The AI Cooling Trade: Vertiv vs Eaton vs Trane vs Carrier
Same physical constraint, four different economic models. The cooling layer is still flat-to-down 20 days into the AI capex confirmation. Here is the company-by-company decomposition — the day before Marvell prints and two days before Dell.
Cooling names with bullish insider signal today
1 of 4
▲ +1 CARR — the only cooling name with insider buying while Vertiv, Eaton, and Trane all flagged bearish
Memory ran +43% in 20 days. Servers ran +37%. Networking ran +24%. The cooling layer — the layer that physically must remove the 132 kilowatts of heat each Blackwell GB200 NVL72 rack throws — sat at +1.6% on Vertiv, -6.1% on Eaton, -7.2% on Trane, and +2.0% on Carrier. All four solve the same physical problem. They are not the same trade. Vertiv is the AI pure-play with a 50%+ data-center revenue mix. Eaton owns the electrical busway. Trane sells commercial HVAC with a small data-center sleeve. Carrier is the smallest data-center exposure of the four — and the only one of the four where insiders are bullish this week. Marvell prints tomorrow after close and Dell prints Thursday after close. The 48-hour window decides whether the cooling layer reprices in June or whether it waits for Q3.
The contradiction
Air-conditioning stocks were the consensus answer to AI heat all month. The headline thesis was clean: every watt of AI compute becomes heat, so the companies that move heat must reprice. The headline ignored the product-mix question. Vertiv is structurally 50%+ AI-data-center pure-play and trades like a data-center stock. Eaton sells electrical busway, switchgear, and commercial buildings — the AI-data-center share of revenue is closer to 25%. Trane Technologies is commercial HVAC plus transport refrigeration with a data-center sleeve closer to 12%. Carrier is the smallest AI-data-center exposure of the four at roughly 8%. The same Nvidia Q1 print pushes Vertiv’s estimates and barely touches Carrier’s. That is why the 20-day returns scatter from +2.0% to −7.2% and why the scanner’s composite scores stay tightly clustered between 40 and 49 — the market is already pricing four different economic models, not one cooling trade. The trade is to decompose, not aggregate.
What the headline says
The AI cooling trade is one trade
4 names, same physical constraint — they should rerate together
What the data says
The AI cooling trade is four trades wearing one label
Vertiv +1.6%, Eaton −6.1%, Trane −7.2%, Carrier +2.0% over 20 days — the dispersion is the alpha
Chapter 01
What Changed for the Cooling Trade This Week
Five dated, sourced events bracket the cooling trade: Nvidia’s Q1 confirmation, two upcoming earnings prints in the next 48 hours, PCE on Thursday, and a 5-day price pattern that suggests the cooling-layer drawdown may already be bottoming.
Five events define the cooling-trade setup as of today. First, Nvidia confirmed AI demand last Wednesday: Q1 FY27 revenue of $81.6 billion, Data Center segment $75.2 billion (+92% year over year), Data Center networking $14.8 billion (+199% year over year). That is the cleanest possible confirmation that the buildout is accelerating, not pausing — the upstream signal every cooling name now needs to translate into orders. Second, Marvell Technology reports Q1 FY27 tomorrow, Wednesday May 27 after the close, at approximately 1:45 PM Pacific. Marvell sells the custom-silicon ASICs and the optical DSPs that go into the same racks that Vertiv and Eaton then cool and power. A strong networking guide from Marvell tomorrow is the cleanest leading indicator that the cooling order book accelerates through summer. Third, Dell Technologies reports Q1 FY27 Thursday May 28 after the close. Dell’s AI server backlog disclosure is the second leading indicator for cooling — every Dell PowerEdge XE9680 rack ordered is roughly 30 kilowatts of cooling demand twelve to eighteen months later. Fourth, on Thursday the Bureau of Economic Analysis releases Personal Income and Outlays / PCE for April. PCE matters because rates set the discount rate on long-duration infrastructure capex — a hotter print pushes back the cooling-trade timeline by tightening conditions; a cooler print compresses it. Fifth, the cooling-layer price action itself: Vertiv ran −3.6% on the 5-day, Eaton was actually +2.5% on the 5-day reversing a prior bleed, Trane was −2.0%, and Carrier was −2.1%. After two weeks of -8% to -10% drawdowns into Memorial Day, the 5-day window shows the bleed has at minimum slowed. The next 48 hours of catalysts decide whether the slowdown is the bottom or the eye of the trade.
