Data Center Power Demand: Cooling Limits and Nuclear Offtake (Investor Guide)

Updated November 7, 2026 with the post-election policy picture. Screen data as of the November 6 close.

Vertiv (VRT) grew third-quarter sales 61% and its organic orders more than doubled. GE Vernova (GEV) now holds 122 GW of gas turbine backlog and slot reservations. Modine (MOD) nearly doubled its data center sales.

At the November 6 close Vertiv sat ~52% below its 52-week high. Modine sat ~72% below. Constellation (CEG) was ~34% off its high. Then election week delivered three governors-elect who ran on promises of data center moratoriums or pauses.

The easy read says the AI power trade finished. The backlogs disagree, and so does the grid.

Who can turn data center power demand into operating megawatts and cash flow, on what schedule and at whose cost?

The short answer. US data centers used an estimated 192 TWh of electricity in 2024, 4.7% of the US total, per Lawrence Berkeley National Laboratory. Its reference case reaches 649 TWh and 11.8% by 2030. Usable megawatts are the bottleneck. Grid connections, equipment, cooling and permits decide who gets paid first.

What data center power demand means for investors

Data center power demand is the electricity a facility draws to run its servers plus everything that keeps them alive: cooling and power conversion plus lighting and backup systems. For investors the useful unit is the operating megawatt. A signed contract, a grid application and an announced campus are all promises. Only energized capacity bills a customer.

Let me flag that distinction now because I'll come back to it in every section. It separates the companies already collecting cash from the ones still waiting on a utility, a turbine or a permit.

IT load vs facility load

IT load is what the servers, storage arrays and network gear consume. Facility load is IT load plus the overhead around it.

The Department of Energy calls the ratio between the two PUE (power usage effectiveness) and defines it as total facility annual energy divided by IT equipment annual energy. A building running at a PUE of 1.2 that hosts 100 MW of IT load draws ~120 MW in total. Cooling is only part of those extra 20 MW, with power conversion losses and other systems mixed in.

One number to keep in your head for every headline you'll read: a 1 GW load running all year uses 8.76 TWh. A 1 GW grid application falls well short of 8.76 TWh of real demand once load factor and redundancy cut into it and the ramp-up takes its time.

Why AI packs megawatts into less floor space

AI training and inference racks pull far more power per square foot than the enterprise servers most data centers were built for. NVIDIA (NVDA) puts its GB200 NVL72 rack at ~120 kW. Its GB300 NVL72 reference architecture calls for up to 142 kW per rack.

Same building and footprint with several times the heat, and that's why power and cooling now move together.

Key terms

The demand stack: where the megawatts show up

The three most cited forecasts measure different things. LBNL puts US data center consumption at 192 TWh in 2024 and 649 TWh by 2030 in its reference case. EIA tracks server electricity inside the commercial sector. The IEA projects global data center use roughly doubling from 485 TWh in 2025 to 950 TWh in 2030.

Which of these should you trust? Each one answers a different question, and the common mistake is stacking them on top of each other as if they shared a denominator.

The national estimates, dated and sourced

Before you quote any of these, a few cautions. LBNL's 2024 number comes out of a model. Its range combines several uncertainties, so don't read it as a statistical confidence band. EIA's share covers servers inside commercial buildings only, which means it can't stand in for a share of all US electricity.

The IEA's global figure comes from a different model again, so dividing it into LBNL's US forecast gives you a number neither source published.

Utility forecasts are a different animal

Grid operators plan for peak demand across the whole system. PJM is the largest US grid operator. Its 2026 long-term forecast projects summer peak reaching 183,008 MW in 2030 and ~222,000 MW in 2036, which works out to ~3.6% a year over ten years.

PJM's revision says more than the headline. It trimmed near-term load growth compared with last year after vetting large-load requests more carefully. ERCOT is doing the same thing in Texas, adjusting utility-submitted peak forecasts because raw requests overstate what will actually connect.

I think better filtering shrinks the paper demand. I don't think it proves the underlying demand went away, and I'd want to see energized megawatts fall before believing that.

