The Real Cost of Cooling

The cooling line item on a data center P&L is a fiction of accounting, not a description of reality.

Operators report cooling as a share of the power bill. That number is real. It is also the smallest cost on a long list of financial exposures that thermal strategy creates. Cooling decisions drive capital intensity, stranded capacity, GPU revenue throughput, water and grid externalities, siting timelines, and an emerging compliance burden that is being priced into asset valuations whether operators are tracking it or not. The cooling bill on the P&L is the visible portion of an iceberg.

The visible portion is itself large. Lawrence Berkeley National Laboratory's 2024 United States Data Center Energy Usage Report documents that cooling systems account for roughly 38 to 40 percent of electricity consumption inside a typical data center, with IT equipment consuming the balance. Uptime Institute's 2024 Global Data Center Survey puts the industry-average power usage effectiveness at 1.56, essentially flat for five consecutive years. As Andy Lawrence, Executive Director of Research at Uptime, observed, average PUE remains stubbornly flat across the industry while newer hyperscaler builds achieve 1.3 or better. At PUE 1.56, every dollar of compute electricity demands roughly fifty-six cents of overhead, the majority of which is cooling.

That direct OpEx is the only part of cooling most operators rigorously track. The argument of this piece is that the other costs (stranded thermal capacity, throttling losses, cooling overbuild in capex, water consumption, grid cost causation, community opposition, regulatory disclosure) are now large enough, and being internalized fast enough, to dominate asset-level returns.

The frame is full P&L, not sustainability. Thermal strategy is a financial discipline.

Start with the easy part. A 30 MW facility running at PUE 1.56 with 70 percent IT utilization burns roughly 184,000 MWh per year on non-IT loads, the bulk of it cooling. At an industrial wholesale-plus-delivery rate of $80 to $130 per MWh in major US markets, that is $15 to $24 million in annual cooling-related power. In tight markets the number is now higher and rising. PJM's 2027/2028 capacity auction cleared at the FERC-approved price cap of $333.44 per MW-day, the third consecutive record, with the market falling 6,625 MW short of the reliability requirement for the first time since the capacity market was created. Cooling power is the single largest controllable line item and the one most exposed to a structurally tightening grid.

The cooling bill is not just power. It contains four other elements that operators tend to bury in non-power line items or in capital depreciation.

Water consumption fees and rights. US data centers directly consumed 17.4 billion gallons of water in 2023 for cooling, projected by LBNL to reach 38 to 73 billion gallons by 2028. Indirect water consumption tied to power generation was another 211 billion gallons in 2023. Direct cooling water has historically priced near municipal industrial rates. In scarce markets it now prices closer to water rights acquisition cost.

Chemical treatment and refrigerant management. Cooling tower water demands biocides, scale inhibitors, and corrosion control. F-gas refrigerants are subject to phasedown under the EU F-Gas Regulation and the US AIM Act, raising both replacement cost and end-of-life liability.

Maintenance contracts. CRAH, CRAC, chillers, cooling towers, and pumps consume preventive maintenance budgets that scale with installed nameplate, not with utilization. An overbuilt cooling plant carries an overbuilt service contract.

Cooling-related electrical losses. Variable-speed drives, pumps, condenser fans, and humidifiers add parasitic loads that PUE captures imperfectly. PUE is an annual ratio that hides hourly mismatch between IT load and cooling response, which is where most of the waste actually lives.

Direct cooling OpEx is real, measurable, and roughly 30 percent of facility power. It is also the smallest cost on this list.

The first hidden cost is stranded thermal capacity. DCIM vendor Sunbird Software reports that customers have used software-driven capacity reclamation to recover up to 40 percent more usable power capacity at existing facilities. The same dynamic applies to thermal capacity. Operators design cooling to peak summer wet-bulb conditions, to N+1 or 2N redundancy, and to a uniform delta-T across the hall, then run a workload that is none of those things most of the time. The result is a building paying maintenance, depreciation, and standby power on cooling tons that will never be called.

