Comfort Systems USA (FIX): The AI Bottleneck Is Shifting to Skilled Labor

Claremont Street Equity Research · AI Infrastructure / Data Center Construction

He who controls the bottleneck shall control the market. The AI buildout has rewarded one kind of company more than any other: the supplier everyone has to go through. In 2025 it was memory and transformers. We think the next bottleneck is already visible — skilled construction labor — and Comfort Systems USA is the company standing in front of it.

1. The thesis: own the bottleneck

967% and 918%. Those are the returns earned by investors who bought SK Hynix and Hyosung Heavy Industries at the start of 2025. Both companies sit in segments of the AI value chain where supply is structurally constrained, and the market has consistently paid up for that position.

We define a true bottleneck by three conditions:

1. Without the product, the value chain stops. GPUs don't ship without HBM memory; data centers don't run without ultra-high-voltage transformers.
2. Demand structurally exceeds supply — not a spike, a sustained imbalance.
3. Supply takes years to expand. HBM capacity takes 2–3 years to add; transformers take 3–5. Long lead times mean the supplier keeps pricing power.

When all three hold, price increases become inevitable. HBM3E prices for 2026 delivery were raised 20% year over year; transformer prices rose 20% in 2025. And because price and volume rise together, EPS grows explosively — which is what separates durable stock appreciation from AI-narrative stocks that correct on any headline. This is the same logic behind how we think about moats: the durable advantage is the one competitors physically cannot replicate quickly.

Comfort Systems USA satisfies all three conditions. Management put it plainly on a recent earnings call: "In 40 years in this industry, I have never experienced a situation like this."

2. Data center construction 101: why MEP is the chokepoint

A data center goes through site selection, permitting, construction, and commissioning — typically 3–6 years from groundbreaking to operation, and 7+ years for projects stuck in grid-interconnection queues. Once construction begins, the work splits into site preparation, shell construction, and MEP: the mechanical, electrical, and plumbing systems that make the building function.

MEP is where the money is. In a typical commercial building, MEP runs about 20% of construction cost. In a data center it's roughly 60–70%, and once facilities transition to liquid cooling it can exceed 80%. Cooling plants, uninterrupted power delivery, and chiller water systems aren't assembly work — they're high-precision installations where a single second of power interruption or a single coolant leak costs hyperscalers millions.

MEP is also fundamentally a human business. Becoming a skilled worker takes roughly six years of education and apprenticeship, licenses are issued state by state, and the work resists automation because every site is different. That is the setup for a labor bottleneck — and the moat for whoever already employs the labor.

3. The bottleneck is shifting from power to people

Until now, the construction labor shortage was masked by the grid bottleneck: projects couldn't break ground because they couldn't get power. That constraint is easing. Hyperscalers have gone all-in on on-site generation — gas turbines, fuel cells, even aircraft and marine engines — and capacity is scaling fast. Bloom Energy plans to double capacity by end-2026; Mitsubishi by 2027. We estimate roughly 60GW of grid-delayed projects from 2027–2030 against ~49GW of new on-site generation supply over the same window.

The moment those delayed projects re-enter the pipeline, the binding constraint becomes construction execution capacity. And the labor math is unforgiving:

• The U.S. construction industry is short roughly 439,000 workers as of end-2025, projected to reach ~499,000 in 2026. 45% of data center contractors reported labor-driven delays in 2025.
• 32% of data center engineering personnel are over 60; only 16% are under 30. About 23,000 experienced workers retire every year.
• Workers can't relocate to demand centers (62% of U.S. data centers sit in just ten states), foreign workers can't transfer licenses, and new entrants need 6+ years to reach deployable skill. Even Google's 30,000-apprentice training initiative won't put workers on data center sites before 2031.
• Data centers, power infrastructure, and manufacturing plants all compete for the same electricians, plumbers, and HVAC technicians across 39 job categories.

The transformer industry already showed how this movie ends: demand up 119% from 2019–2025, supply capped by skilled labor, and prices up 77%. Data center MEP is a harsher version of the same setup — transformers could at least be imported; on-site construction labor cannot.