For the grid-wide economics of who earns money from a power shortage, see our Power Grid Masterclass (free registration).

The grid bottleneck: queues, transformers and turbines

A data center needs a grid connection and the transformers and switchgear to use it, plus enough generation behind it. Each piece has its own waiting line. The median power project completed in 2025 took more than five years from interconnection request to operation, per LBNL. Transformer and switchgear lead times ran 31 to 54 weeks in early 2026.

Interconnection queues

LBNL's Queued Up 2026 edition counted more than 8,200 projects and ~2,060 GW of generation and storage waiting for grid connections at the end of 2025. Only ~13% of the capacity requested from 2000 to 2020 was operating by the end of 2025. These are queues for power plants and batteries, separate from data center load requests. The supply side has to clear them before new demand can be served, which is why I track them alongside the load forecasts.

Capacity prices

PJM's auction for the 2028/29 delivery year cleared at the $325 per MW-day price cap and came in 6,831 MW short of its reliability requirement. The two prior auctions cleared at $333.44 and $329.17.

A capacity price pays plants to be available at peak, and it sits apart from the per-MWh price on your power bill. The obvious objection is that a 6,831 MW shortfall means blackouts are coming. It means the system is tight at the price ceiling regulators set, which is a narrower claim.

Equipment lead times and turbine backlogs

Gilbane's Q1 2026 market survey put pad-mounted transformers at 31 to 51 weeks and low-voltage switchgear at 32 to 54 weeks. Large engineered transformers are a separate and slower product class.

On the generation side the backlogs run deep.

Read the GE Vernova line carefully. Nearly half of that 122 GW is reservations, which hold a factory slot and can still lapse. Siemens Energy's figure includes service work, so it can't be compared with GE Vernova's gigawatts. No company's backlog divided by factory output gives you a reliable delivery date either.

Gas turbines for data centers as bridge power

When the grid can't connect a campus fast enough, developers look at on-site gas generation. That can shorten the wait but brings its own constraints: an air permit, a gas pipeline agreement and a turbine slot that may itself be years out.

In my view bridge power just moves the bottleneck.

Liquid cooling: when air runs out of room

Dense AI racks produce more heat than moving air can carry away efficiently, so NVIDIA's rack-scale systems are designed for liquid cooling. Direct-to-chip cooling pipes liquid to cold plates on the processors, while immersion submerges hardware in dielectric fluid. A CDU controls the coolant loop. DOE said in 2023 that cooling can reach up to 40% of facility energy.

Rack density by NVIDIA platform

You'll see 190 to 230 kW figures for Rubin racks in trade coverage, but those are third-party estimates and NVIDIA hasn't published an official operating-demand number for that rack.

Before we go further, a caution on thresholds. No single kW level marks the point where air cooling stops working everywhere, because it depends on heat flux, airflow, inlet temperatures and the building. The densest AI systems ship with liquid in mind either way.

Direct-to-chip vs immersion vs CDU

Direct-to-chip cooling runs coolant through cold plates bolted to the high-power processors. Fans and room air still handle the lower-heat components.

Immersion cooling submerges suitable hardware in a dielectric fluid. That changes how technicians service the equipment and narrows which hardware can go in the tank.

A CDU sits between the two worlds. It delivers controlled, clean coolant to direct-to-chip or immersion equipment and keeps the IT loop separate from facility water.

None of these removes the final step because heat still has to leave the building through chillers, dry coolers or towers. Liquid at the chip makes the system more efficient while every watt of that heat still has to go somewhere.

How far can efficiency go? DOE's 2023 COOLERCHIPS announcement said cooling can account for up to 40% of a data center's energy use. ARPA-E's research target for the same program is cooling energy below 5% of IT load. Treat the first as an old upper bound and the second as a goal, and use neither as today's industry average.

What cooling growth looks like in supplier results

Two points for investors. Vertiv's growth includes more than liquid cooling, so the headline sales number overstates pure liquid exposure. Modine shows what fast growth can cost, with data center sales up 95% while segment gross margin fell 470 basis points. Growth that dilutes margins needs a payoff later, and I want to see Modine's margin recover over the next few quarters as proof.

