AI’s Next Bottleneck Is Heat
Every watt that powers an AI chip becomes heat. That turns Vertiv, Eaton, Trane, and Carrier into part of the AI trade — even when the market is selling them.
AI rack heat density vs. traditional server rack
16×
▲ +16 132 kW vs ~8 kW per rack — same square foot, sixteen times the heat
Nvidia just confirmed the largest capex cycle since the internet buildout. The cooling stocks sold off anyway: Vertiv −13%, Eaton −4%, Trane −4%, Carrier −4%. Here is the physical story behind the contradiction: an Nvidia GB200 NVL72 rack dissipates ~132 kW of heat — sixteen times more per rack than the servers a traditional data center was built to cool. That is not a chip problem. It is an air-conditioning problem. Why the market is pricing the lag, not the demand. And why one of the four cooling names is already telling you which way this resolves.
The contradiction
Every watt of electricity that powers an AI chip eventually becomes heat. An Nvidia GB200 NVL72 rack dissipates roughly 132 kilowatts — sixteen times more per rack than the equipment a traditional data center was built to cool. That is a physical fact, not a forecast. Yet the week Nvidia confirmed the buildout is real, the four companies that build the cooling for AI factories — Vertiv, Eaton, Trane, Carrier — lost between 4% and 13%. Either the cooling layer is mispriced and will catch up when their order books reflect the new GPU demand, or the market is correctly pricing a structural problem: that cooling is a volume business with hyperscaler buying power compressing margins. Both can be true for different names. The next two weeks tell you which.
What the headline says
Nvidia confirms AI demand
$81.6B Q1 revenue, $91B Q2 guide, Data Center networking +199% YoY
What the data says
But the cooling stocks sold off
VRT −13%, ETN −4%, TT −4%, CARR −4% on the same week
Chapter 01
Nvidia Confirmed the Buildout. Cooling Stocks Sold Off Anyway.
The four companies that build the cooling and electrical infrastructure for AI factories lost between 4% and 13% in the same five days Nvidia printed a single-quarter record and guided up. That is the contradiction this article unpacks.
Friday closed the most bullish week for AI demand confirmation in two years. Nvidia reported $81.6 billion in Q1 FY27 revenue, $75.2 billion of it from the Data Center segment, and Data Center networking revenue growing 199% year over year. The company guided Q2 to $91 billion, implying another 11.5% sequential acceleration on top of an already-record quarter. Five of the eight layers of the AI infrastructure stack rallied immediately: connectivity silicon, networking, memory, compute, and generation. One layer did the opposite. Vertiv — the publicly-traded leader in AI data-center power and cooling — closed the week down 12.8%, the worst single-stock drawdown anywhere in the AI infrastructure complex. Eaton, Trane Technologies, and Carrier all sold off between 3.9% and 4.5%. On a 20-day basis the picture is worse: Eaton is down 10%, Trane down 8%. Carrier, in fact, is essentially flat over 20 days (+2%) — the only one of the four resisting the broader sell. The question this article answers is not whether AI demand is real (Nvidia just confirmed it is). The question is why the layer that physically removes heat from AI factories — a layer with no plausible substitute and no offshoring out — keeps selling.
Cooling & Electrical Stocks: 5-Day and 20-Day Returns Through 2026-05-22
Source: MarketDecode scanner, returns ending 2026-05-22
Worst drawdown
VRT −13%Vertiv — the AI power & cooling pure-play
Month-long bleed
ETN −10% (20d)Eaton sold for a month while the rest of the stack rallied
The anomaly
CARR 2% (20d)Only cooling name flat over 20 days — see Section 5
Chapter 02
Why It’s a Heat Problem, Not a Power Problem
Every watt of electricity that powers an AI chip eventually becomes heat. An Nvidia GB200 NVL72 rack dissipates roughly 132 kilowatts — sixteen times more per rack than the eight kilowatts a traditional data center was built for. That is what makes cooling a physical-build industry, not a software one.
There is a tendency to talk about AI infrastructure as if it were a chip problem with a software wrapper. The physical reality is simpler. A modern Nvidia GB200 NVL72 rack — the standard unit of the Blackwell-class AI factory — draws approximately 132 kilowatts of electrical power continuously when it is training or serving an AI model. By the laws of thermodynamics, essentially all of that 132 kilowatts comes back out of the rack as heat. Not as light, not as motion, not as stored chemical energy. Heat. A traditional enterprise server rack, by comparison, dissipates roughly five to ten kilowatts — the design point most existing data center floors were built to handle, with raised-floor air cooling. Sixteen times the heat in the same square foot. That ratio is what makes cooling and electrical infrastructure a separate trade from chips, not a derivative of chips. The chips are designed in Taiwan and shipped in weeks. The cooling has to be specified, ordered, manufactured, shipped, and physically installed inside a building that may not yet exist. Two-to-four quarters per cycle. And the cooling has to be redesigned for the new density — air alone cannot remove 132 kilowatts per rack at acceptable efficiency, which is why direct-to-chip liquid cooling and rear-door heat exchangers are now standard on Blackwell-class deployments. That redesign is itself a product cycle that has to ship.
