks. Compliance costs are already visible: maintaining A100/A800/H20 product lines for different markets has pushed R&D and certification overhead from ~3% to 5%+ of revenue. If dual-use classification extends to inference GPUs, every product launch becomes a legal negotiation. Competitively, AMD's MI300X and Intel's Gaudi 3 will absorb de-Nvidia procurement. Huawei's Ascend has crossed 40% domestic substitution; further Washington hostility permanently closes Nvidia's incremental market, forcing R&D toward allied markets—precisely what Washington wants. Within 12-24 months, global AI compute fractures into 2-3 incompatible ecosystems. Nvidia's valuation migrates from growth to utility. The real risk is not lost orders—it is CUDA de-standardization. That is the 2001 Microsoft antitrust case, replayed in silicon.