Neither company disclosed a numerical backlog for the September quarter in the sources we used, so this guide doesn't cite one.

Nuclear power for data centers: firm power on slow timelines

Nuclear plants run 24/7 outside refueling outages, which suits data centers with constant load. Near-term deals mostly restart or extend existing reactors. Newly built advanced reactors and SMRs look like a 2030s story, according to the Office of Nuclear Energy at DOE. A signed offtake agreement is not an operating reactor, and an existing-plant contract often preserves output without adding new supply.

Restarts and existing plants

A 20-year nuclear contract sounds like new supply, and you'd be forgiven for reading it that way. For each line ask whether the contract creates new megawatts or locks in megawatts that already exist. Both can be good business, but only the first adds supply to the grid.

For a deeper look at the nuclear fuel chain and its listed names, see The Nuclear Stack (GNG Pro).

Advanced reactors and SMRs are mostly a 2030s story

The named advanced projects have real sponsors and long timelines. Google (GOOG) with Kairos Power and TVA target the 50 MW Hermes 2 plant for 2030 inside a broader goal of up to 500 MW by 2035. X-energy's Cascade project with Energy Northwest starts at 320 MWe and could expand to 960 MWe. In October 2026 Amazon (AMZN), X-energy and American Electric Power (AEP) announced plans to explore 12 Xe-100 reactors targeting 960 MW in Indiana.

The NRC's current generic schedule allows 18 months for a construction permit review, and that's only the review deadline. The review is only one step: financing, construction and the approvals needed to load fuel also matter. Can nuclear power data centers next year then? Existing plants already do, and new reactors mostly need until the 2030s.

Co-location and who pays for the grid

Putting a data center next to a power plant sounds like a way around the grid, but federal regulators haven't let it work that simply. In December 2025 FERC told PJM to write clearer co-location rules. In June 2026 it opened large-load proceedings covering cost allocation and transparency along with stranded-cost protection, with responses due November 16 for five grid operators and November 20 for SPP.

FERC itself says these show-cause orders neither make final findings nor impose new rates. I'll be reading those filings closely.

After the election: whose power is already usable?

Dated update, November 7, 2026.

Governors-elect in Texas and Ohio campaigned on data center moratoriums or conditions, and Georgia's governor-elect campaigned on a pause. Campaign platforms are not enacted orders, and the scope and grandfathering rules don't exist yet. The useful split runs between capacity already operating and capacity still waiting on approvals.

What was actually promised

The obvious worry for anyone holding a supplier is that every campaign promise turns into a cancelled order. You may have seen an estimate that ~45% of planned US capacity is exposed to possible restrictions. This guide skips it because the underlying project list and its treatment of grandfathered projects weren't available to check, and exposure to a proposal still isn't cancelled capacity.

We also aren't claiming voters rejected AI, since reporting supports rising political resistance to data centers but can't isolate why people voted the way they did.

Operating megawatts vs pipeline megawatts

Restrictions can make existing, permitted capacity scarcer and more valuable. They can also delay unbuilt campuses while interest and holding costs keep running. The same company can sit on both sides.

IREN (IREN) shows how that works. Its November 5 results say Horizon 1 was delivered to Microsoft (MSFT) while Horizons 2 through 4 were in late-stage construction. Its 2 GW Sweetwater hub in Texas carries a conditional ERCOT classification, and its high-voltage substation is energized with 300 MW gross of data center construction targeted for Q4 2027. An energized substation is not finished, accepted IT capacity.

IREN also pushed its target of more than $4 billion in annualized revenue run-rate from November to January, possibly February 2027. The company tied that to design changes and supplier timing. That delay predates the election.

Who gains and who gets hurt, by business type

Treat the table as scenario analysis. The post-election proposals are not enacted law, and the table assigns no probabilities. New York's existing executive action is a separate measure.

The GNG exposure map: screen data, not recommendations

The listed companies tied to data center power fall into six groups: utilities, owners of nuclear and independent power, gas turbine makers, grid equipment suppliers, cooling suppliers and builders. The table shows each name's GNG screen data at the November 6 close. Several popular power names screen as Hold or Sell today.

