In late February, after a week of back-to-back roadmap reviews, I found myself standing in the aisle at Micro Center in Cambridge, staring at an NVIDIA RTX 4090 locked behind glass like a Cartier watch. The price tag—still hovering north of $1,700—felt less like consumer electronics and more like a proxy for geopolitics.

I don’t usually write about public markets. My portfolio is mostly boring: VTI, a little VXUS, some ill-timed dabbles in individual names that taught me humility. But this quarter, the conversations changed. Not “Which model are you using?” or “How many H100s did you secure?” Instead: “Can you get enough power?” “How are you cooling it?” “What’s your interconnect story?”

The focal length shifted. The GPUs are still the headliners. But the transitive closure of AI—the second-order effects—is where the real capital is moving.

I keep thinking about something Carlota Perez wrote in Technological Revolutions and Financial Capital: after the installation phase of a new technology, capital rotates from the flashy frontier into the infrastructure that makes it durable. Railroads weren’t the end; they were the forcing function for steel, telegraph, standardized time. The internet wasn’t just Pets.com; it was fiber, data centers, AWS.

Net net: Act 1 of AI was GPUs. Act 2 looks a lot like power grids, cooling systems, and the unsexy precision tools that make semiconductors possible.

1. The Substation Problem

On Monday I took the 7:42 a.m. commuter rail from West Newton into South Station. I’ve ridden that line long enough to know where it stalls—just before Yawkey, when freight traffic gets priority. That pause always reminds me: infrastructure sets the cadence. You can have the fastest train in the world; if the signaling system is old, you wait.

Data centers are hitting the same edge of pain.

A single AI training cluster can draw 50–100 megawatts. Hyperscale campuses are asking utilities for 500 megawatts or more—roughly the load of a mid-sized city. Dominion Energy in Virginia has warned about data center demand reshaping load forecasts. ERCOT in Texas is fielding interconnection requests that look like science fiction.

This is where companies like Eaton (ETN) sit. Circuit breakers, switchgear, transformers—the stuff no one tweets about. I’ve walked through a data hall in Ashburn; the loudest thing isn’t the servers, it’s the hum of power conversion equipment. The boring boxes matter.

If GPUs are Stradivarius violins, then power management is the concert hall. Without acoustics, you just have noise.

I don’t own ETN (yet), but I’ve started reading their earnings transcripts the way I used to read liner notes from Blue Note records—looking for the quiet session musicians. You can feel the order backlog in the language. You can feel utilities renegotiating timelines.

The pattern is familiar from software architecture: once you scale, the constraints move from logic to throughput. From cleverness to physics.

2. Cooling Is the New Compute

Last summer, I visited a co-location facility outside Chicago, near Elk Grove Village. The building looked like a Costco with no windows. Inside: hot aisle containment, CRAC units, a forest of pipes sweating under fluorescent light.

Air cooling is reaching its limits. Liquid cooling—direct-to-chip, immersion—has moved from lab curiosity to procurement checklist. I’ve had more hallway conversations about coolant chemistry than model architecture.

When Jensen Huang holds up a new GPU, the subtext is watts per rack. Heat density. The thermodynamics of ambition.

This is where the ecosystem expands: pumps, valves, specialty gases. MKS Instruments (MKSI) doesn’t make headlines like NVIDIA, but they make subsystems—pressure control, vacuum solutions—that sit inside semiconductor tools. ASM International (ASMI) focuses on atomic layer deposition, the slow, meticulous layering process that enables smaller nodes.

Atomic layer deposition is a lesson in patience. You grow films one atomic layer at a time. It’s the opposite of blitzscaling. It reminds me of practicing scales on a Yamaha U1 upright in my old apartment in Somerville—metronome ticking at 60 BPM, hands learning distance by repetition. The incremental difference compounds.

Semiconductor manufacturing is like that. You don’t leap to 2nm; you approach it asymptotically, one process improvement at a time. The companies supplying that precision feel less glamorous, more convex in the long run.

