ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
NewsTopic · AI基建瓶颈迁移论4 min read·2026-10-19

AI Investing Writer Rob May: The Bottleneck in AI Infrastructure Is Migrating from Chips to Power, Cooling and Networks

AI investing writer Rob May presents a constraint-migration framework: once one bottleneck is relieved, the economics shift to the next thing that can't keep up. He cites roughly 120 GW of scheduled data center capacity through August 2028 and 24-to-48-month utility interconnection queues as evidence that the binding constraint has shifted from silicon to power, thermal management, networking and edge inference, and offers a three-question test for evaluating AI infrastructure companies. The author also acknowledges that the second-order trade is quickly becoming consensus, and his third-order judgment involves his own commercial interest.

ImportanceMaterialEvidenceE3 inspectableWrite-upDeep

2026-10-19 03:24 GMT+8

Everyone knows AI needs compute — and that itself is the investment problem. When consensus is already fully priced in, you need perfection just to break even. AI investing writer Rob May proposes a framework in the latest issue of Investing in AI: the most durable, highest-margin AI opportunities won't be in foundation models or GPUs, but in the physical and digital infrastructure required to keep those GPUs running.

The idea comes from Howard Marks's second-order thinking. First-order thinking is simple and reactive: AI is growing, so buy the chipmaker. Second-order thinking asks the next question: if millions of AI chips get deployed, what breaks? What has to adapt? What becomes indispensable?

History keeps validating this pattern. In the 1920s, the first-order trade was Ford, GM and the tire makers; but the bigger economic story was everything mass car ownership made possible — the interstate highway system, the suburban real estate boom, motels, drive-through fast food. Nobody looking at a Model T was underwriting McDonald's, but the car was a big reason McDonald's became McDonald's. The mobile internet ran the same play: in 2007 the first-order trade was Apple and the cell tower companies, but always-on, GPS-enabled connectivity is what made Uber (2009) and Instagram (2010) possible. The bottleneck quietly shifted from the handset to cloud storage and mobile data networks.

Sign in / Sign up
ReadingResearchRadarInvestment framework

Rob May calls this pattern constraint migration: relieve one bottleneck and the economics move to the next thing that can't keep up. He cites his own earlier analysis showing that roughly 120 GW of new data center capacity was scheduled for delivery through August 2028, while historically only about half of scheduled capacity lands on time. The reason isn't money or chips — it's transformers, gas turbines and utility interconnection queues that now run 24 to 48 months. Flex CEO Revathi Advaithi said it plainly on their last earnings call: AI is increasingly becoming an infrastructure story, and more specifically, a power story.

Within this framework, he argues the constraint is unfolding across four layers: power — AI data centers draw far more power per rack than traditional cloud facilities, the grid wasn't built for that, forcing a scramble for bespoke generation and heavy upgrades to high-voltage transmission; thermal — next-generation GPU racks throw off more heat than fans and air conditioning can move, forcing a migration to direct-to-chip liquid cooling and immersion cooling, but liquid cooling doesn't eliminate the heat, it just moves it out of the rack to the roof and the water supply; networking — tying 100,000+ GPUs into one coherent supercomputer requires advanced optical transceivers, silicon photonics and ultra-fast switching at a scale nobody has built before, and at this size the network is the computer; edge inference — not every query should travel to a gigawatt campus, latency and cost push a growing share of inference to phones, PCs, cars and industrial machines, and small models make that practical.

He highlights what he calls the boring bull market: legacy industrial and utility-like companies that spent decades as dull, low-multiple businesses suddenly sit on the critical path of the most important technology buildout of our lifetimes. He watches three sectors — HVAC and industrial cooling, power and uranium, fiber and optical components. Private and public markets play this differently: venture bets are in materials and physics breakthroughs like next-gen cooling fluids, SMRs and photonics startups — long-dated and capital-intensive, but where a breakthrough could reset a constraint entirely; public markets focus on established electrical equipment makers whose order books just became strategic.

Rob May also acknowledges that the second-order trade itself is quickly becoming consensus — the power and cooling story is about to become the next crowded lane. He therefore asks the third question: once power and cooling get solved, where does the constraint go next? His bet is utilization and orchestration — making expensive GPU fleets actually busy instead of expensive and idle. This happens to be the domain of Taskrouter, the inference-optimization tool he is involved with, and readers should factor that interest into their assessment of this judgment.

He closes with a test framework: to evaluate any AI infrastructure company, ask three questions — what is the binding constraint today? Does this company own it? Is it positioned for the next one? If a company can't answer yes to at least two of those, it's riding the boom, not capturing it.

Sources

  • 原始出处investinginai.substack.com↗
Topic

AI基建瓶颈迁移论

AI基础设施的核心约束正经历“瓶颈迁移”：随着硅片供应缓解，经济价值与限制因素迅速转向电力、散热、网络互联及推理效率。投资博主 Rob May 提出这一框架，指出解除一个瓶颈后，经济学将转移到下一个跟不上的环节。当前证据显示，尽管约 120…

4 steps1 piecesUpdated 2026-10-19 04:44

How it got here

  1. 2026-08-16Rob May 发布《Is There An Inference Glut Coming?》，分析指出虽然算力供给翻倍，但电网互联排队需 24-48 个月，且推理需求因智能体工作流增加而呈指数级增长，导致电力而非芯片成为主要限制。
  2. 2026-09-01HyperAccel 发表社论《The Next AI Infrastructure Bottleneck Is Memory, Not Compute》，论证在推理时代，内存访问和数据移动效率取代纯计算能力成为关键瓶颈，提出通过架构优化提升有效带宽利用率。
  3. 2026-09-09读者 Mike Verge 撰文深化 Dell 案例分析，验证“约束迁移”理论：Dell 虽销量巨大但因不拥有稀缺组件（芯片/内存）而毛利压缩，指出 Vertiv 等公司正通过收购提前锁定电力与冷却这两个下一阶段的瓶颈。

What is still unsettled

  • 电网互联排队时间是否会因政策干预或小型模块化核反应堆（SMR）的部署而显著缩短？
  • 随着推理负载占比超过训练，内存带宽瓶颈是否会导致现有 HBM 架构被新型近存计算架构快速替代？
  • 边缘推理（手机/PC端小模型）能否真正分流数据中心压力，还是仅作为补充而无法改变集中式供电的根本矛盾？
Read the full topic →
Continue researching →