Meta's Muse lifts server CPU demand, with AMD EPYC orders extending into 2028

Agentic AI is not another GPU story. Persistent multi-turn reasoning, tool-call routing, and cross-model orchestration are fundamentally CPU-bound workloads. The data center BOM is rebalancing: CPU-to-GPU ratios are shifting from 1:8 toward 1:2. AMD EPYC's 128-core density and AVX-512 make it the default orchestration substrate. Orders locked through 2027-2028 signal that hyperscalers now treat CPU as a strategic constraint, not a commodity line item. The competitive counterpunch is three-pronged: Intel's Granite Rapids ships but lacks DC-segment execution credibility; ARM-based silicon (Graviton, Ampere) is eroding cloud-native share; and the quietest threat is NVIDIA bundling Grace CPU with Blackwell GPU, effectively absorbing the orchestration layer into its own stack and suppressing standalone CPU demand. Supply-chain risk is the wildcard. Advanced-node capacity in Taiwan, China is already saturated by GPU and mobile. A simultaneous CPU demand spike risks a 2021-style shortage with 52+ week lead times and minimal buffer stock. Within 18 months, rack-level power budgets will be rewritten as single-socket TDP crosses 400W. Agent-optimized server SKUs will emerge as a distinct category. The GPU monoculture of 2023-2025 is fracturing.