The AI Datacenter 4-Tier Framework — How to Allocate Capital Across a $7T Buildout

Bottom Line

AI infrastructure is not a one-stock GPU bet. It's a build-out across four physical layers — contractors, equipment, cooling, and hidden infrastructure — that every GPU rack depends on. Hyperscalers (Microsoft, Amazon, Google, Meta, Oracle) are spending over $700B on AI infrastructure this year alone, and McKinsey projects $7T in global datacenter spending through 2030. That's about 23% of 2025 US GDP.

Why You Have to Look Beyond the GPU

The first time I read the $7T figure I didn't quite believe it. Then I asked the question that flipped it for me: who physically builds all of that? Before any GPU lights up, somebody has to grade the land, pour the concrete, run the transmission lines, install the switchgear, lay the fiber, build the cooling loop. The chip is the last 1% of the bill of materials.

From my angle, AI infrastructure decomposes cleanly into four layers.

The 4-Tier Framework

Tier 1 — Contractors

The companies that physically build the datacenters and the mechanical, electrical, and plumbing systems inside them. My three picks: Comfort Systems USA (FIX), IES Holdings (IESC), Quanta Services (PWR).

FIX moved datacenter revenue from one-third of total two years ago to over half last quarter, with backlog at a record $122.4B — nearly double year-over-year. A $1,000 invested in spring 2021 would be roughly $22,000 today. A 22-bagger in five years for an HVAC contractor is the kind of stat that still doesn't feel real.

IESC is the smallest of the three at a $12.6B market cap. Revenue compounded 23% annually for five straight years; EPS multiplied nearly 8x. As long as the communications segment keeps printing 35%+, hyperscaler demand is intact.

PWR sits on a $48B backlog — larger than the entire market cap of most other names in this basket combined. Management raised guidance into the $35B revenue range, and a major slice of backlog is locked under long-term master service agreements. Add the AEP partnership tied to a $72B capital plan, and PWR is the credibility anchor of the entire layer.

Tier 2 — Equipment Makers

Switchgear, transformers, power distribution gear — the equipment that goes inside every datacenter regardless of which GPU is in the rack.

Powell Industries (POWL) was a 0%-operating-margin no-name in 2021. Four years later it's at ~20% margins, revenue more than doubled, and the company executed a 3-for-1 split. Latest quarter: orders up 63% YoY, backlog at a record $1.6B.

Eaton (ETN) is the $167B giant. Most recent quarter saw datacenter orders up over 200% YoY and Americas Electrical backlog at a record $13.2B. The Boyd acquisition adds liquid cooling that's expected to generate $1.7B in annual sales, 80%+ from datacenters.

Tier 3 — Cooling

NVIDIA's next-gen AI racks pull 132 kW — nearly 10x a traditional rack. Air cooling cannot handle that. Liquid cooling is the only answer. I broke down the cooling layer separately.

Tier 4 — Hidden Infrastructure

Fiber, copper, behind-the-meter power. One or two steps removed from the GPU and largely absent from analyst models — which is exactly why some of the best risk/reward names live here. I cover Belden, Mueller, and Bloom Energy in a separate post.

If I Were Allocating $100 Today

Tier	Allocation	Why
1. Contractors	$30	Most direct hyperscaler exposure + cleanest balance sheets
2. Equipment	$25	Best capital efficiency + largest backlog visibility
3. Cooling	$20	Highest-growth subsegment of the buildout
4. Hidden	$25	Contrarian setups and value names

The heavier weight goes to the names with the cleanest balance sheets and the most direct hyperscaler exposure. Smaller weights go to the higher-volatility names. Same framework I use for the rest of my portfolio.

What Breaks This Thesis

First, hyperscaler capex deceleration. Backlogs run through 2028, but if AI ROI starts being seriously questioned, new orders dry up first.

Second, power bottlenecks. You can finish a building and still wait years for the grid to interconnect — which is exactly what makes a name like Bloom Energy interesting.

Third, valuation. Several of these have already turned into multi-baggers. My posture is not to wait forever, but to know the names cold and scale in on pullbacks.