AI
Compute is no longer the scarce input; power and memory are
Anthropic's $3.4 billion generator purchase, TSMC's study of 3D DRAM at Nvidia's request and a Senate bill that would make data centres pay for grid upgrades describe one bottleneck from three directions. Power equipment and advanced memory are both multi-year lead-time items, and buying them outright moves capital from operating expense onto the balance sheet. Nvidia's argument that the memory wall needs a whole-stack fix points the same way. For 2027 planning, the constraint that decides schedules may be interconnection queues and DRAM allocation rather than accelerators.