The second hidden cost is GPU thermal throttling. NVIDIA specifies a maximum junction temperature of 83°C for the H100 SXM under sustained load. Exceeding that triggers automatic clock reduction. A cluster running at a throttled state can lose roughly 25 percent of its theoretical maximum performance under aggressive AI training loads. For a hyperscaler renting H100s at $2 to $4 per hour, a 25 percent throughput loss on a 1,024-GPU training cluster running a multi-week pre-training run is in the seven-figure range per run. For a colocation tenant on a fixed power-by-the-kW lease, throttling is a direct conversion of paid power into waste heat.

The economics generalize. H100 SXM has a 700W TDP. B200 reaches 1,000W per GPU. An 8-GPU server chassis dissipates 5.6 to 8 kW. Air cooling in a 4U server chassis maxes out near 350 to 400W per GPU, well below modern flagship parts. Operators running these chips on air-cooled infrastructure either restrict density, accept throttling, or invest in retrofit cooling. Each of those is a P&L decision, and none of them shows up on the cooling line.

The third hidden cost is conservative overprovisioning at the rack and row. ASHRAE TC 9.9 has progressively expanded recommended and allowable thermal envelopes (A1 through A4 classes) to encourage higher inlet temperatures and outdoor air economization. Operator behavior has lagged. Uptime Institute's 2024 data shows average rack densities still below 8 kW with most facilities not yet running racks above 30 kW. Industry estimates put recoverable stranded thermal capacity at 15 to 30 percent of installed cooling through active control. That is OpEx and capex savings invisible on the P&L until the next expansion forces the question.

A data center built today costs $9 million to $15 million per MW of IT load fully equipped, with Northern Virginia at the upper end and West Texas and tertiary markets toward the lower end. At Cushman and Wakefield's 19-market average of $11.7 million per MW, a 30 MW facility represents roughly $350 million of capital before IT equipment. HVAC, mechanical, and cooling systems comprise 15 to 20 percent of total development costs in air-cooled designs. Electrical comprises 40 to 45 percent. Building fit-out the balance.

For a $350 million facility, cooling capex is therefore $50 to $70 million. In a liquid-cooled AI factory the share rises. Schneider Electric's analysis finds liquid-cooled designs at 20 kW per rack achieve roughly 10 percent capex savings versus air on a per-watt basis, and 14 percent at 40 kW per rack, by compacting IT and eliminating containment and CRAH/CRAC footprint. Direct-to-chip cooling carries a $2,500 to $4,500 per kW capex premium versus pure air at the rack level when sized for chip-level cold plates, CDU deployment, and server integration. Cushman and Wakefield documents 40 percent chiller price escalation since Q3 2021, with lead times exceeding 28 to 46 weeks for many mechanical components.

The stranded-capital math is straightforward. If a 30 MW facility commits $60 million to cooling infrastructure and operationally uses 60 percent of installed cooling capacity, $24 million of capital sits idle. At a 10 percent hurdle rate that is $2.4 million per year of unproductive capital cost, plus depreciation, plus maintenance on the unused capacity. For investors evaluating asset-level IRR on data center REITs and infrastructure funds (Digital Realty, Equinix, KKR's CyrusOne, Blackstone's QTS, Brookfield's Compass, Stonepeak's Cologix), cooling overbuild is a measurable drag on yield.

The forward direction is more, not less, cooling capital intensity. NVIDIA's GB200 NVL72 reference design assumes direct liquid cooling. Microsoft began Azure fleet deployment of direct-to-chip cooling in July 2025. The global data center liquid cooling market is valued at around $5.51 billion in 2025 and is expected to reach $19.03 billion by 2030, a CAGR of approximately 23 percent. Operators that treat cooling capex as a one-time, peak-sized investment rather than a dynamic, controlled, demand-following asset will continue to strand capital at a rising rate.

Water has been the most subsidized input in data center economics. That is over.

LBNL's 2024 report quantifies direct water consumption at 17.4 billion gallons in 2023 and projects 38 to 73 billion gallons by 2028. Hyperscalers will account for roughly half. The cost of that water has been low historically because municipal rates and water rights pricing did not internalize scarcity. They are starting to.

