AI infrastructure is increasingly constrained not only by GPU availability and electricity, but by the ability to remove, distribute, and economically manage heat. For NVIDIA, thermal management has therefore become a strategic adjacency to accelerated computing rather than a peripheral facilities concern. As rack densities move beyond the practical limits of conventional air cooling, customers must deploy direct-to-chip liquid cooling, coolant-distribution units (CDUs), pumps, manifolds, heat exchangers, chillers, controls, and facility-level heat rejection 2,23,45. The opportunity extends across the data-center stack: thermally efficient semiconductors, advanced memory, server and rack design, power delivery, cooling infrastructure, software monitoring, and waste-heat recovery.
The evidence base is recent, spanning July 28 to August 11, 2026. Several claims dated December 11, 2026 are forward-dated relative to the current reporting window and should therefore be treated as lower-confidence scenario or research evidence rather than established market facts. The strongest signals are the repeated, two-source findings that liquid cooling is gaining share, that direct liquid cooling becomes economically compelling around 20–30 kW per rack, that two-phase immersion and CDUs are among the fastest-growing categories, and that thermal resistance remains a material barrier for advanced memory and stacked systems 2,13,24,26,28,47.
Thermal Density Is Becoming the Binding Constraint
AI compute growth increases the thermal and electrical content of every deployed rack. Air cooling reportedly loses efficiency above roughly 30–40 kW per rack 45, while direct liquid cooling becomes more compelling around 20–30 kW per rack 13. At densities of 60–100 kW or more, cooling, electrical, reliability, and operational risks rise sharply 46. Higher density can reduce data-hall area and the number of power-distribution components required for a given IT capacity 13, but those benefits diminish above approximately 20–25 kW per rack 13. Density is therefore not an unconditional advantage. It is a trade-off between space efficiency and thermal resilience.
The operational margin is narrow. High-density cold-plate systems may tolerate loss of coolant circulation for only a few seconds, compared with several minutes for immersion-cooled tanks 13. Higher power density shortens thermal ride-through times and reduces the margin for error 13. Direct liquid cooling may also require additional UPS capacity for mechanical loads while increasing capital costs, commissioning complexity, retrofit requirements, and supplier dependence 13. Legacy facilities often lack the plumbing required for liquid cooling 45, and retrofits are expected to cost more than incorporating the technology into a new build 13.
This distinction matters for NVIDIA because the addressable market for GPUs is increasingly determined by genuinely deployable, thermally enabled capacity—not by floor space or nominal power availability alone 43. A customer may have electrical service and physical room for a new cluster yet lack the fluid distribution, heat rejection, controls, or operational procedures required to run it safely.
The Transition Is Toward Integrated Thermal Architectures
The market is moving from air cooling toward liquid cooling, but the durable conclusion is not a simple air-versus-liquid substitution. Future facilities are likely to combine multiple thermal layers across compute hardware, servers, racks, CDUs, and facility infrastructure 2,25,30,31. Direct-to-chip cooling is described as the leading lower-complexity approach for 2025–2026 deployments and represented approximately 42.85% of liquid-cooling revenue share in one market estimate 45.
Single-phase immersion remains the leading immersion technology. Its appeal rests on compatibility with secondary water loops, the absence of phase-change-fluid management and specialized pressure containment, and the possibility of installation by standard mechanical contractors 47. Two-phase immersion offers materially better heat transfer and can support rack densities that single-phase systems cannot handle efficiently 47. It is also identified as the fastest-growing cooling segment, with CDUs another rapidly expanding component category 47.
The higher performance of two-phase systems comes with greater operational complexity 45. Some fluids create PFAS-related environmental and regulatory risks 45. The phase-out of PFAS-based Novec fluids could produce substitution risk, fluid shortages, requalification expense, and stranded compatibility investments 47. Single-phase direct-to-chip systems may avoid part of this exposure 45, but the absence of a unified international standard for engineered dielectric fluids remains a market-wide qualification problem 47.
What the marketing materials do not show you is the licensing and qualification surface area embedded in the cooling architecture. Customers increasingly prefer complete thermal-management solutions over standalone immersion tanks 47. They favor suppliers that can provide fluid certification, turnkey integration from rack to heat-rejection loop, modularity, remote monitoring, software-controlled flow management, and support for multiple fluid types 47. Established infrastructure companies consequently have an advantage over single-product specialists 47.
For NVIDIA, this supports a platform-oriented strategy. The company’s existing emphasis on systems, networking, power, and software as complements to accelerator silicon can extend into partnerships and integration with suppliers such as Vertiv, Schneider Electric, nVent, and Flex 2,20. The practical priority belongs to the ecosystem that can qualify and operate the complete system, not merely the component with the highest theoretical heat-transfer rate.
