In late February, after a week of budget reviews and one too many 7:12 a.m. Caltrain rides from Mountain View to 4th & King, I found myself rereading a dog‑eared copy of The Power Broker. Not because Robert Caro has anything to say about GPUs, but because he understood a simple pattern: when demand compounds faster than governance, infrastructure becomes destiny.

I don’t usually write about trades. I prefer books and bass lines and the slow accretion of craft. But over the past year, I’ve been logging a different kind of set list in my notebook—utility capex announcements, transformer lead times, backlog disclosures from semi equipment vendors. It reads like a tour schedule for the next five years. Net net: if AI Act 1 was about model releases and valuation multiples, Act 2 is about power, cooling, and the unglamorous tools that make silicon behave.

The high reference here is obvious. Carlota Perez’s Technological Revolutions and Financial Capital lays out the cadence: installation period (frenzy, speculation), then deployment (infrastructure, standardization, boring money). The low detail is that my garage subpanel in Redwood City is already maxed at 100 amps, and PG&E wants a small novella’s worth of paperwork for an upgrade. Systems scale only when copper does.

1. Power: The Grid Is the GPU’s Shadow

Last October, I drove up I‑5 to see a friend in Quincy who works on substation design. He walked me through a yard full of transformers the size of studio apartments. Lead times, he said, have stretched from 12 months to 24–36 for certain high‑voltage units. Steel cores, specialized labor, shipping bottlenecks. It sounded like trying to source a vintage Neve 1073 preamp—except the preamp powers a city.

AI data centers are now quoting 100–300 MW campuses as table stakes. Dominion Energy, NextEra, Duke—pick your utility—are fielding requests that look like small towns arriving all at once. The forcing function isn’t hype; it’s training clusters that behave like industrial loads. Jensen Huang talks about “AI factories.” Factories consume power in convex bursts: build once, run hard.

I track three buckets here:

• Utilities with regulated returns and visible rate base growth (NextEra Energy, Dominion, Duke). Boring, steady, politically negotiated.
• Grid components and electrical equipment (Eaton, Schneider Electric, ABB). Switchgear is the new picks‑and‑shovels.
• Generation with optionality—gas turbines (GE Vernova), nuclear life extensions (Constellation), and the slow burn of renewables plus storage.

The pattern recognition is simple: you can’t fine‑tune a model if you can’t fine‑tune a grid. Power is the transitive closure of AI demand.

On my commute, I’ve started noticing substations the way I used to notice record stores—anonymous, fenced, essential. They are the rhythm section. You don’t clap for them, but the song collapses without them.

2. Cooling: Thermodynamics Always Wins

Two summers ago, I toured a colocation facility in Santa Clara. The most memorable room wasn’t the white‑hot rack rows; it was the mechanical plant—CRAH units, chillers, a maze of pipes. The engineer joked that data centers are just HVAC companies that happen to store bits.

As rack densities push past 40–60 kW and toward 100 kW with liquid cooling, the focal length shifts from chips to thermodynamics. Vertiv and Schneider aren’t glamorous brands in the way NVIDIA is, but they sit at the edge of pain where heat meets uptime. If you’ve ever had a MacBook throttle during a Logic Pro session, you understand the micro version of the problem. Multiply by 10,000.

Here the portfolio spans:

• Thermal management and liquid cooling specialists (Vertiv, Schneider Electric).
• Industrial chillers and heat exchangers (Trane Technologies).
• REITs and operators adapting facilities (Equinix, Digital Realty) who must retrofit at scale.

There’s a music analogy I can’t shake. In jazz, the drummer’s ride cymbal sets the cadence while the rest of the band improvises. Cooling is that ride cymbal—steady, relentless. Ignore it and the soloist falls apart.

The incremental difference between 1.2 and 1.1 PUE doesn’t trend on X, but over a 20‑year asset life it compounds like Buffett’s favorite aphorism. Small thermodynamic wins are capital allocation wins.

3. Semi Tools: The Quiet Kings of the Fab

If power is rhythm and cooling is tempo, semi equipment is harmony—the structure that makes the melody possible. I spent a Saturday in San Jose last month at a used bookstore on Stevens Creek, and I picked up a battered biography of Andy Grove. His paranoia wasn’t theatrical; it was operational. Control the process.

ASML, Applied Materials, Lam Research, KLA. The names read like a syllabus for how modernity is etched. Extreme ultraviolet (EUV) lithography is no longer a curiosity; it’s table stakes at advanced nodes. High‑NA EUV looms. Deposition, etch, metrology—these are not equivalence classes; each step has its own bottleneck, its own moat.

What I like about semi tools in a 2026–2030 frame:

• Backlog visibility tied to foundry capex (TSMC, Samsung, Intel’s CHIPS‑fueled buildout).
• High switching costs—once a process is qualified, you don’t casually swap vendors.
• Geopolitical forcing functions that encourage regional fabs (Arizona, Dresden), which in turn require duplicate tool sets.

This is the less romantic part of AI. No keynote glow. Just vacuum chambers and service contracts. But practice beats performance. The fabs are where the practice happens—thousands of incremental differences that add up to yield.

The Turn: From Model to Molecule

What changed for me wasn’t a new earnings call or a flashy demo. It was noticing my own attention. In 2023, I refreshed arXiv and watched model benchmarks. In 2025, I refresh utility commission filings and transformer order books. The zeitgeist shifted from software abstraction to physical constraint.

There’s a humility here. We can debate open vs. closed models, fine‑tuning strategies, the right data flywheel. None of it matters without electrons, chilled water, and photolithography at nanometer precision. AI is not weightless. It is industrial.

On the train last week, somewhere between Millbrae and San Bruno, I felt a clean line connect Caro’s bridges to Huang’s GPUs. Infrastructure is moral gravity. It decides who gets access, at what price, and how quickly. It is also slow, regulated, and capital intensive. That’s precisely why it can be durable.

I don’t think this is a trade for adrenaline. It’s a portfolio for patience. The convexity is lower than a pre‑IPO model company, but the drawdowns may be kinder. And if Perez is right about installation and deployment phases, we are early in the boring buildout that funds the next decade.

Lessons I’m Carrying (2026–2030)

• Follow the bottleneck, not the buzz. When demand outruns infrastructure, the constraint holders accrue pricing power.
• Boring can be convex. Regulated utilities with visible capex can compound in ways that don’t require heroics.
• Thermodynamics is undefeated. Cooling and power are not accessories; they are first principles.
• Process beats narrative. Semi tools win through qualification cycles and service depth, not headlines.
• Zoom the focal length. Models are quarterly. Grids and fabs are decadal.

Grace notes:

• Currently reading: Robert Caro, The Power Broker (again); Carlota Perez, Technological Revolutions and Financial Capital.
• Listening while writing: Brian Eno, Ambient 1: Music for Airports.
• Watching: CPU and transformer lead times more than model leaderboards.

If there’s a community tilt here, it’s simple: support the people who build the substrate. The electricians, the mechanical engineers, the field service techs in cleanroom bunny suits at 2 a.m. Show up for the infrastructure. It’s less glamorous than a demo day. It’s also what makes the future run.

#InfrastructureMatters #TechAndCulture #PowerAndCooling #MusicInTech #SystemsThinking #SemiConductorInsights #CommunityOfBuilders #ThermodynamicsWins #BoringCanBeBeautiful #LongTermInvesting