Cooling-Trade Setup: 5-Day Returns Heading into the Marvell + Dell Window
Source: MarketDecode scanner, 5-day returns ending 2026-05-26 (scanner_universe.json); next earnings dates: Marvell IR (Q1 FY27 May 27 1:45 PM PT), Dell IR (Q1 FY27 May 28 after close); PCE date: Bureau of Economic Analysis release schedule
Catalyst #1
MRVL Wed AMCQ1 FY27 print at 1:45 PM PT — networking-silicon guide is the leading cooling-order signal
Catalyst #2
DELL Thu AMCQ1 FY27 print — AI-server backlog disclosure is the second leading cooling signal
Macro catalyst
PCE Thu 8:30 ETDiscount-rate read — hot print pushes infrastructure capex timing back, cool print pulls it forward
Chapter 02
The Mechanism: Same Physical Constraint, Four Different Economic Models
All four names sell into the same physical problem — heat must leave the data center and power must enter it. But the share of revenue that is actually AI-data-center pure-play differs by roughly 6× across the four. That is the mechanism that explains the 20-day return dispersion.
The single most useful number for understanding why Vertiv, Eaton, Trane, and Carrier trade differently is "what fraction of your revenue actually comes from AI data centers right now?" The answer ranges from approximately 50% for Vertiv to approximately 8% for Carrier — a 6× spread among four names that the headline news flow lumps together. Vertiv is structurally the AI-data-center pure-play: roughly half of trailing-twelve-month revenue is in data-center thermal management (rear-door heat exchangers, direct-to-chip liquid loops, chillers, busway, switchgear, lithium-ion UPS), with the bulk of that increasingly tied to hyperscaler and AI workloads. Eaton is the most diversified of the four: roughly 25% AI-data-center share through its electrical busway, low-voltage switchgear, and PDU business, with the balance in commercial buildings, aerospace, and vehicle electrification. Trane Technologies is largely commercial HVAC and transport refrigeration with a data-center sleeve closer to 12%, anchored by chiller and CRAH sales into purpose-built data centers but constrained by the company’s exposure to commercial real estate, schools, hospitals, and food-cold-chain customers. Carrier is the smallest of the four for AI data centers at approximately 8% — commercial HVAC and refrigeration are the primary businesses, with residential HVAC, light commercial, and refrigerated transport (after the 2024 portfolio reshape) doing the volume. The same Nvidia Q1 print therefore pushes Vertiv’s 2026 estimates meaningfully, nudges Eaton’s, barely touches Trane’s, and is almost noise for Carrier. The mechanism is the product-mix gradient — and the 20-day return dispersion in Section 3 maps almost directly onto it.
AI Data-Center Revenue Share by Name (MarketDecode editorial estimates)
Source: MarketDecode editorial estimates aggregated from each company’s most recent 10-K segment disclosures, investor day presentations, and analyst day commentary (2025-2026). Ranges: VRT 45–55%, ETN 20–30%, TT 10–15%, CARR 5–10%. Methodology in private/methodology/cooling-layer-revenue-mix.md
Pure-play above 30%
VRT onlyThe cleanest AI-cooling reaction function — +1–3% on a strong NVDA day historically
Diversified industrial
ETN, TT, CARRAI-cycle exposure but rate-cycle and commercial-construction exposure dominate the day-to-day
Mechanism summary
Mix > priceThe single best predictor of relative performance across these four is the AI-revenue-mix gradient, not technicals or earnings beat
Chapter 03
The Investable Map: Four-Name Scanner Dashboard (2026-05-26)
20-day returns spread from -7.2% to +2.0%. Composite scores cluster between 40 and 49 — the scanner sees them as similar in absolute strength but different in trajectory. The insider signal is the cleanest divergence: Carrier is the only one of the four with bullish insider activity this week.
