The memory that made Nvidia untouchable is scarce, expensive, and sold out to 2027. Qualcomm, Intel, and Google are each betting they can route around it. Qualcomm packs 768GB of cheap mobile memory onto a single inference card, Intel builds the first data-centre GPU that skips HBM altogether, and Google turns dead Pixels into a server. Three bets, one target: the memory tax that underwrites Nvidia's lead.

Nvidia and Cerebras are redesigning AI chips to use less high-bandwidth memory, threatening the long-term growth of a memory chip boom driven by AI demand.

Makers of power semiconductors, including Germany’s Infineon, US-based Texas Instruments and China’s Yangjie Technology, have raised prices.