The Physics Problem Nobody Could Ignore Forever
Data centre cooling has always been a physics problem dressed up as an engineering challenge, and for most of the industry’s history the engineering solutions were good enough to keep pace with what physics demanded. Raised floors channelled chilled air through hot-aisle containment. Precision air conditioning units maintained the ambient temperature envelopes inside which server hardware could operate reliably. Computer room air handlers cycled billions of cubic feet of conditioned air every hour across facilities that consumed power as aggressively as small cities. The entire architecture rested on a comfortable assumption: that the thermal density of computing equipment would remain within a range that moving air could manage. That assumption held, with periodic strain, through three decades of Moore’s Law-driven density increases. It did not survive the arrival of AI accelerators at hyperscale. When Nvidia launched its Blackwell B200 GPU commercially in early 2025, delivering a thermal design power of up to one kilowatt per chip, and when successor architectures on the Rubin Ultra platform are anticipated to push per-chip TDP beyond 1,400 watts by 2027, the gap between what air can move and what heat these chips produce became a structural design constraint rather than a temporary challenge requiring incremental adjustment.
The response from American hyperscalers was fast and, by historical standards, decisive. AWS, Microsoft Azure, Google Cloud, and Meta publicly committed to multi-billion-dollar liquid-cooled AI campus expansions across Northern Virginia, Iowa, Phoenix, and the Pacific Northwest, accelerating a shift in facility design philosophy that had been building for years but that the AI acceleration wave compressed into a three-year transition period. Direct-to-chip liquid cooling — in which coolant plates contact hot components directly without submerging the entire server — consolidated as the dominant architecture for current-generation AI clusters across these campuses. The appeal of direct-to-chip is that it fits within existing data centre form factors, works with standard server designs from major original equipment manufacturers, and supports rack densities between 40 and 150 kW without requiring the facility-level redesign that more radical alternatives demand. It solved the immediate problem that the Hopper and early Blackwell generations created without asking hyperscalers to bet their facility roadmaps on an entirely unproven operational model. Behind that consolidation, a quieter and structurally more significant competition is now underway between the two architectures that claim to handle what direct-to-chip cannot: full immersion cooling.
Why Liquid Cooling Became Inevitable, Not Optional
The transition from air to liquid cooling is not a technology preference — it is a thermal physics mandate that the semiconductor industry’s roadmap turned into a deployment requirement. Nvidia’s B200 GPU alone demonstrates a TDP of 1,200 watts, while Intel’s next-generation Jaguar Shores processor was anticipated to match or exceed similar power levels in 2025 and 2026, and planned AI accelerator generations beyond these will push chip-level heat generation into ranges that no conventional air cooling architecture can dissipate without unacceptable energy penalty and floor space consumption. Traditional air cooling at these densities requires so much airflow volume and conditioning that the infrastructure supporting the cooling consumes a significant fraction of the power that the servers themselves draw a PUE spiral that makes high-density air-cooled AI clusters economically irrational at the scale hyperscalers are building. Liquid coolants conduct heat up to a thousand times more effectively than air under comparable conditions, meaning that a liquid cooling system can remove the same heat load with dramatically less energy and infrastructure footprint than its air-cooled equivalent. Dell’Oro Group’s January 2026 Data Center Liquid Cooling Advanced Research Report identified the liquid cooling market as reaching close to three billion dollars in 2025, doubling in a single year, and projected continued scaling toward approximately seven billion dollars by 2029. That trajectory does not represent an optional technology adoption curve it represents the infrastructure industry catching up with the physics problem that AI accelerator manufacturers created on a product roadmap that did not wait for the cooling ecosystem to follow.
Microsoft’s Fairwater AI campus programme illustrates what this transition looks like in practice for a hyperscaler that has committed to it at scale. The Atlanta site, which became operational in October 2025, and the Wisconsin follow-on, expected online in early 2026, use closed-loop liquid cooling systems that eliminate operational water consumption addressing simultaneously the thermal management requirement and the water controversy that has accompanied conventional evaporative cooling architectures in markets like Johor and Phoenix. Oracle adopted a comparable approach, deploying closed-loop, direct-to-chip cooling specifically to reduce water consumption and address community concerns associated with traditional evaporative systems. These facility design decisions are not cosmetic sustainability moves they reflect a recognition that the operational water consumption model that accompanied air-plus-evaporative-cooling architectures at scale was generating genuine community and regulatory friction in the US markets where hyperscalers need to build most urgently. Liquid cooling, in its closed-loop direct-to-chip form, resolves both the thermal problem and the water problem simultaneously, which explains why it consolidated so quickly across the major American AI campus programmes of 2025 and 2026.
