Neel Somani, a researcher and technologist with a background in large-scale computational systems, notes that modern AI deployment is no longer limited by raw processing power alone. Thermal management has become one of the defining economic questions in data center design, particularly as next-generation AI hardware begins to exceed the limits of conventional cooling architecture.
As artificial intelligence infrastructure expands beyond traditional enterprise computing, engineers and investors are confronting a basic physical constraint: heat.
“The conversation around AI infrastructure often focuses on chips, models, and power contracts,” says Neel Somani. “But in practice, the cooling layer determines whether that compute can actually operate at scale.”
The challenge is growing quickly. High-density AI racks built for modern accelerator systems now demand thermal performance far beyond what air-based systems were designed to handle. As that threshold rises, cooling is shifting from an operational detail into a capital allocation decision with direct consequences for energy pricing, asset valuation, and long-term grid participation.
The Thermal Wall and the Risk of Stranded Assets
For decades, air cooling defined the economics of data center design. Raised floors, chilled aisles, and high-volume fan systems were sufficient when rack densities remained within predictable limits. That assumption no longer holds.
Traditional air cooling begins to lose effectiveness at roughly 40 to 50 kilowatts per rack. Above that level, airflow requirements increase dramatically, forcing facilities into diminishing returns where more power is spent moving air rather than supporting computation.
Modern AI systems now routinely exceed that threshold. Emerging high-density deployments built around advanced accelerator clusters increasingly require 132 kilowatts per rack, while some next-generation configurations move beyond 200 kilowatts.
At those levels, the physical properties of heat transfer become decisive. Water can absorb roughly 3,500 times more heat per unit volume than air, allowing far greater thermal movement with significantly less mechanical effort.
To remove 100 kilowatts of heat using air requires near hurricane-force airflow through a rack environment. A liquid system can remove the same heat load using a controlled closed-loop flow of roughly 10 gallons per minute.
That difference creates a new category of infrastructure risk: stranded thermal assets.
A facility may have available floor space, utility power, and network connectivity, yet still remain commercially obsolete if it cannot host modern AI hardware because cooling systems cannot support the thermal load.
“Some of the most valuable existing facilities are discovering that available megawatts alone no longer guarantee relevance,” Somani explains. “If the thermal envelope is wrong, the building effectively loses access to the highest-value workloads.”
This is increasingly affecting older enterprise campuses where retrofit costs now compete directly with greenfield development economics.
Marginal Cost and the New PUE Arbitrage
The financial argument for liquid cooling begins with a difficult upfront calculation.
Retrofitting a facility for liquid cooling often costs between two and three million dollars per megawatt, depending on piping architecture, coolant distribution units, structural modifications, and redundancy requirements.
For many operators, that creates hesitation. Air systems are familiar, easier to maintain, and already integrated into operating procedures.
Yet operating economics increasingly favor liquid systems over time because cooling energy consumption falls sharply once heat transfer efficiency improves.
Liquid-cooled environments can reduce cooling energy demand by 10 to 40 percent, depending on load profile and ambient climate conditions.
That improvement directly affects Power Usage Effectiveness, the industry metric comparing total facility power to compute power.
Air-cooled facilities often operate between 1.5 and 1.7 PUE under high-density stress conditions. Liquid-cooled environments increasingly achieve 1.02 to 1.10.
That difference becomes economically significant at the utility scale.
In a 100-megawatt campus, even a fractional reduction in PUE can translate into millions of dollars in annual energy savings.
Operators increasingly describe this as PUE arbitrage: capital spent once to permanently reduce the non-compute portion of power consumption.
Hardware reliability also shifts.
Stable liquid temperatures reduce thermal cycling, one of the major drivers of solder fatigue, component stress, and leakage current.
The result is longer component life, especially in GPU-intensive deployments where thermal instability can shorten replacement cycles.
“Temperature stability matters more than many procurement teams initially realize,” says Somani. “When thermal variance falls, hardware behaves more predictably, maintenance intervals improve, and depreciation curves can extend.”
That makes cooling not only an efficiency investment but also a hardware longevity strategy.
Grid Reliability and Market Participation
The global power implications are becoming difficult to ignore.
International projections suggest data center electricity demand could approach 945 terawatt-hours by 2030, representing roughly three percent of total global electricity demand.
AI is accelerating that trajectory.
Large inference clusters draw power with highly concentrated load profiles, creating new stress points for local grids.
Liquid cooling changes how these facilities interact with power systems because thermal inertia improves operational flexibility.
Air systems respond quickly to load spikes but offer limited buffering. Liquid systems hold thermal capacity longer, allowing operators to modulate compute behavior more gradually during grid stress events.
This has created interest in treating liquid-cooled facilities as partial thermal batteries.
In markets such as Texas, where ERCOT pricing volatility can spike dramatically during peak demand, operators increasingly explore demand-response participation.
Facilities capable of reducing non-critical load or shifting thermal draw during high-price intervals may receive direct market compensation.
Liquid systems improve that response window because coolant loops retain usable thermal stability even during rapid operating changes.
The effect is subtle but increasingly valuable.
Instead of acting as passive power consumers, advanced data centers become controllable industrial participants inside electricity markets.
Regulators are noticing this as well.
In power-constrained hubs such as Northern Virginia and Dublin, permitting discussions increasingly include thermal efficiency expectations because local substations are approaching delivery limits.
Facilities that demonstrate lower non-compute power demand gain stronger permitting leverage.
The Water-Energy Paradox
At first glance, liquid cooling appears to create a sustainability contradiction: replacing air with water in an already resource-intensive industry.
In practice, the opposite often occurs.
Traditional air cooling at a large scale frequently depends on evaporative cooling towers, which consume significant freshwater through continuous evaporation.
Many high-efficiency air systems achieve thermal performance only by increasing water use indirectly.
Closed-loop liquid cooling changes that balance.
Because coolant circulates in sealed systems, freshwater consumption can fall by 70 to 90 percent depending on facility design.
That distinction matters in regions where water permitting is becoming as sensitive as electrical access.
The economics extend further when heat reuse becomes possible.
Liquid cooling produces stable outlet temperatures often between 45 and 60 degrees Celsius, warm enough for secondary industrial use.
District heating systems, greenhouse operations, and nearby industrial processes can absorb that waste heat as a usable thermal product.
In parts of Europe, this has already shifted from experimental practice to regulatory expectation.
Waste heat is increasingly viewed not as a byproduct but as an asset.
“Once thermal output becomes predictable, it stops being waste,” Somani says. “It becomes another energy stream that can be monetized or regulated.”
That creates an emerging ESG advantage.
Facilities able to demonstrate thermal reuse may improve compliance positioning while offsetting a portion of cooling infrastructure costs.
A New Economic Layer Beneath AI
The AI era is often framed as a race for larger models and faster chips, yet beneath that race sits a quieter economic reality.
Every new watt of compute creates a corresponding thermal liability.
As rack densities rise, cooling is no longer a facilities discussion delegated to engineering teams after procurement decisions are made.
It now influences real estate value, capital planning, utility negotiations, regulatory approvals, and long-term infrastructure competitiveness.
The economics of AI increasingly depend not only on how efficiently systems calculate, but on how intelligently they remove heat.
