Memory and power bottlenecks reshape the AI chip race

Nvidia still leads the AI chip market on training performance and software maturity, but supply constraints are increasingly shaping what customers can actually deploy. High bandwidth memory and TSMC’s CoWoS packaging are sold out into 2027, limiting shipments even before export controls, customer allocation and manufacturing yield are considered.

China remains the most contested front. Beijing is reportedly prepared to allow a limited group of firms to buy Nvidia’s H200, but the cap may fall below 200,000 chips, less than half of what buyers requested. Huawei’s Ascend chips have become viable for inference, as shown by DeepSeek’s use of Ascend hardware for serving R2, but the same hardware could not reliably complete a frontier training run.

Intel is pursuing a different opening with Crescent Island, an inference-focused data center GPU built around cheaper LPDDR5X memory rather than HBM. The broader race now depends less on a single benchmark than on memory access, packaging capacity, China policy, chip yields and power availability for data centers.