The scanner data tells a layered story when you look at all four names side-by-side rather than one at a time. On 20-day returns, the order is Carrier +2.0%, Vertiv +1.6%, Eaton −6.1%, Trane −7.2%. On composite ranking, the order is Eaton 49, Carrier 46, Trane 44, Vertiv 40 — a 9-point spread, tight enough that the scanner is not declaring a clear winner. On 14-day RSI, the order is Vertiv 49, Eaton 47, Carrier 46, Trane 41 — Trane is the only name approaching oversold territory. On insider signal, the order breaks open: Carrier flags bullish (insider buying detected over the trailing window); Vertiv, Eaton, and Trane all flag bearish (insider selling, neutral options activity, or both). The insider divergence is the most unusual signal in the table because Carrier carries the smallest AI exposure of the four, yet insiders are accumulating while the rest of the layer is bleeding. There are three rational explanations: (a) insiders see commercial HVAC ex-AI strengthening faster than the AI piece can decay; (b) insiders are reacting to a portfolio reshape positive that has nothing to do with AI; (c) the AI thesis is being mis-priced by the market in the WRONG direction — Carrier is more AI-exposed than headline numbers suggest, or it acquires its way in. Each is testable through the next two earnings prints (Carrier reports October 27, 2026, but interim disclosures may move ahead of that). The chart below shows the four names side-by-side across the four scanner metrics.
Cooling-Layer Decomposition — 20-Day Returns by Name (2026-05-26 snapshot)
Source: MarketDecode scanner, 20-day returns ending 2026-05-26 (scanner_universe.json). Composite scores: ETN 49, CARR 46, TT 44, VRT 40. RSI(14): VRT 49.0, ETN 46.5, CARR 46.0, TT 40.8. Insider signal: CARR bullish; VRT / ETN / TT bearish.
Only bullish insider
CARRLowest AI exposure of the four, but the only insider-buying signal this week
Highest composite
ETN 49Best absolute strength reading, but 20-day return is −6.1% — quality at a discount
Most oversold
TT RSI 41Closest to oversold; lowest 20-day return of the four; the contrarian setup if Marvell + Dell deliver
Chapter 04
The Margin Question: Where Does Hyperscaler Buying Power Compress Gross Margin First?
Hyperscaler customers concentrate. Four buyers (Microsoft, Amazon, Google, Meta) account for the bulk of AI-data-center cooling and power equipment demand. That concentration compresses gross margin on commodity products and protects gross margin on specialized ones. Three of the four names face commodity exposure; one is structurally better protected.
The most under-discussed risk in the cooling trade is gross margin. Nvidia’s Data Center gross margin is approximately 75%. The cooling layer’s gross margin is approximately 25–35% depending on product mix. That gap exists because Nvidia sells a deeply differentiated product (CUDA ecosystem + Blackwell architecture + the optical transceivers and NVLink switches that only it ships) while the cooling layer sells products that, at the commodity end, are interchangeable. A 132-kW rack can be cooled by Vertiv, Schneider, Stulz, or Coolcentric — the hyperscaler procurement teams know it, and they negotiate accordingly. The result is gross-margin compression on the most commoditized SKUs (rear-door heat exchangers, in-row CRAC units, standard PDUs) and gross-margin defense on the most differentiated SKUs (liquid-cooling distribution units paired with proprietary software, high-density lithium-ion UPS with long warranty, switchgear engineered to Tier IV uptime targets). The four-name product-mix question becomes: which company’s revenue mix is concentrated in the differentiated end of the catalog? Eaton has the strongest differentiation in low-voltage switchgear and busway — these are engineered-to-order products with long lead times and stickier customer relationships. Vertiv has growing differentiation in liquid cooling and lithium UPS but a long tail of commodity products. Trane Technologies has differentiation in chillers and integrated building management but data-center sleeve is small. Carrier has the least differentiation in pure-AI-data-center catalogs but the strongest non-AI tailwinds. The chart below scores each name’s gross margin profile on a 1–10 scale where 10 is "best protected from hyperscaler buying power."