How to read it. GNG Quant is GNG's quantitative screen, separate from the Wall Street rating column and from any Vulcan MK5 rating. Forward P/E and EPS estimate changes come from Finnhub's crowdsourced estimates, so treat them as estimates. Talen (TLN) is the only Sell in the table even with Wall Street at Buy.

Data is from the GNG database at the November 6 close and isn't a live quote. CEG and Vistra (VST) still show June 30 as their latest quarter in the dataset, and so does Southern (SO).

Siemens Energy and Mitsubishi Heavy Industries trade outside the US and aren't in the GNG dataset, so they have no row.

Notice what the table says as a whole. You might expect names with rising estimates to trade near their highs. Yet Vertiv plus four builders with rising estimates sit ~31% to ~75% below theirs: Comfort Systems (FIX), Quanta (PWR), EMCOR (EME) and Sterling (STRL). The grid equipment group of Eaton (ETN), Hubbell (HUBB) and nVent (NVT) sits ~24% to ~42% off its highs, with Trane (TT) ~20% off.

The three utilities show small estimate changes and sit within 10% of their highs, led by Dominion (D) at ~4% off. I read that as the market pricing execution and policy risk into the suppliers and treating regulated utilities as the safer seat. Rising estimates on falling prices is exactly the setup the falsifiers below are built to test.

If you want our earlier ranked list of AI power and utility names, see My 2026 Top 10 AI Power and Utilities Refresh (free registration). This guide builds on that list without replacing it.

Seven things that would break the power thesis

The "power is the binding constraint" thesis fails if efficiency outruns workload growth, customers cut spending, queue reform reveals much less real demand, new supply catches up, equipment shortages ease, policy blocks enough projects, or suppliers can't turn growth into cash.

Two cautions on reading these. A smaller queue can mean better filtering with the same real demand underneath. Capacity prices that fall only because of a regulatory cap don't prove supply is adequate either.

If you only track one falsifier, make it the last one. Power can stay physically scarce while a supplier's stock still disappoints because scarcity helps only the companies that convert it into free cash flow.

The Vulcan checklist for the power trade

Before you treat any company as a data center power winner, run it through these questions in order.

1. Is the megawatt operating? Energized and accepted capacity that's billing beats contracted, which beats announced.

2. Who pays for the grid connection? Check the tariff, the deposits and any minimum bill. Cost allocation is where regulators are focused right now.

3. Is the backlog firm? Separate firm orders from reservations and options.

4. Does growth convert to cash? Compare revenue growth with operating cash flow and margins.

5. Which timeline applies? Grid equipment runs in quarters and turbines in years, while new reactors run into the 2030s.

6. What is the policy exposure by site? A state headline can't tell you which projects are grandfathered.

7. What does the screen say today? Estimates and revisions move. Recheck before acting on anything here.

For a cash-flow case study on hyperscaler spending, see Google Doubled Its Cloud Business and Lost 10% Overnight. For the financing side of the AI capital cycle, see AI Stocks Are Selling Off. I Added to Nebius Anyway.

FAQ

How much electricity do US data centers use?

Lawrence Berkeley National Laboratory estimates US data centers used 192 TWh in 2024, 4.7% of US electricity. The figure is a modeled estimate and excludes crypto mining.

How much power will data centers need by 2030?

LBNL's reference case puts US data center use at 649 TWh in 2030, 11.8% of US electricity, with a range of 521 to 843 TWh. Globally the IEA projects 950 TWh in 2030, up from 485 TWh in 2025.

Is AI electricity demand the same as data center power demand?

No. AI is a growing share of data center demand. Data centers also run cloud, storage, streaming and enterprise workloads. The IEA expects AI-focused data center demand to roughly triple between 2025 and 2030.

Why do AI data centers need liquid cooling?

AI racks pack far more power into the same space. NVIDIA rates its GB300 NVL72 rack at up to 142 kW and designs its rack-scale systems for liquid cooling. Air alone can't carry that much heat away efficiently.

What is the difference between direct-to-chip and immersion cooling?