4. Why Comfort Systems USA wins the labor era

Houston-based Comfort Systems operates ~45 subsidiaries and 170+ locations, concentrated in the eastern U.S. and Texas — exactly where data centers are being built. It is one of the few contractors that can deliver turn-key MEP (mechanical, electrical, and plumbing as one integrated package), which matters because coordinating separate specialty firms creates idle time, and a one-month delay on a 60MW project can cost roughly $1.42 billion in lost revenue. Its 1Q26 backlog grew 80% year over year.

Three advantages make it the top pick in a labor-constrained market:

Modular construction. Instead of building MEP systems on-site, Comfort Systems pre-builds complete systems in factories, cutting time-to-completion roughly in half and on-site staffing needs by up to 80% — the single highest-leverage answer to scarce labor. Modular revenue already runs at 18% of the company, capacity is expanding from 3.0M to 4.0M sq ft by end-2026, and the floor space is committed to its two largest hyperscaler customers, who order directly rather than through general contractors. Its closest competitor, EMCOR, has acknowledged on its own conference call that it has no modular construction experience.

Workforce retention. In a business where the workers are the moat, Comfort Systems keeps them: average tenure of ~6 years versus EMCOR's ~4.6 and the national average of 4.2, supported by a 4-year paid apprenticeship, an internal university with 1,000+ courses, and a non-union merit shop model with no labor disputes since 2002.

Aggressive reinvestment. The company has completed 11 acquisitions since 2022 — buying regional footholds, skilled crews, and modular technology (TAS Energy, Summit Industrial) — and guides 2026 CapEx of ~$1.5 bn, far above EMCOR's sub-1%-of-revenue posture, with management signaling 5% of revenue going forward. It has generated positive free cash flow in every year but one over the past five, funding all of this without leaning on debt.

5. The numbers

We model revenue from the company's labor capacity rather than from end-market demand alone — in this market, the constraint on revenue is how many skilled workers you can field, not how many orders you can win. Technology-segment revenue (data centers) grows from 45% of the mix in 2025 to ~80% by 2028.

Estimates summary. U.S. dollars in thousands. E = estimate.

The margin expansion is the operating-leverage story: when orders exceed capacity, the company selectively accepts only the highest-margin projects, and modular construction squeezes more revenue out of every worker. 1Q26 already printed 56.5% revenue growth and 121.3% EPS growth year over year.

6. What it's worth

We value FIX on its own trading history rather than against peers — power-bottleneck names monetize their leverage differently, and other construction firms lack comparable data center exposure. Applying a 36.23x target P/E (its 3-month average during 4Q25, the quarter data center orders inflected and backlog first crossed $6 bn) to 2027E EPS of $147 gives a target price of $5,326 — 190% upside from $1,835. Against 2028E net income growth of 49.6%, that multiple implies a PEG of 0.73.

7. Key risks to watch

• The bottleneck thesis depends on power easing. If on-site generation scales slower than expected, delayed projects stay delayed and the labor constraint never binds.
• Hyperscaler CapEx is the demand engine. A pullback in AI infrastructure spending — a DeepSeek-style efficiency scare, tariffs, or a capex digestion year — hits backlog growth directly, as the early-2025 correction showed.
• The estimates are aggressive. We model revenue more than tripling and net margins doubling by 2028. Execution shortfalls in modular capacity expansion or labor recruitment would compress both.
• Multiple risk. A 36x earnings multiple on a construction company prices in sustained hypergrowth; any growth deceleration likely de-rates the stock faster than earnings fall.
• Concentration. Modular expansion is anchored by two hyperscaler customers. A change in either relationship would matter.

---

Claremont Street research is for informational and educational purposes only. It is not investment advice or a recommendation to buy or sell any security. Figures are estimates drawn from company filings and our own modeling and may be incorrect. Do your own research before investing. Prices and figures as of May 23, 2026.