In Arizona, the situation is the most concrete. Mesa's 2019 Large Customer Sustainable Water Allowance ordinance caps industrial water use and requires developers to bring supplemental supply to exceed the cap. A Mesa spokesperson told Grist that large water users have provided roughly 7,800 acre-feet of additional water supplies, equivalent to roughly 2.5 billion gallons, through water rights purchased on the open market. The Tucson City Council voted 7-0 on August 6, 2025 to reject Project Blue, an Amazon-linked data center; the developer subsequently committed to an air-cooled, zero-water design.

In Newton County, Georgia, Meta's $750 million Stanton Springs data center has produced years of well-water disputes with neighboring residents. Mike Hopkins, Executive Director of the Newton County Water and Sewerage Authority, told the New York Times that the county is pursuing more than $250 million in recycling facility upgrades, describing the work as a "race against the clock." Hopkins disclosed that nine companies have applied for data centers in the county, some requesting up to 6 million gallons per day, more than the county's current total consumption.

In Virginia, the JLARC December 2024 report documents that data centers consumed roughly 2.1 billion gallons in 2023, an 86 percent increase since 2019, with Loudoun County alone accounting for roughly 899 million gallons. JLARC found current water use sustainable but flagged that the absence of standardized facility-level disclosure prevents regulators from assessing cumulative drought vulnerability. JLARC recommended Virginia "expressly authorize local governments to require and consider water use estimates for proposed data center developments." That is a leading indicator of water disclosure becoming a permitting precondition.

The capex implication is dry coolers, closed-loop systems, immersion, and direct-to-chip with adiabatic backup. Each reduces water consumption but increases mechanical capex and power consumption. The OpEx implication is that water priced as a scarce input rather than a municipal commodity changes the cooling architecture decision. Operators selecting evaporative cooling on the basis of historical water cost are underwriting a future expense and a permitting risk that the cooling line on the P&L does not capture.

Indirect water consumption is the other side of the ledger. LBNL estimates 211 billion gallons in 2023 of indirect water tied to electricity generation for data centers, roughly 12 times direct use. Cooling architecture that reduces facility power therefore reduces water consumption upstream, by a factor of roughly twelve. For now, indirect water is a reputational and permitting issue rather than a line item. Direct water cost is already an OpEx line in scarce markets.

Grid impact has shifted from externality to direct charge. The single most consequential financial event in US data center economics in the past two years is the PJM capacity auction repricing.

PJM's 2025/2026 auction cleared at $269.92 per MW-day, an 833 percent increase over the prior year's $28.92, the largest single-year jump in PJM's 27-year history. PJM-wide capacity costs rose from $2.2 billion in 2024/25 to $14.7 billion in 2025/26, with PJM's independent market monitor attributing the majority of the increase to data center load. The 2026/2027 auction cleared at the FERC-approved cap of $329.17 per MW-day. The 2027/2028 auction hit the updated cap of $333.44 per MW-day while falling 6,625 MW short of the reliability requirement, the first such shortfall in PJM history. Cumulative cost projections through 2033 run from $100 billion to $163 billion. In the December 2025 auction specifically, Monitoring Analytics attributed $6.5 billion (40 percent) of the $16.4 billion in total clearing costs to data center load.

The regulatory response is large-load tariffs. AEP Ohio's data center tariff, approved by the Public Utilities Commission of Ohio on July 9, 2025, requires data center customers with peak loads of 25 MW or more to pay for at least 85 percent of their contracted power even if actual usage is less, with a 12-year commitment structure (4-year ramp-up plus 8-year minimum) and exit fees equal to three years of minimum charges. Texas SB 6, signed by Governor Abbott on June 20, 2025, applies to large loads of 75 MW and above and mandates an ERCOT-controlled curtailment mechanism (the "kill switch") that allows grid operators to isolate co-located loads in emergencies, along with interconnection cost recovery and financial commitment requirements. PUCT's draft implementing rule §25.194 published March 12, 2026 would impose a non-refundable $50,000 per MW interconnection fee and 100 percent Contribution in Aid of Construction for direct interconnection costs.

