Meta Muse Drives Server CPU Demand, AMD EPYC 2027 Capacity Sold Out
nashnova research
Meta's personal AI agent Muse hit 700,000 users in ten days, stress-testing agentic AI's appetite for CPUs — AMD's latest EPYC processors are sold out for 2027, and server CPU lead times have stretched to 25–30 weeks.
How much compute does Muse actually consume?
Meta assigns each Muse user a dedicated virtual machine: 2 vCPUs, 8 GB RAM, 100 GB SSD.
Altimeter partner Freda Duan modeled a scenario of 100 million DAUs at two hours per day: average load requires roughly 8.33 million VMs, peak load (2.5×) about 20.8 million, translating to some 12.5 million physical CPU cores.
This means → a single consumer-grade AI agent could absorb CPU resources equivalent to a hyperscale data center.
Why does agentic AI suddenly make CPUs critical?
Traditional LLMs concentrate compute on GPUs; CPUs handle only orchestration. Agentic AI — AI that autonomously plans tasks and calls external tools — adds heavy task planning, tool invocation, sandbox execution, vector-database retrieval, and I/O coordination, all CPU-bound.
TrendForce notes that in agentic workloads, CPU-side orchestration accounts for 50%–90% of total latency; the CPU-to-GPU ratio may shift from roughly 1:8 toward 1:1 or lower.
Arm estimates traditional cloud architecture needs about 30 million CPU cores per gigawatt of data-center capacity; agentic AI could push that to 120 million — roughly a fourfold increase.
In plain terms = GPUs used to be the star and CPUs the stagehand. Agentic AI has massively increased the CPU's workload, leveling the two roles.
How large could the CPU market become?
BOCOM International cites AMD and Intel executives: the market expects the CPU-to-accelerator ratio to move from 1:8 toward 1:1.
Using 2028 accelerator shipment forecasts as a base, the firm estimates that at a 1:2 ratio, roughly 50.47 million accelerators would require over 25 million data-center CPUs.
At $5,000 per unit, the data-center CPU market would reach approximately $125 billion.
This means → CPUs alone could underpin a hundred-billion-dollar market — and 1:2 is the conservative scenario.
Why is AMD EPYC the most direct beneficiary?
AMD's new EPYC 9006 series (codenamed Venice) is built on Zen 6 architecture and TSMC's 2 nm process — AMD's first data-center microarchitecture designed from the ground up for AI workloads, with PCIe Gen6 support and memory bandwidth up to 1.6 TB/s.
Industry channel sources report EPYC 9006 2027 capacity is fully sold out; AMD has begun taking 2028 orders.
Meta deployed millions of EPYC 9006 CPUs before Muse's launch — making it both the largest customer and a key driver of the capacity squeeze.
What is the ramp schedule and revenue outlook?
The EPYC 9006 series will roll out in four variants: SP7 (Q4 2026), SP8 (H1 2027), Venice-X and Verano (H2 2027).
Morgan Stanley forecasts 2027 shipments of roughly 6.75 million units, up sharply from about 1.25 million in 2026.
At an average price of $7,691 across 31 SP7 and SP8 SKUs, that implies revenue of approximately $51 billion.
This reflects a CPU demand cycle that is not a forward-looking narrative — sold-out capacity and stretched lead times are already here.
市场有风险，内容仅供研究参考，不构成投资建议。