Single-Phase Immersion: The Architecture That Solved the Present
How Mineral Oil Became a Data Centre Medium
Single-phase immersion cooling submerges entire server boards in a thermally conductive, electrically non-conductive dielectric fluid — typically a synthetic hydrocarbon, engineered ester, or mineral oil derivative that remains in liquid state throughout the entire heat absorption and rejection cycle. The heat that server components generate transfers directly to the surrounding fluid, which circulates to an external heat exchanger where the thermal load is rejected to the environment or captured for heat reuse applications. No fans are required inside the tank. No air conditioning units manage ambient temperature within the server enclosure. No chilled water loop chases a temperature differential across a CRAC unit floor space that could otherwise carry revenue-generating compute. Green Revolution Cooling, the Austin-based company that has operated as the immersion cooling authority across twenty-one countries, builds its entire product philosophy around this architecture arguing that eliminating the need for chillers, computer room air conditioning, air handlers, humidity controls, and conventional cooling components simultaneously reduces design cost, build cost, energy cost, and maintenance overhead in a single infrastructural decision. GRC and UNICOM Engineering expanded their collaboration in March 2026 to deliver turnkey immersion-cooled solutions specifically for AI, HPC, and GPU-as-a-Service deployments in production environments ranging from large campuses to factory production lines, remote depots, and distributed warehouse environments a scope that reflects how far the single-phase architecture has moved from its origins in bespoke high-performance computing installations toward a standardised, channel-delivered infrastructure product.
The operational logic of single-phase immersion is straightforward and, by the standards of frontier data centre technology, well-validated. Mineral oil at entry-level costs between nineteen and thirty-eight dollars per gallon, making it accessible for operators willing to accept its functional trade-offs, while engineered single-phase fluids offering better dielectric properties and lower viscosity run from one hundred fifty to four hundred dollars per gallon. Single-phase systems consistently support rack densities of two hundred kilowatts per rack using mineral oil and considerably higher with purpose-engineered fluids densities that direct-to-chip systems reach only at the upper end of their operational envelope and that air cooling cannot approach without prohibitive energy penalty. The primary operational friction in single-phase immersion derives from the mineral oil’s solvent properties: continuous fluid flow across server components can, over extended operation, dissolve labels, identification markings, and printed circuit board flux materials that become suspended in the oil and progressively contaminate it, reducing its dielectric integrity and eventually requiring fluid replacement at both cost and downtime. Engineered fluid formulations and better server designs produced for immersion environments address this problem through chemistry and hardware modification rather than architectural change, but the operational discipline required to manage a large-scale immersion deployment remains meaningfully different from the discipline governing a direct-to-chip or air-cooled facility, which is one reason single-phase adoption at hyperscale has proceeded more slowly than its thermal advantages might suggest.
The Density Advantage and Its Real-World Ceiling
The density advantages that single-phase immersion delivers in a controlled lab or benchmarking environment do not transfer directly into hyperscale deployment without resolving a set of operational challenges that become larger as the scale of deployment increases. Server hardware designed for air-cooled environments does not drop into a single-phase immersion tank without modification optical modules, storage devices, and power supplies all require redesign or replacement with immersion-compatible variants, and the supply chain for immersion-optimised components at hyperscale remains thinner and slower than the supply chain for standard air-cooled or direct-to-chip components. NADDOD’s analysis of single-phase versus two-phase architectures in commercial data centre environments confirms that single-phase systems retain the largest market share due to installation familiarity, but notes that the ecosystem of component suppliers designing specifically for immersion deployment is still building to the depth that hyperscale procurement timelines require. Server withdrawal and maintenance in single-phase mineral oil environments requires allowing fluid to drain from the withdrawn hardware before service personnel can work on it, creating a multi-hour maintenance window for any component-level intervention — an operational model that suits certain deployment patterns but that conflicts with the continuous availability requirements of active AI training clusters. GRC’s product roadmap heading into 2026, as described by the company’s leadership, envisions two possible trajectories: acquisition by a data centre infrastructure corporation seeking to integrate the patented immersion technology into a broader thermal management portfolio, or independent growth toward a company with a diversified liquid cooling product range, regional engineering teams, and potentially a technology licensing programme. That strategic optionality reflects a company operating at the frontier of a market that has validated its technology but has not yet resolved the deployment model at hyperscale.