The Maryland Office of People's Counsel filed a FERC complaint on May 7, 2026, projecting that data-center-driven transmission upgrades will cost Maryland customers $1.6 billion over ten years, equivalent to approximately $345 per residential customer, $673 per commercial customer, and $15,074 per industrial customer.

The behind-the-meter response is large. Constellation Energy's 20-year power purchase agreement with Microsoft to restart the Three Mile Island Unit 1 reactor (rebranded Crane Clean Energy Center) restores 835 MW of dispatchable carbon-free power for Microsoft's PJM-region data centers, with a $1.6 billion refurbishment cost. AWS signed for 960 MW from Talen Energy's Susquehanna plant. Meta signed a 20-year, 1,121 MW PPA with Constellation for its Clinton, Illinois plant beginning June 2027.

The strategic pattern is clear. The most aggressive operators are buying their way around grid cost causation. The defensive pattern, available to any 10 to 50 MW operator, is to reduce the grid demand denominator. Cooling is the largest non-IT load. A facility that runs at PUE 1.3 instead of 1.56 reduces its grid footprint, its capacity auction exposure, and its tariff base by roughly 17 percent on the overhead component.

Permitting friction is now a measurable cost in development timelines and capital exposure.

Data Center Watch, the tracking effort by 10a Labs, reports that in Q2 2025 alone, 20 data center projects worth an estimated $98 billion were blocked or delayed amid local opposition, more than the total for all previous quarters since 2023.

Loudoun County voted 7-2 on March 18, 2025, to end by-right zoning for data centers and require Special Exception (SPEX) approval through Planning Commission and Board of Supervisors hearings. DC Byte shows only 28 percent of tracked Loudoun capacity was live in 2024, with 72 percent planned, a decrease attributed to power constraints, new zoning, and community backlash. MultiState tracks more than 50 cities and counties that have enacted bans or moratoria as of mid-2026.

In Memphis, xAI's Colossus 1 data center operated as many as 35 unpermitted gas turbines for primary power before being pressured to remove all but 15 under a permit issued by the Shelby County Health Department on July 2, 2025.

The financial implication is the conversion of permitting risk into a quantifiable line item. Data center development timelines have extended by 18 to 36 months in contested jurisdictions. Community benefit agreements, transmission cost-sharing, environmental impact studies, and reputational risk now appear in pro forma. For 10 to 50 MW operators outside the hyperscaler tier, this matters because the regulatory and community precedent set against megaprojects accrues to all data centers. The cost of being "another data center" in a county with one bad actor is now real.

The EU Energy Efficiency Directive (EED) 2023/1791 and Commission Delegated Regulation (EU) 2024/1364, dated March 14, 2024, require all data centers with installed IT power demand of 500 kW or more to report annually to a European database. The first reports covering 2023 were due September 15, 2024; subsequent reports are due May 15 annually. Required disclosure includes floor area, installed power, energy consumption, PUE, water usage, temperature setpoints, waste heat utilization, and renewable energy use. The Commission is preparing a Data Centre Energy Efficiency Package targeted for Q2 2026 adoption that will introduce a rating scheme and "launch the work on minimum performance standards." The regulatory direction is unambiguous: disclosure today, minimum standards tomorrow, stranded asset risk for non-compliant facilities.

Ireland's Commission for Regulation of Utilities imposed a de facto data center connection moratorium in November 2021. In December 2025, CRU published its Large Energy Users Connection Policy Decision Paper, lifting the moratorium but requiring new data centers to install on-site or proximate generation matching full electricity demand, source 80 percent of annual electricity from additional Irish renewables within six years, and participate in the wholesale electricity market. Data centers consumed approximately 22 percent of Ireland's electricity in 2024.

Singapore imposed a moratorium on new data centers in 2019, lifted in 2022. The May 2024 Green Data Centre Roadmap targets at least 300 MW of additional capacity contingent on green energy use, with PUE targets below 1.3 for new builds.

