AI Infrastructure News · June 25, 2026
First, custom inference silicon is mounting a credible challenge to Nvidia's GPU monopoly—OpenAI and Broadcom's Jalapeño, Qualcomm's new CPU-accelerator stack ($3.9B Modular acquisition), SambaNova ($10B), and Groq ($650M raise) all signal an architectural pivot away from incumbency. Second, power has become the binding constraint on hyperscaler growth, not compute: Tesla and Sunrun's 16 GW virtual power plant, Microsoft's GE Vernova turbine deals, and a 300% surge in gas turbine pricing over three years expose dispatchable energy as the real bottleneck. Third, China is structurally decoupling from US compute architectures: LineShine's CPU-only 2.198-exaflop supercomputer claimed TOP500 dominance without a single GPU, while Trump's 25% H200 tariff prompted Beijing to block Nvidia purchases outright, costing Nvidia ~$30B and accelerating a shift away from US-dependent silicon.
16 GW power plant · 300% turbine price surge · 2.198 exaflops China · $30B Nvidia loss · $50B MGX fund