NVIDIA Faces TPU Inference Squeeze

NVIDIA’s data center dominance is facing fresh headwinds: Google has just unveiled its next-gen TPUs, sharpening the battle over AI chip supremacy by splitting training and inference workloads and targeting costs where it hurts. This isn't a clean break—Google’s still offering NVIDIA’s best GPUs and broadening compatibility—but the 8i TPU’s deep integration and memory innovations mean big customers like Anthropic are already shifting, and if other hyperscalers follow suit, NVIDIA could see real pressure on its inference margins.

But here’s the catch: NVIDIA’s China business, once a growth juggernaut, is now mired in regulatory quicksand. Despite the U.S. allowing H200 chips to ship with a 25% tariff, China’s blocking them at customs, accelerating the rise of domestic chipmakers like Huawei and Cambricon. Official channel sales have stalled, gray-market risks are mounting, and NVIDIA’s China market share has dropped from 95% to under 60%. The risk isn’t just lost sales today—it’s a long-term shift as China builds its own supply chain and locks out U.S. chips.

Amid this, NVIDIA is doubling down on software and interconnects, quietly investing in startups like Point2 Technology and rolling out new tools like Nemoclo to secure enterprise AI deployments. With the market already pricing in near-perfection—NVDA shares at record highs, options betting on it becoming the world’s largest company—any earnings wobble could send sentiment south fast. Featuring reporting from Tom’s Hardware and insights on the ground in Seoul.

Powered by Apisod.com