Cooling is similar. It’s not a single breakthrough; it’s a portfolio of tweaks: better heat exchangers, smarter flow control, redesigned racks. The zeitgeist celebrates the model release. The balance sheet rewards the plumbing.

3. Energy Isn’t Abstract

I grew up thinking of energy as a utility bill my dad grumbled about at the kitchen table. Now it’s a strategic constraint discussed in boardrooms.

If AI clusters need continuous, reliable power, then the question becomes: where does it come from? Natural gas, nuclear, renewables plus storage? Suncor (SU) shows up in some AI-adjacent investor decks not because oil is trendy, but because hydrocarbons still underpin grid reliability.

I’m not ideological about this. I’m pattern-seeking. When load growth inflects, marginal supply matters. The grid wasn’t built for this demand curve.

I think about this when I’m on I-90, passing the Mystic Generating Station, stacks visible against a gray sky. Infrastructure hides in plain sight until you need more of it.

The same dynamic shows up in software teams. When usage spikes, you discover the hidden coupling. The database you ignored. The queue that backs up. The one engineer who understands the legacy system and becomes your most important hire.

Energy is the legacy system of civilization. We’re rediscovering its importance under AI load.

4. The Picks and Shovels, Again

I’ve also watched Palantir (PLTR) ride the AI narrative, repositioning from data integration to AI platform. There’s something instructive there. Not every beneficiary of Act 2 is physical. Some are orchestration layers—software that helps institutions operationalize AI across messy, real-world systems.

It echoes what happened with AWS. The raw compute mattered. But the management layer—the APIs, the services, the abstraction—captured durable value.

The equivalence class here isn’t “AI companies.” It’s “constraint solvers.” Power distribution. Thermal management. Deposition tools. Data orchestration.

If Act 1 was about securing GPUs, Act 2 is about removing bottlenecks. It’s less about raw intelligence and more about throughput.

There’s a humility in that.

I’ve felt a version of this in my own career. Early on, I optimized for visible output—features shipped, talks given, projects launched. Mid-career, I find myself drawn to enabling work: hiring well, documenting decisions, aligning incentives. The unglamorous scaffolding.

It’s less legible on Twitter. It’s more durable in reality.

The Turn

What’s been sitting with me is this: every technological revolution eventually becomes a utilities story.

Electricity. Rail. The internet. Cloud.

The capital rotation isn’t just financial. It’s psychological. We move from fascination to maintenance.

That’s not a downgrade. It’s maturation.

When I zoom out—focal length wide—the $2 trillion figure floating around AI feels less like a bubble number and more like a systems upgrade. Systems upgrades are expensive. They also require coordination across domains that rarely talk: utilities and chip fabs, software engineers and chemical engineers, policymakers and procurement officers.

The music metaphor writes itself. A band can have a virtuosic guitarist. But if the rhythm section drifts, the whole thing collapses. Act 2 is rhythm section energy.

And rhythm sections don’t chase spotlight. They keep time.

Some Principles I’m Carrying

• Follow constraints, not headlines. The bottleneck is often upstream of the hype.
• Infrastructure sets cadence. Whether it’s the MBTA or a data center, throughput beats brilliance.
• Incremental difference compounds. Atomic layer deposition, better cooling loops, tighter power management—small gains at scale.
• Enablement is leverage. The tools and systems that let others perform can be more convex than the performers.
• Maturation beats mania. When capital rotates to plumbing, it’s a sign the technology is settling in.

Grace notes:

• Currently reading: Technological Revolutions and Financial Capital by Carlota Perez, and Chip War by Chris Miller.
• Listening while writing: Bill Evans’ Sunday at the Village Vanguard.

If Act 1 was about securing the violin, Act 2 is about tuning the hall, reinforcing the stage, and making sure the lights stay on.

I’m trying to invest—and work—accordingly.

#AIInfrastructure #EnergyEconomics #SystemsThinking #DataCenterDynamics #PrecisionEngineering #MusicalMetaphors #TechAndCulture #IncrementalInnovation #UtilityScale #CommunityOfPractice