Direct-to-chip cooling pumps liquid through cold plates mounted on processors. Immersion cooling submerges servers in a dielectric fluid. They need different hardware, maintenance and building designs.

What is a CDU in data center cooling?

A coolant distribution unit controls coolant temperature, flow and pressure and can isolate the IT cooling loop from facility water. It supports direct-to-chip and immersion systems. It doesn't generate cooling on its own.

Can nuclear power data centers soon?

Existing reactors already do, through long-term contracts and restarts like Constellation's Crane project, targeted for 2027. New advanced reactors and SMRs are mostly expected in the 2030s, per DOE's Office of Nuclear Energy.

Does a nuclear power purchase agreement create new electricity?

Sometimes. A restart like Crane adds supply back to the grid. A contract on an existing plant like Clinton mostly preserves output that already exists. Clinton's new capacity is a 30 MW uprate.

Why are gas turbines used for data centers?

On-site gas turbines can power a campus while it waits for a grid connection. They need air permits, a gas supply and a turbine slot. GE Vernova alone reports 122 GW of gas backlog and reservations.

Will data center projects be cancelled after the 2026 election?

Some may be delayed, moved or cancelled. It depends on the orders and laws that actually pass and on each project's stage. Campaign platforms in Texas, Ohio and Georgia show intent. They are not enacted rules yet.

Sources

All sources retrieved November 7, 2026.

• LBNL / DOE, United States Data Center Energy Usage Report: 2025 Update, June 18, 2026

• EIA, Today in Energy (AEO2026 server electricity), May 19, 2026

• IEA, Key Questions on Energy and AI, April 16, 2026

• PJM, 2026 long-term load forecast

• ERCOT market notice M-B051826-01, May 18, 2026

• LBNL, Queued Up: 2026 Edition, June 2026

• PJM, 2028/2029 capacity auction results, July 14, 2026

• NJ BPU, investigation of PJM's capacity market, August 12, 2026

• Gilbane, Q1 2026 Market Conditions Report

• GE Vernova, Q3 2026 results, October 28, 2026

• Siemens Energy, Q3 FY2026 earnings release, August 5, 2026

• Mitsubishi Heavy Industries, 1H FY2026 presentation, November 6, 2026

• NVIDIA, MGX technical blog (GB200 NVL72)

• NVIDIA, GB300 NVL72 reference architecture, March 2026

• NVIDIA, DGX Vera Rubin NVL72 hardware guide

• DOE, COOLERCHIPS announcement, May 9, 2023

• ARPA-E, COOLERCHIPS program

• DOE FEMP, data center cooling and PUE

• Vertiv, coolant distribution units explainer

• Saint-Gobain, data center liquid cooling explainer

• Vertiv, Q3 2026 results (SEC exhibit 99.1), October 28, 2026

• Modine, Q2 FY2027 results, November 3, 2026

• Constellation, Crane Clean Energy Center and Q2 2026 results

• Constellation and Meta, Clinton agreement, June 3, 2025

• Talen Energy and Amazon, Susquehanna, June 11, 2025

• Google, Kairos Power and TVA, Hermes 2, August 18, 2025

• X-energy, Cascade with Energy Northwest

• Amazon, X-energy and AEP in Indiana, October 2026

• Meta, nuclear agreements with Vistra, TerraPower and Oklo, January 2026

• NRC, generic schedules

• DOE Office of Nuclear Energy, advantages and challenges of nuclear-powered data centers, April 8, 2025

• FERC, large-load cost transparency and customer protection

• FERC, PJM co-location order, December 18, 2025

• Gina Hinojosa campaign, data centers priority

• Amy Acton campaign, conditional moratorium, July 27, 2026

• Fortune, anti-data-center governors, November 4, 2026

• Reuters via Fidelity, election results, November 4, 2026

• New York Governor, data center permit pause, July 14, 2026

• IREN, Q1 FY27 results (SEC), November 5, 2026, and Sweetwater Batch Zero release, September 8, 2026

• GNG Research database, November 6, 2026 close. Forward estimates from Finnhub (crowdsourced)

This guide is educational research. It is not a recommendation to buy or sell any security.