OpenAI, a Longtime Nvidia Customer, Plans to Buy Google’s AI Chips - The Information

This isn't a chip swap—it's the first structural crack in CUDA's 15-year monopoly. Technical chain: The real cost isn't hardware procurement; it's the 18-month re-architecture of training pipelines from CUDA kernels to TPU sparse-core scheduling. Once complete, software lock-in is permanently weakened. Upstream, HBM and CoWoS advanced packaging (Taiwan, China capacity) remain a shared bottleneck, but TPU's ICI interconnect imposes different bandwidth-density requirements than NVLink, forcing packaging-line adaptation. Compliance: Under tightening US export controls, single-source dependency has escalated from commercial risk to strategic vulnerability. OpenAI is writing "supply-chain sovereignty" into procurement logic—the same playbook that drove semiconductor multi-sourcing after the 2022 entity-list actions. Market game: The real loser isn't Nvidia—it's AMD. Google pushing TPUs to external customers bypasses AMD's "CUDA-compatible" positioning entirely via vertical integration. Nvidia's likely response: accelerate Triton compiler openness to weaponize software stickiness against hardware substitution. 12–24 month outlook: At least three hyperscalers will run dual-chip architectures. The market shifts from "one superpower" to "duopoly + custom ASICs." CUDA's share won't collapse, but its pricing premium faces its first material compression.