Per-Rack Heat Dissipation: Traditional vs AI Compute Density (kW)
Source: NVIDIA Blackwell platform datasheet (GB200 NVL72 spec, 132 kW typical); Uptime Institute Global Data Center Survey 2025 for traditional rack averages; Rubin estimate is mid-2026 industry expectation, NVIDIA roadmap
Standard AI rack
132 kWGB200 NVL72 typical — the Blackwell-class unit
Density ratio
16×AI rack vs. legacy server rack heat output
Next generation
~200 kWRubin platform expected mid-2026 — 25× legacy density
Chapter 03
Why Cooling Sells When GPUs Rally: the Order-Book Lag
A GPU ships in weeks. A cooling unit, electrical distribution module, or liquid-cooling system takes two to four quarters from spec to install. The cooling layer is selling the lag, not the demand.
Capital sorted the eight AI infrastructure layers this week along one axis: how many quarters between Nvidia’s confirmed shipment and your own confirmed revenue. Connectivity silicon (Astera Labs) sits on the same printed circuit board as the GPU — the retimers and CXL chips ship the same week. Revenue lag: effectively zero quarters. ALAB rallied 28%. Networking silicon (Marvell) and memory (Micron) ship into the same server SKU. Revenue lag: one quarter. MRVL and MU rallied 5–8%. Servers (Dell) and generation (Constellation) sit at one-to-two quarters. They rallied 4–7%. Cooling and electrical sit at roughly three quarters at the optimistic end and four-plus when grid interconnection studies or building permits are involved. That is the structural reason the cooling layer is selling in a week when everything closer to the chip rallied. The market is not denying AI demand — it just discounted the cooling layer by the time-value of three to four extra quarters of waiting plus the execution risk those four quarters introduce. Vertiv’s next earnings print is in October, five months away. Until then, the order-book signal is invisible.
Revenue-Conversion Lag by Layer: Cooling Is the Cold End of the Stack
Source: MarketDecode editorial framework, 2026-05-22, based on supplier disclosures and historical industry order-to-revenue cycles
GPU-to-cooling lag
~3 quartersSpec → order → manufacture → ship → install
Next VRT print
October 2026Five months before the order book is visible
Layers already converting
5 of 8Compute, memory, networking, servers, connectivity
Chapter 04
The Margin Question: Volume Without Pricing Power
Even if cooling order books pick up, hyperscaler buying power may compress gross margins toward 25–35% — well below Nvidia’s 70%+. The market is pricing both the lag AND the margin compression. Only one of those is recoverable on a 12-month view.
There is a second reason the cooling layer keeps selling that is harder than the lag and may not resolve at all. Cooling, power distribution, and electrical equipment are partly commodity categories. Vertiv and Eaton manufacture standardized 20-foot precision air handlers, busway, switchgear, and uninterruptible power supplies. Trane and Carrier sit closer to commercial HVAC. The buyer set is small — a handful of hyperscalers plus a long tail of enterprise colocation. The largest buyers (Microsoft, Google, Amazon, Meta) negotiate against each other and against the equipment makers. Historical gross margins in cooling and electrical equipment have run 25–35%, with a few specialized niches (liquid-cooling distribution units, immersion tanks) running 40–45%. Nvidia’s gross margin is over 70%. AMD’s is over 50%. The unit economics of the chip layer are simply different. So the cooling layer faces a compounding question: even when AI cooling volume eventually scales, do margins scale with it, or does hyperscaler buying power compress every incremental dollar of revenue toward a structural mid-thirties margin? This is what the equity market is currently pricing. The drawdown is not only “wait three quarters for the order book.” It is “the order book may come, and the margins may not follow.” Carrier’s relative strength — flat on 20 days while peers bleed — is the early signal. Carrier’s product mix is more biased to specialized commercial HVAC than to commodity equipment. If the early differentiation inside the cold layer is by margin profile, Carrier is the name that does it first.
Gross Margin Profile by AI Infrastructure Layer (TTM, %)
Source: Company 10-Ks and most recent quarterly filings, retrieved 2026-05-22 (NVDA Q1 FY27, AVGO Q1 FY26, MRVL Q4 FY26, MU Q2 FY26, others trailing-twelve-months)
Top of stack
NVDA 75%GPU gross margin — the benchmark
Cooling band
27–35%TT / VRT / CARR — the cliff is structural
Spread
~40 ptsNVDA gross margin vs cooling average
Chapter 05
One Cooling Name Is Different. Two Dates Resolve It.
