US Approves NVIDIA H200 Exports to China; AI Capex Cycle Broadens

US export controls on advanced semiconductors, a cornerstone of the Trump administration's China strategy, showed unexpected flexibility when the Department of Commerce approved sales of NVIDIA's H200 chips to 10 Chinese companies. The decision came during the Trump-Xi summit in Beijing and suggests that US policymakers are willing to compartmentalize restrictions in favor of broader trade negotiations. H200 chips are designed for inference workloads and long-context AI applications, a tier below the most cutting-edge training chips but still highly valued by Chinese tech giants building out cloud infrastructure.

Cisco's restructuring announcement provided a contemporaneous signal of how broad the AI infrastructure buildout has become. CEO Chuck Robbins noted that AI infrastructure demand is now scaling beyond model training into long-context inference capacity, with switches, optics, and networking gear becoming bottleneck points. This narrative expansion matters because it widens the addressable market for semiconductor and infrastructure vendors beyond NVIDIA and AMD into companies like Broadcom (AVGO), with structural tailwinds lasting years rather than quarters. Traders have positioned accordingly, with NVDA gaining 20% in seven days and approaching a $6 trillion market capitalization, while AVGO and ARM Holdings also rally on the broadening demand thesis.

The market is pricing in a multi-year capex cycle that extends well beyond the initial AI data-center buildout. Meta's announced $21 billion agreement with CoreWeave for long-term inference capacity, combined with JPMorgan's bull-case target raise for Taiwan equities (recommending them as the "most pure-play exposure to the global AI buildout"), signals institutional conviction in sustained demand. Investors are positioning in mega-cap tech, chip designers, and infrastructure names under the assumption that AI spending will continue to accelerate through 2026 and 2027.

Risk narratives emphasize that NVIDIA valuation multiples are stretched on trailing profitability, and that a sharp slowdown in enterprise cloud spending or a shift toward inference-optimized chips (which may be less profitable for NVIDIA) could compress margins. The H200 export approval, while positive in headline terms, also suggests US policy is willing to let China access lower-tier advanced chips, potentially accelerating China's domestic AI development and reducing future NVIDIA TAM in the world's largest market. Retail overexposure to NVDA (with some traders joking about 'full porting' into the stock) raises execution risk if sentiment turns.