Hyperscaler Buying-Power Defense by Name (1 = most exposed, 10 = best protected)
Source: MarketDecode editorial framework, 2026-05-26 — scoring blends customer concentration disclosure, engineered-to-order revenue share (10-K segment data), warranty and service revenue share, and competitive intensity in each name’s top-3 SKU categories. Methodology in private/methodology/cooling-layer-margin-defense.md
Best-defended
ETNEngineered-to-order switchgear + diversified customer base = least hyperscaler-buying-power exposure
Highest growth, mid defense
VRTAI pure-play comes with high customer concentration — growth but procurement leverage works against gross margin
Smallest exposure, smallest leverage
CARRLimited AI exposure means limited hyperscaler concentration, but also limited AI-cycle upside
Chapter 05
Bull vs Bear: Six Paired Evidence Points
The bull case for the cooling layer rests on confirmed Nvidia demand, GB200 heat math, and the $630B 2026 capex projection. The bear case rests on insider signals bearish in 3 of 4 names, RSI bleeding, October-November earnings being 5 months out, and the absence of a near-term catalyst that does not depend on Marvell + Dell delivering tomorrow and Thursday.
The cleanest way to assess a sector-battle trade is to put the bullish evidence and the bearish evidence on the same chart, paired by category. Six categories matter here. Demand: Nvidia’s $14.8 billion in pure networking revenue (+199% YoY) confirms cluster-scale buildout (bull); but Marvell and Dell prints could still come in soft or in-line and reset expectations (bear). Heat math: every 132 kW rack must shed heat to a liquid loop, which mathematically grows the cooling-vendor SOM regardless of stock price (bull); but Trane and Carrier have minimal exposure to liquid-cooling SKUs, so the math benefits Vertiv and Eaton disproportionately, not the layer as a whole (bear). Insider signal: Carrier showed bullish insider activity this week (bull); but three of four cooling names flagged bearish, the strongest negative-signal cluster in the cooling layer in three months (bear). Technicals: Trane RSI 41 approaches oversold and Eaton 5-day already turned positive (bull); but Vertiv 5-day is still −3.6% with no clean reversal pattern yet (bear). Calendar: Marvell tomorrow + Dell Thursday + PCE Thursday creates a three-event catalyst window in 48 hours (bull); but cooling-layer earnings themselves do not arrive until October 21 (VRT, GEV) and October 27 (CARR) — a five-month gap between catalysts (bear). Macro: rates have been stable to easing, supporting long-duration infrastructure capex (bull); but a hot PCE Thursday would push the cooling-layer discount-rate back up and stretch the timeline (bear).
Bull vs Bear Evidence — Paired Strength by Category (0 = no evidence, 10 = strong)
Source: MarketDecode editorial scorecard, 2026-05-26 — each bull and bear scored independently against the same evidence base; bull total 41/60 = 68%, bear total 36/60 = 60%. Methodology in private/methodology/bull-bear-paired-scoring.md
Bull total
41 / 60Confirmed demand + heat math + near-term catalyst window = the strong corners
Bear total
36 / 60Insider signals + earnings 5 months out + Marvell/Dell soft-print risk = the durable risks
Decision
Asymmetric longBull case rests on 48-hour confirmation; bear case rests on the absence of a catalyst between now and October — size accordingly
Chapter 06
The Read: What Marvell Tomorrow and Dell Thursday Tell Us
The cooling layer’s next move is decided in 48 hours. Marvell’s networking-silicon guide is the leading order-book indicator; Dell’s AI-server backlog is the second leading indicator. Both confirming pulls the cooling layer forward into June. Either soft pushes it to Q3 earnings season.
