# HBM demand — X 热门讨论 (2026-09-22 00:02 UTC)
## @ZTheTrader (⚡️The Traveling Trader⚡️) · 09-21 20:03 · ♥50 ↻2 💬5 I see a lot of folks saying the CPU trade came out of NOWHERE. How is the market just finding out that CPUs are a major AI bottleneck? Let me try to explain.
The market had the ingredients right but the ratio wrong. From 2023-25 the assumption was that inference means the GPU does the work and the CPU just boots the box and feeds it. Training ran about 1 CPU per 8 GPUs. Serving tightened that to maybe 1:3-4.
Then agents showed up and the math completely changed.
An agent spends most of its time outside the model call. Tool calls, headless browsers, sandboxes staying alive after you close the app, retrieval, multi-agent orchestration. The GPU handles one step of that. Almost everything wrapped around it though is CPU work. Some research puts CPU-side tool processing at 50-90% of total latency in an agentic loop.
So the math changed in real time:
Training 1:8, Plain inference 1:4, Agentic now approaching 1:1.
Arm's version of the same number: 30M CPU cores per gigawatt in a normal AI datacenter vs 120M in an agent-era one. A 4x jump in host compute intensity.
Why it stayed a story instead of a trade for so long: custom ARM silicon was supposed to absorb this, but agentic demand got big enough to stress BOTH custom ARM and merchant x86, and even NVIDIA ended up shipping Vera as a standalone CPU.
Intel was psychologically un-investable after years of share loss, so the first reaction to rising CPU demand was "AMD takes it," and not "the entire CPU category is short." And capital followed the binding constraint in order: GPUs, packaging, HBM/DRAM, power. CPU wafers looked fine until those higher margin products already ate the foundry capacity ahead of them.
2026 turned the story into actual, tangible numbers. AMD doubled its server CPU TAM to $120B by 2030 on agentic demand. EPYC sold out through year end, 30+ week lead times. Intel filling only 40% of backlog, 6-8 month waits in parts of Asia.
Nobody disputed the input was needed; the sizing was just misunderstood. https://x.com/ZTheTrader/status/2102126566423863571