When AI workloads shift from training to agents, the most valuable chip in the data center may no longer be the GPU.
Bank of America Securities just raised its 2030 server CPU total addressable market (TAM) forecast to more than $210 billion, up from roughly $170 billion, and lifted the expected CAGR from 30% to 36%. The core logic: agentic AI is turning the CPU from a GPU “sidekick” into the control plane of AI inference. The CPU-to-GPU ratio, BofA argues, will tighten from about 1:4 in the training era to roughly 1:2 in the inference phase, and approach 1:1 as agentic AI matures.
That single ratio is a structural signal for the whole compute supply chain — and it pairs with the software-side story that agent orchestration (the “harness”) is becoming as important as the model itself, a theme we covered in our analysis of MiniMax’s Agent strategy.
Why agents are CPU-hungry
Training was a GPU story. CPUs handled data preprocessing, tokenization, batching and feeding accelerators — support work. The numbers tell the tale: from 2022 to 2025, AI accelerators grew at a 139% CAGR while server CPUs grew just 14%. By 2025, accelerators consumed 85% of data center compute spending, CPUs only 15% (versus 26% in 2021).
Agentic workloads invert that pattern. An agent is not a single question-and-answer round trip; it is a multi-step loop: planning, context retrieval, tool calls, state management, API interactions, model routing and result evaluation. These are sequential, latency-sensitive, CPU-class operations — not the parallel matrix math GPUs excel at. The bottleneck therefore widens from raw GPU FLOPs to the broader infrastructure: orchestration, memory management, tool execution and data movement.
BofA is explicit that this is not CPU replacing GPU — demand expands on both sides. GPUs keep the core role in matrix operations and heavy inference; CPUs take on orchestration. Because the bottleneck broadens, the total data center TAM grows with it.
What the 2030 map looks like
- Market structure: server CPU TAM splits into roughly $30B traditional/IaaS, $90B AI compute/head nodes and $90B agent-specific nodes — about 86% of the market is AI-driven.
- ARM is the big winner: commercial ARM CPUs (NVIDIA Vera, ARM’s own designs, Qualcomm) take ~38% of value share, plus ~9% for custom ARM (AWS Graviton, Google Axion, Microsoft Cobalt) — nearly half combined.
- AMD is BofA’s top pick: Turin’s 5.0GHz peak frequency, and Venice (2H 2026) with up to 256 cores/512 threads. BofA estimates the EPYC 9965 (Turin, 192 cores) delivers 2.37x the rack-level performance of NVIDIA Vera on agentic workloads — Venice could stretch that to 3.30x. Target price $620.
- Intel is the pressure point: value share falls from ~34% to ~22% by 2030 (unit share still leads at ~36%), with the 18A-P-based Coral Rapids platform its 2028 hope.
The hardware echo of the software story
This is the flip side of the harness debate. Just as model companies are realizing agents need an orchestration layer around the model, the industry is realizing the data center needs a control plane. If agent runtime cost and latency are dominated by orchestration and memory movement, then CPU core count, memory bandwidth and unified memory matter as much as TOPS and FLOPs.
The same logic extends to the edge: this is why on-device agent chips are being designed around CPU-NPU co-design and unified memory rather than raw peak TOPS. And with open-weight frontier models like Qwen3.8 becoming deployable in more places, the frontier is moving toward distribution — which only raises the value of orchestration-class compute.
What to do with this
- If you deploy agents at scale: budget for CPU and memory, not just GPU. Profile your agent loops — you may find you are orchestration-bound, and the fix is cheaper than another accelerator.
- If you buy infrastructure: watch AMD Venice and NVIDIA Vera launches in 2H 2026. The head-node CPU is becoming a first-class purchase decision, not an afterthought.
- If you build edge or device agents: apply the same reasoning at the edge — CPU-NPU co-design beats a bigger TOPS number.