What the single-phase architecture has accomplished, beyond its specific technical merits, is to establish the fundamental vocabulary of a new facility design paradigm. Data centres built around single-phase immersion do not look or operate like their air-cooled predecessors. Tank rows replace rack rows. External heat rejection systems replace chiller plants and cooling towers. Fluid management infrastructure distribution manifolds, heat exchangers, monitoring systems for fluid quality and dielectric integrity replaces the dense array of CRAC units and air handlers that define conventional data centre cooling plant. The capital investment required to build this infrastructure from scratch, rather than to adapt existing facilities, is one reason the immersion transition has proceeded primarily through greenfield campuses rather than retrofits of legacy sites. For American hyperscalers building new AI-specific campuses in Phoenix, Iowa, and Prince William County that are designed from the foundation layer up, the argument for immersion-ready design is stronger than for operators managing large existing footprints where the infrastructure sunk cost in conventional cooling systems is a genuine transition barrier.
Two-Phase Immersion: Theoretically Superior, Practically Complicated
The Promise of Boiling Coolant
Two-phase immersion cooling represents the most thermodynamically efficient approach to data centre heat rejection that current engineering can deliver. It submerges server hardware in a dielectric fluid with a low boiling point, allowing the chip’s own heat to vaporise the fluid directly at the component surface a mechanism called nucleate boiling that transfers heat with exceptional efficiency because the phase transition from liquid to vapour absorbs substantially more energy per unit of mass than simple convective heating of a liquid. The vapour rises through the tank, contacts a condenser surface at the top of the enclosure, returns to liquid form, and falls back onto the submerged hardware in a continuous passive cycle that requires no pump to circulate the primary coolant. This pump-free architecture reduces mechanical complexity and eliminates a category of potential failure, while the phase-change heat transfer mechanism allows two-phase systems to handle thermal flux densities that exceed what single-phase fluid circulation can manage making it the architecture of choice wherever chip-level heat generation reaches the extreme upper bound of what any liquid cooling approach must address. Power usage effectiveness ratios approaching the theoretical minimum of 1.0 are achievable under well-engineered two-phase conditions, which makes the technology genuinely compelling for operators whose energy costs and sustainability commitments make PUE reduction a primary design constraint. Uptime Institute’s analysis of the competitive landscape confirmed that two-phase immersion had attracted substantial renewed interest precisely as runaway silicon thermal power pushed direct-to-chip systems toward their own density limits.
The operational case for two-phase immersion is not, however, primarily a question of physics. Physics strongly favours it. The challenge is the supply chain for the fluids that make the technology work, which 3M’s December 2022 announcement comprehensively destabilised and which has not recovered to the degree that large-scale commercial deployment requires. Accelsius launched the NeuCool IR150 at Data Center World 2026 as the industry’s first fully integrated rack-level two-phase direct-to-chip cooling solution a product positioned to bring two-phase technology to hyperscalers and enterprise operators in a standardised, channel-delivered form factor and simultaneously announced the NeuCool HyperStart program, through which hyperscale operators and neocloud providers can validate two-phase direct-to-chip liquid cooling solutions for high-density, large-scale deployments. Several hyperscale AI cloud providers engaged with Accelsius under the programme as they built cooling roadmaps for next-generation AI infrastructure. That engagement reflects genuine interest at the hyperscale level in two-phase technology, but the cautious framing validation programmes, cooling roadmaps accurately characterises the state of commercial readiness: serious evaluation, not committed deployment at the scale of direct-to-chip or single-phase implementations.
The PFAS Crisis and the Supply Chain Collapse
The single most consequential event in the commercial history of two-phase immersion cooling was not a data centre fire, a hyperscaler procurement decision, or a technology failure — it was a legal settlement. On 20 December 2022, 3M announced it would stop manufacturing all per- and polyfluoroalkyl substance chemicals by the end of 2025, facing more than four thousand lawsuits and a twelve-and-a-half-billion-dollar settlement with more than eleven thousand US public water systems alleging PFAS contamination in drinking water. The last date to place a new Novec order was 31 March 2025. Novec 7100, Novec 649, and Fluorinert FC-72 — the fluorocarbon fluids that made two-phase immersion cooling thermodynamically elegant and practically deployable at scale exited commercial availability on a court-driven timeline that had nothing to do with the data centre industry’s readiness for an alternative. For operators and vendors who had spent years and significant capital validating two-phase immersion deployments using Novec-family fluids, 3M’s exit created what Motivair Cooling described as an immediate risk of obsolescence. The US Environmental Protection Agency’s classification of certain PFAS substances as hazardous materials added a regulatory layer to the supply chain problem, signalling that the legal and regulatory environment for PFAS-based fluids in commercial applications would not improve even if alternative production sources materialised.
