TSMC excludes CPUs from AI accelerator growth forecast

TSMC's decision to exclude CPUs from its AI accelerator growth metric is not an accounting footnote—it is a margin defense play. Agentic AI is fundamentally redefining the CPU's role: from passive instruction execution to active orchestration of inference pipelines. The CPU-GPU boundary is dissolving, yet TSMC deliberately maintains a narrow definition because CPU gross margins sit far below CoWoS-packaged AI ASICs. Including them would dilute the high-growth narrative that underpins investor pricing of its advanced-node premium. Technically, NVIDIA's Grace-Hopper and AMD's MI300A heterogeneous integration prove that AI server compute allocation increasingly depends on CPU-side scheduling. TSMC's exclusion signals that its cutting-edge capacity remains tightly coupled to GPU and ASIC workloads, while CPU foundry at 12nm and 7nm has not entered its core growth engine. On competition, Intel and AMD are aggressively positioning AI CPUs to capture cloud infrastructure budgets. TSMC's framing effectively tells the market: CPUs are a legacy market, not a growth story. Over the next 18 months, as agentic AI inference chains lengthen, CPU share in AI BOMs will climb from roughly 15% toward 25%. At that inflection point, the definitional debate escalates from an accounting question to a geopolitical compliance issue—if US export controls extend to AI-server CPUs, TSMC's narrow scope shifts from financial strategy to existential risk.