Thermal Innovation Begins Before the Facility
The most effective heat is often the heat that is never generated. SiC and GaN power semiconductors are described as reducing conversion losses, heat generation, and thermal-management requirements 29,36. Integrated or near-load power delivery can reduce electrical losses, shorten current-travel distance, improve power density, and simplify customer design 22. An 800VDC architecture may avoid intermediate AC-conversion losses and reduce copper usage by 45% 45,49. These developments affect NVIDIA’s total system economics because lower power loss reduces both electricity expense and the cooling equipment required around each accelerator cluster.
At the package and memory level, materials innovation is being positioned as a response to rising chip power density and cooling requirements 10. Materials that improve integrated-circuit thermal efficiency could reduce energy consumption or improve chip performance 9. Discovered Materials focuses specifically on thermal management in semiconductor materials, while other firms pursue broader AI-driven materials discovery 9. Claimed benefits include lower data-center operating energy through improved heat dissipation 9. Failure to solve cooling constraints, however, could impair the commercial prospects of these materials-discovery companies 8.
The Samsung zHBM claims illustrate both the opportunity and the evidentiary risk. Samsung claims that zHBM could reduce thermal resistance by more than 50%, a point supported by five sources 24,26,28. The broader claim that a threefold improvement in energy efficiency could improve data-center economics has two sources 28. Yet zHBM may still face thermal barriers 27, and the claim that glass substrates could reduce power consumption by approximately 30% has not been independently verified 19. NVIDIA should benefit if these technologies reduce the power and cooling burden of high-bandwidth memory and advanced packaging. Vendor claims should not be treated as realized platform-level savings until they are validated in production systems.
Cooling Economics: Water, Energy, and Capital
Cooling decisions cannot be reduced to a single water-use metric. In recirculating or closed-loop systems, coolant can repeatedly circulate through pipes and heat exchangers 41, and cooling water is generally not consumed in the same manner as water used for direct consumption 12. Once-through or open systems may return warmed water with only minor evaporative losses 12. Nuclear plants in some countries use recirculating cooling towers that dissipate heat through air 12, while closed cooling at Stenungsund chemical facilities both dissipates heat and reduces water usage 12.
The underlying constraint remains site-specific. Evaporative cooling can be energy-efficient while consuming substantial water 48, and water-based cooling can consume significant quantities through evaporation or dissipation 37. Eliminating evaporative cooling does not necessarily reduce total environmental water consumption if more electricity-intensive systems increase indirect water use through power generation 11. Low-water alternatives may require more electricity and higher construction costs 41. The contrast between Nebius’s reported 0.018 liters per kWh and a 0.47-liter benchmark 32, and xAI’s reported use of more than one million gallons per day at Colossus 38, demonstrates the dispersion in operating profiles.
Regulatory and social-license pressure will shape where GPU capacity can be deployed. A Chilean court required review and redesign of a proposed data-center cooling system from water cooling to air cooling 17. Google has committed to advanced air cooling at Tarluvada to reduce water use 39,44. These examples coexist with evidence that air cooling can increase electricity demand and emissions 1, and that lower-water systems may raise capital expenditure. The relevant commercial metric is therefore total resource intensity: electricity, water, capital, and reliability. Water consumption in isolation is an incomplete measure.
Waste Heat Is a Potential Value Stream, Not Yet a Proven One
Waste-heat recovery is emerging as an infrastructure opportunity. Excess data-center heat can be recovered through heat pumps or heat exchangers for district heating 12, particularly where facilities can connect to existing district-heating networks 12. Nebius’s Finland operations are associated with high-temperature operation, renewable energy, advanced cooling, and heat reuse 3. Its proposed model would direct waste heat to homes and communities 3. EU Directive 2023/1791 also encourages data centers to recover waste heat 14.
A more integrated proposal links data-center HVAC waste heat through liquid-to-air loops to controlled-environment vertical farming 6. The model reports a 68.4% reduction in vertical-farm thermal operating expenditure and a 31.2% reduction in total unit production cost 6,7. These are modelled, single-source claims rather than demonstrated NVIDIA-relevant economics. The concept is nevertheless strategically relevant because it turns cooling infrastructure into an energy-integration mechanism rather than a pure cost center 7.
The margin here is dangerously thin. Heat reuse depends on geography, temperature grade, network connectivity, seasonal demand, heat-pump economics, and contractual arrangements. The economic significance of Nebius’s heat-reuse activity remains uncertain without disclosed contracts or cost data 3. For NVIDIA, waste-heat recovery should be treated as an enabling feature that may improve customer permitting and operating economics, not as a direct material revenue stream.