The search for viable PFAS replacements is underway at several chemical companies, including Chemours, Solvay, Honeywell, and Envirotech, but progress has been materially slower than the timeline created by 3M’s exit. Solvay’s Galden, a commonly cited Novec replacement, is itself a PFAS compound subject to the same regulatory trajectory that made Novec commercially untenable meaning it functions as a transitional fluid under existing regulations rather than a long-term solution to the underlying chemistry problem. Schneider Electric’s analysis of the PFAS phase-out and its implications for US data centre operators, published in May 2026, concluded that while natural alternatives to PFAS are under active research, the pace of progress is slow and the pressure is intense. The immersion cooling fluids market reflects this tension directly in its pricing structure: engineered single-phase fluids fill the gap that two-phase PFAS exit created for operators who need density above mineral oil’s ceiling, running at one hundred fifty to four hundred dollars per gallon, while two-phase alternative fluids from Chemours and its competitors have not yet reached the volume production scales at which pricing would support widespread commercial deployment. Two-phase immersion cooling carries the best thermal physics of any commercially evaluated data centre cooling architecture, and its commercial trajectory is contingent on a chemistry problem that the semiconductor industry, not the data centre industry, will ultimately need to solve.
The Competitive Landscape: Who Is Building the Immersion Market
Vendors Competing for a Market That Is Still Taking Shape
The competitive structure of the immersion cooling market reflects its developmental stage: a combination of specialist vendors who built the technology before hyperscaler demand arrived, established infrastructure companies acquiring those specialists as the market becomes strategically significant, and integrated AI infrastructure providers absorbing cooling capability as part of a broader platform play. Trane Technologies’ acquisition of LiquidStack, announced in February 2026, is the clearest expression of the second dynamic a major thermal management corporation acquiring a liquid cooling specialist to expand its end-to-end portfolio and production capacity specifically for hyperscale and AI-driven data centre deployments. Trane’s entry signals that the liquid cooling market has graduated from a specialist niche to a mainstream thermal management category that large industrial companies regard as strategically essential to their data centre infrastructure business, alongside chiller plants, cooling towers, and building management systems. Submer Technologies, the Barcelona-founded immersion cooling specialist with significant US commercial presence, responded to the same market signal from a different direction transforming from a cooling specialist into what the company describes as a full-stack AI infrastructure provider, integrating modular data centres, GPU compute, and sovereign cloud services into a single product line that addresses the entire infrastructure stack above the facility layer. Submer’s partnership with Anant Raj in India, positioned around national AI sovereignty goals, illustrates the broader ambition: immersion cooling as the physical foundation of a complete AI infrastructure offering rather than a component within someone else’s facility design.
GRC operates as the most mature single-phase immersion specialist in the market with deployments across twenty-one countries, approaching the commercial scale question from a product and manufacturing perspective rather than a platform integration perspective. The company’s March 2026 expansion of its collaboration with UNICOM Engineering delivering fully validated, production-ready immersion-cooled solutions that remove multi-vendor integration complexity addresses a genuine barrier to hyperscale adoption: the difficulty of sourcing, integrating, and validating immersion cooling systems from multiple component vendors within the procurement and commissioning timelines that hyperscale campus buildouts impose. By delivering a turnkey solution combining UNICOM Engineering’s immersion-optimised computing platforms with GRC’s patented single-phase cooling systems, the collaboration offers hyperscalers and neocloud operators a single-vendor accountability model that is more compatible with large-scale procurement than a best-of-breed assembly from multiple specialists. GRC’s product roadmap also includes the ICEraQ Nano, launched in November 2025, targeting edge deployments in small data rooms and communications closets a move toward the distributed end of the compute topology that signals the company’s view that immersion cooling’s commercial future extends well beyond hyperscale campuses into a broader infrastructure ecosystem.