Carrier’s composite score of 64 with insider buying is the first signal that internal capital believes at least one cooling name is mispriced. Marvell on May 27 and Dell on May 28 grade the volume-vs-margin question for the whole layer.
Inside the four-name cold layer, the early differentiation is small but real. Carrier carries a MarketDecode composite score of 64 as of last Friday — the only cooling name above the 60 bullish threshold — and the underlying factor breakdown includes positive insider-buying activity. Vertiv’s composite score is 44. Eaton’s is 44. Trane’s is 52. The spread is the first signal of internal capital separating one of the four names from the others. The interpretation is not that Carrier wins and the rest lose; it is that one of the four cooling-and-electrical names has crossed the threshold where insiders see a mispricing, and Carrier’s product mix — more biased toward specialized commercial HVAC than commodity cooling equipment — is the most defensible against the margin compression risk in Section 4. The catalyst chain that grades the whole layer runs two days. Marvell reports Q1 FY27 on Tuesday, May 27 after the close. Marvell is the next layer downstream from Nvidia in networking silicon and custom AI ASICs. A confirmation of accelerating networking revenue, anywhere close to Nvidia’s +199% Data Center networking read-through, confirms that the cluster-buildout pace is accelerating — which directly extends GPU shipments, which directly extends the cooling order-book buildup. Dell reports Q1 FY27 on Wednesday, May 28 after the close. Dell is the server layer. A clean print confirms or denies whether AI server unit growth comes with maintained gross margin. If Dell margins hold, the structural-margin-compression thesis from Section 4 weakens, and the cooling layer reprices broadly. If Dell margins compress, the cliff is real and only Carrier-type product mixes work.
Composite Scores Inside the Cold Layer (2026-05-22)
Source: MarketDecode composite scoring, 2026-05-22
May 27 catalyst
Marvell Q1Networking demand confirmation → extends cooling buildout
May 28 catalyst
Dell Q1Server margin test → read-through to cooling margins
Same-day macro
April PCERate path matters for long-duration physical capex
Chapter 06
The MarketDecode Read
Heat is real. Lag is real. Margin compression is real. Three of four cooling names are mispriced for the lag. One is mispriced for the structural margin risk. Marvell and Dell tell you which is which by next Thursday.
Heat is a physical constraint on AI infrastructure, not a financial story. An Nvidia GB200 rack puts out sixteen times the heat of a legacy server rack, and the Rubin platform expected mid-2026 pushes that to roughly twenty-five times. There is no path to running the AI demand Nvidia just confirmed that does not also run through the four companies in this article and the broader cooling and electrical supply chain. The order-book lag is real — cooling sits at quarter three of the conversion chain, and Vertiv’s next print is in October. The margin compression risk is also real — cooling and electrical have run 25–35% gross margins historically, and hyperscaler buying power tends to compress, not expand, that band. The two failure modes are different. If the layer is selling the lag, then by the Q3 2026 prints (Vertiv’s October, Eaton’s August), the order books will be visibly larger and the layer reprices in advance of those prints — likely sometime between June and August. If the layer is selling the margin, the lag clears, the volume arrives, the gross margin does not expand, and the layer stays cheap because the equity story was always volume-without-margin. Carrier is the early signal. Marvell and Dell are the trigger. The next two-week window grades both questions on the same Tuesday and Wednesday — May 27 and May 28 — and we will know which version of the story is operative by the close on Thursday.
Cold Layer Synthesis: Returns, Margins, and Composite — 2026-05-22
Source: MarketDecode scanner, 2026-05-22 — 4 cooling names, 20-day windows except VRT which is the worst 5-day in the stack
Verdict
3 mispricedVRT, ETN, TT — selling the lag, not the demand
Watch
CARR signalInsider buying + product mix defensible against margin risk
Resolution
May 27–28Marvell + Dell + PCE same 48-hour window
Resolution window — 1 month
What would confirm or invalidate this read
Confirmation
Marvell’s May 27 print shows Data Center networking revenue growing at 80%+ year over year (confirms cluster-buildout pace, extends cooling order book); Dell’s May 28 print holds AI-server gross margin within 100 bps of consensus (denies the structural-margin-compression thesis for the cooling layer); and Vertiv, Eaton, or Trane recover at least 50% of their 20-day drawdown within 15 trading days after the prints. Two of three confirm = the cooling layer is mispriced on lag and reprices by mid-June. Carrier outperforming the other three by 500+ bps within the same window = the margin-vs-volume sort begins inside the layer.
Invalidation
Marvell misses on networking revenue, OR Dell’s AI-server gross margin compresses by more than 200 bps versus the prior quarter, AND VRT/ETN/TT extend their drawdowns by another 5% over the following 10 trading days. That combination says the cooling layer is correctly cheap on a structural-margin basis, the buildout volume goes to commoditized suppliers, and the equity beneficiaries are upstream (the chip layer, the connectivity layer) rather than physical-infrastructure.