The cooling layer carries a structural underweight relative to the rest of the AI infrastructure stack. Memory ran +43% in 20 days. Servers ran +37%. Networking ran +24%. The cooling layer sat between −7% and +2% over the same window. That is the gap the layer is supposed to close — but timing matters. Two earnings prints in the next 48 hours are the decision point. First, Marvell tomorrow at 1:45 PM Pacific. Watch for three things: (a) what Marvell guides on Data Center revenue for the next two quarters — anything materially above $1.5 billion per quarter means the AI cluster buildout is accelerating, which translates roughly four-to-six months later into cooling orders; (b) custom-silicon (ASIC) commentary — if hyperscaler ASIC adoption is accelerating, the rack count grows, and the cooling SOM grows linearly with rack count; (c) optical DSP commentary — if the optics-to-cluster ratio is rising, networking density per rack is going up, which raises the average kW per rack, which raises the cooling spend per rack. Second, Dell on Thursday. Watch for two things: (a) AI server backlog disclosure — every dollar of AI server backlog is roughly $0.05 of cooling-vendor revenue six to twelve months later (the rule of thumb hyperscaler procurement teams quote); (b) margin guidance — if Dell maintains or expands AI-server gross margin, the entire infrastructure layer including cooling is being priced rationally; if margin compresses, the cooling layer’s margin question (Section 4) gets pushed forward. The watchlist below specifies the four exact conditions and the resolution windows. If three of four trigger, the cooling layer reprices through Vertiv by mid-June and dispersion widens. If only zero to one trigger, the layer waits for the October-November earnings season.
The 48-Hour Watchlist — What to Look For and When
Source: MarketDecode watchlist 2026-05-26 — confirmation: any 3 of 4 triggers pull the cooling-layer rerate forward into June; invalidation: 0 of 4 trigger AND VRT 5-day turns more negative pushes the rerate to Q3 (October-November earnings season).
Confirmation threshold
3 of 4Three watchlist triggers = cooling layer reprices through VRT first by mid-June
Invalidation
0 of 4 + VRT lowerIf none of the triggers fire and VRT keeps bleeding, the rerate waits for October earnings
Asymmetric pick
CARRLowest AI exposure but only bullish insider signal — if the rerate happens, CARR underperforms VRT but cushions the downside if it does not
Resolution window — 3 months
What would confirm or invalidate this read
Confirmation
Three of four watchlist triggers fire within the 48-hour window: (1) Marvell Q1 FY27 Data Center segment guide above $1.5B per quarter on the May 27 print; (2) Dell AI server backlog above $5B disclosed in the May 28 print; (3) PCE in-line or cooler than annualized 2.7% on the May 28 release; (4) VRT 5-day return flips positive by Friday May 29 close. If 3 of 4 confirm, the cooling layer reprices through Vertiv first by mid-June, with Carrier outperforming on the insider-signal divergence and Eaton on margin defense.
Invalidation
Zero or one of the four watchlist triggers fires AND Vertiv 5-day return turns more negative by Friday May 29 close. Additional bear confirmations: Marvell Data Center guide below $1.3B/quarter (the cooling-layer leading indicator falters); Dell AI backlog flat or down quarter-over-quarter; PCE annualized above 3.0% pushing the discount rate against long-duration infrastructure capex; or CARR insider signal flips to bearish on the next refresh, killing the anomaly thesis. If 3 of these resolve bearish, the cooling layer waits for the October-November earnings season for any meaningful rerate.