Execution, Measurement, and Platform Implications
Thermal architecture is only useful if the surrounding infrastructure is delivered on schedule. Large AI facilities require coordination with utilities, PJM, PPL, NJDEP, planning commissions, equipment suppliers, and other authorities 33. Nebius’s contracted power exceeds 3.5 GW, but contracted capacity is not necessarily energized or revenue-producing 33. Its Vineland project reportedly exceeds 300 MW and requires behind-the-meter generation 33. Delays in permitting, equipment delivery, construction, or utility upgrades can delay both energization and revenue recognition 33. Similar timing risks affect power and cooling projects involving Amphenol 21.
Measurement quality is becoming a competitive differentiator. Facility power is the sum of IT, cooling, electrical, and auxiliary loads 15. Thermal KPIs can diagnose recirculation, airflow imbalance, insufficient heat removal, and overcooling 4. Integrated monitoring can synchronize inlet and outlet temperatures, CPU temperatures, server power, fan speeds, and cooling-unit data 4. Dashboard visibility alone, however, does not guarantee accurate thermal or energy assessment 4. Internal sensors may also have coarse granularity or calibration drift 16.
This creates an opening for NVIDIA to make thermal telemetry, workload scheduling, and system optimization part of the broader accelerated-computing platform. The objective is not simply to report nominal GPU performance. It is to validate delivered compute under the actual power, cooling, and uptime conditions of the facility.
Implications for NVIDIA and Investors
Thermal management is becoming a market-enabling layer for accelerated computing. NVIDIA does not need to become a cooling-equipment manufacturer for this transition to matter. The value of its GPUs increasingly depends on a qualified, integrated system capable of delivering power, cooling, networking, and uptime at scale. Liquid cooling can remove substantially more heat than air cooling and may reduce cooling energy and improve PUE 23,46, but it also requires a broader bill of materials and greater systems integration.
The competitive landscape should therefore be assessed at the platform level. CDU, cold-plate, rack, power-distribution, monitoring, and heat-rejection suppliers may capture incremental value as AI factories devote a substantial share of physical space to “grey space” rather than IT “white space” 45. Flex’s JetCool acquisition added cold-plate capability, and its CDUs are undergoing customer qualification 2, but its liquid-cooling revenue remains unproven at scale 2. Vertiv is expanding testing and coordination capabilities, including large-chiller testing under high-density loads and centralized cooling management 18,23. These suppliers can improve NVIDIA system deployment while also creating qualification bottlenecks and potential margin-sharing across the AI infrastructure stack.
There is a second-order tension between efficiency and demand. More efficient chips, power architectures, workload scheduling, and model compression can reduce energy per computation 34,35,40,42. Lower compute cost, however, can stimulate additional AI usage, preserving or increasing aggregate demand for NVIDIA accelerators. Higher rack density similarly reduces facility footprint while intensifying cooling, reliability, and capital requirements. Lower-density racks may provide better server efficiency and cooling performance than designs optimized solely for density maximization 5. Customers will therefore optimize for delivered compute per dollar and per unit of power, not simply GPUs per rack.
The investment view is constructive but conditional. Thermal innovation can broaden NVIDIA’s long-term market by making high-density AI deployment feasible across more sites and by supporting higher-performance architectures. The principal risks are delayed infrastructure build-outs, insufficient grid or cooling capacity, fluid and environmental regulation, unverified vendor efficiency claims, and a shift toward more compute-efficient architectures that reduces accelerator demand at the margin.
Investors should monitor five indicators: production qualification of liquid-cooling systems; deployed rack power densities; customer PUE and water intensity; the share of revenue tied to fully integrated systems; and evidence that waste-heat reuse produces contractual value. The strongest conclusion is structural. Thermal management is now a prerequisite for NVIDIA’s growth trajectory, while economic capture across the surrounding ecosystem remains unsettled. The industry has once again confused a press release with a production timeline; the decisive evidence will come from energized facilities, qualified systems, and measured operating performance.
Key Takeaways
- Thermal capacity is becoming a prerequisite for AI capacity. Direct liquid cooling becomes compelling around 20–30 kW per rack, while air cooling becomes increasingly constrained above roughly 30–40 kW 13,45.
- The opportunity is platform-level. CDUs, cold plates, power delivery, monitoring, heat rejection, and facility integration are likely to matter as much as standalone cooling hardware 30,47.
- Water efficiency is not a standalone ESG metric. Closed-loop and air systems can reduce direct water use but may increase electricity consumption, capital cost, or indirect water impact 11,41.
- NVIDIA’s upside is substantial but execution-dependent. Thermal innovation can unlock higher-density deployments, while permitting, retrofit, qualification, reliability, and unverified efficiency claims remain binding constraints 13,33,43.