The IEA's Energy and AI report projects global electricity consumption for data centers will double to roughly 945 TWh by 2030 in the Base Case. EPRI's 2025 update projects US data centers will consume 9 to 17 percent of US electricity by 2030, with Virginia facilities consuming 41 to 59 percent of state electricity.

For a 10 to 50 MW operator, the compliance cost is real but bounded. The strategic cost is asymmetric. A facility that cannot demonstrate metered, sub-system level energy and water performance against ASHRAE thermal guidelines, EED reporting fields, or local large-load tariff requirements faces a stranded-asset risk path: rising tariffs, denied connection upgrades, refused expansion permits, eventual divestment discount. Cooling control is the substrate on which most of these disclosures rest.

The visible cooling line on the P&L is necessary but insufficient. Seven questions a financially literate executive should ask the operations team at the next review:

One. What is the ratio of installed cooling capacity to peak IT load, and the ratio to average IT load? What does the difference represent in capex carried and OpEx serviced?

Two. What percentage of the cooling power bill is fan and pump energy that varies in real time with control, versus chiller energy that is staged less frequently? How is each metered?

Three. How many GPU-hours per quarter are lost to throttling or to administratively de-rated power to prevent throttling? What is the revenue or chargeback equivalent?

Four. What is the facility's water consumption, water cost per gallon, and water rights cost in the current market? What is the substitution capex for dry-cooler or closed-loop architecture?

Five. What is the facility's exposure to the relevant capacity auction or large-load tariff per MW of contracted capacity? What is the OpEx reduction available from a 10 percent improvement in PUE?

Six. What is the facility's compliance posture against EED Annex VII (if EU), ASHRAE TC 9.9 thermal envelope, and the local utility's large-load tariff disclosure requirement?

Seven. Is the BMS, DCIM, and SCADA stack producing data that the CFO can use, or producing alarms that the facilities manager can use? Who owns the control loop above those systems?

The last question is the diagnostic. Most operators have monitoring. Few have control. PUE is reported annually. Cooling decisions are made minute by minute. The economic value of thermal control compounds at the frequency of those decisions.

The structural direction of data center economics is unambiguous. Power is scarcer and more expensive. Water is a priced and permitted input. Grid cost causation is a tariff line, not a subsidy. Disclosure is regulated. Community license is monetized. Cooling capex is rising on a per-MW basis as densities climb past the air-cooled threshold. The thermal envelope of the modern AI factory is narrower than the envelope of the cloud-era data hall, and the cost of operating outside the envelope (throttling, equipment derating, accelerated aging) is converging with the cost of operating inside it.

Inside this environment, the operators that win are those that treat cooling as a dynamic, controlled asset rather than a peak-sized, set-and-forget overhead. Google DeepMind reported that its machine-learning cooling control system "was able to consistently achieve a 40 percent reduction in the amount of energy used for cooling, which equates to a 15 percent reduction in overall PUE overhead after accounting for electrical losses and other non-cooling inefficiencies" in production deployment at Google data centers. That outcome is achievable not because Google has access to special physics, but because Google closed the control loop between IT load, thermal response, and facility actuation, and ran it at the frequency of the underlying decisions.

For a 10 to 50 MW operator running existing BMS, DCIM, and SCADA, the equivalent prize is reachable. Recoverable thermal stranded capacity in the range of 15 to 30 percent, cooling power reduction in the range of 10 to 25 percent, water consumption reduction in the range cited by hyperscalers for closed-loop architecture, and avoided capex on the next cooling plant expansion are all on the table when control replaces monitoring. The capital structure of data centers makes these numbers material to IRR.

The cooling bill on the P&L is the visible 30 percent of an iceberg. The other 70 percent is capex carried on idle plant, throughput lost to throttling, water priced as a scarce input, grid cost causation assigned by tariff, community opposition assigned by permitting timeline, and disclosure assigned by directive. None of that is hypothetical. All of it is being priced into asset valuations now.

The CFO who treats cooling as a facilities-management problem will discover, on the next refinancing or the next expansion permit, that the market has already treated it as a financial-strategy problem.

Arcestra.AI is the control layer that closes the cooling P&L gap.