Agent Era Reshapes Compute Bottlenecks
coinfanking · reddit · 2026-07-11
This article argues that as AI enters the agent era, the hardware bottleneck shifts from GPU-centric training logic to CPU, memory, and network for running agents.
Key Judgment
- Training was primarily a GPU story; an agent is more like "one task, one set of computation, memory, and communication," thus CPU demand resurges.
- The author expects hardware spending in the agentic era over the next two to three years to be roughly 3x that of the training era.
CPU: The Underestimated Inflection Point
- The article estimates the CPU market could grow from $350-400 billion today to over $2000 billion by 2030.
- Citing Cloudflare: serving 100 million knowledge workers would require about 10 million CPUs; globally, it could be close to 1 billion CPUs.
- The conclusion: 2025 may be the year the CPU demand inflection point truly emerges.
Memory: From Narrative to Financial Reports
- The article highlights Micron's quarterly performance as proof that memory is no longer just narrative but real price and profit.
- Adjusted gross margin is nearly 80%, compared to the historical range of 30%–50%.
- The author states storage prices have risen roughly 7x from the cycle bottom, directing profits to Micron while squeezing budgets of large customers like NVIDIA, Alphabet, and Microsoft.
Structural Changes
- Agents require not only HBM but also standard DRAM and NAND, adding a new demand curve.
- Meanwhile, HBM production consumes significant traditional DRAM capacity, further tightening supply.
- Micron's CEO believes HBM supply shows no clear signs of recovery in the short term; the author projects tightness could extend through 2028.
Contract and Investment Perspectives
- The article mentions that Micron has signed 16 strategic customer agreements, mostly fixed-price, about 3-year terms, totaling at least $100 billion, extending to 2030.
- The author still believes it's better to invest in the equipment chain (e.g., KLA, Lam Research, Applied Materials) rather than directly betting on storage commodities.
Additional Tail Risk/Opportunity
- The end mentions that humanoid robots may require roughly 10x the memory of today's AI systems, potentially extending this cycle into the next decade.
More from Infra
- Anthropic S-1 filing reveals 85% H1 gross margin and 5,300 employees — Jsevillamol · 2026-10-26
- HBM demand shifts: low-stack inference plans shelved, Rubin to go 12-high only — BenBajarin · 2026-10-26
- Anthropic generated $2.08 revenue per $1 of compute spend in Q2 2026, up from $0.63 — cedric_chee · 2026-10-26
- CircuitHub unveils The Grid, a factory-scale robot making PCBs in the US twice as fast as Shenzhen — garrytan · 2026-10-26
- Ubicloud publishes "Fixing Tool-Call Accuracy in vLLM" — petrusenko_max · 2026-10-26
- Meta approves nonvoting Class C shares as 2026 capex guidance hits $135B–$145B — thione · 2026-10-26