What Dell’Oro’s Architecture Projection Actually Means
Dell’Oro Group’s January 2026 advanced research report on the data centre liquid cooling market made a projection that contains more information in its qualifications than in its headline numbers. Single-phase direct liquid cooling is expected to remain the prevailing architecture for most deployments through the end of the decade, the report found, with two-phase direct liquid cooling expanding gradually, with adoption accelerating once chip-level TDPs and thermal flux exceed the practical limits of single-phase systems. Immersion cooling, the report noted separately, is finding its place through selective adoption where its architectural trade-offs are justified by performance or operational requirements. Decomposing what this projection actually implies for American hyperscale campus design requires attention to the phrase “practical limits of single-phase systems” — because those limits are not fixed. They are a moving target set by GPU manufacturer roadmaps, and the Rubin Ultra platform anticipated to push per-chip TDP beyond 1,400 watts by 2027 may cross single-phase system thresholds well before 2030, rendering the “end of the decade” framing for single-phase dominance more compressed than a linear reading suggests. The competitive dynamics that Dell’Oro identifies Vertiv leading the liquid cooling market with CoolIT, nVent, and Boyd maintaining strong positions reflect a market where established infrastructure companies have moved faster than immersion specialists to commercialise direct-to-chip solutions, creating the market share gap that single-phase immersion vendors are competing to close.
The ecosystem is also undergoing what Dell’Oro characterises as rapid change from new entrants, expanding vendor capabilities, and increased merger-and-acquisition activity reshaping the competitive landscape. Trane’s LiquidStack acquisition, Submer’s platform expansion, GRC’s UNICOM partnership, and Accelsius’ NeuCool product launch all occurred within a compressed twelve-month window, reflecting the industry’s collective recognition that the cooling architecture decisions being made in 2025 and 2026 will lock in facility design patterns for the AI campuses being built through 2030. Vertiv, the established leader, faces a more competitive landscape heading into the second half of the decade than it did when direct-to-chip consolidated its early market position. The outcome of the architecture competition between single-phase immersion, two-phase immersion, and direct-to-chip will not be determined by any single technology’s thermal performance superiority — it will be determined by the combination of fluid supply chain resolution, server OEM design compatibility, operational management maturity, and the pace at which the PFAS alternative chemistry matures into a commercially reliable supply base.
The Path from Pilot to Production
What Winning at Hyperscale Actually Requires
The immersion cooling market as it stands in 2026 operates in a state that its participants variously describe as validation, early large-scale implementation, and selective adoption all of which are accurate, and all of which describe a technology that has demonstrated its core thermal performance claims but has not yet resolved the operational, supply chain, and hardware compatibility challenges that hyperscale procurement requires. Dell’Oro’s assessment that immersion cooling is finding its place through selective adoption where architectural trade-offs are justified is the clearest formulation of where the market sits: valuable in specific contexts, not yet universally deployable at the speeds and volumes that American hyperscaler campus construction timelines demand. The path from this position to production-scale adoption across the industry requires several parallel progressions that cannot wait for each other sequentially fluid supply chain development, server OEM programme expansion, operational management framework maturation, and facility design standardisation must all advance simultaneously rather than in series. The PFAS resolution challenge is the most externally constrained of these progressions, because it depends on chemical engineering timelines at Chemours, Honeywell, and their competitors that data centre demand can accelerate through purchasing commitment but cannot compress through facility design decisions.
The vendors that will win the American immersion market are those that solve the deployment complexity problem rather than simply the thermal performance problem. GRC’s turnkey collaboration with UNICOM Engineering, Accelsius’ NeuCool HyperStart validation programme, and Submer’s full-stack platform integration all represent attempts to reduce the friction between a hyperscaler’s decision to evaluate immersion cooling and its ability to deploy a validated, production-ready system within a campus buildout timeline. That friction reduction is commercially more valuable than incremental PUE improvement for operators who have already committed to liquid cooling and are now deciding between direct-to-chip and immersion architectures for their next facility generation. The operators most likely to make full immersion their standard architecture first are those building genuinely new campuses in Phoenix, in Iowa, in the Pacific Northwest markets where land acquisition and grid connection are the primary timeline constraints rather than operators managing large existing footprints where the transition cost from legacy air cooling to immersion represents a sunk cost calculation that direct-to-chip sidesteps more cleanly.
The Architecture Decision That Will Define the Decade
The competition between direct-to-chip, single-phase immersion, and two-phase immersion cooling for American AI campus architecture is not a transient technology choice that will resolve cleanly by one architecture winning outright. It is a layered market that will distribute across architecture types based on workload characteristics, facility vintage, operator risk tolerance, and the pace at which the two central challenges PFAS alternative fluid availability and server OEM immersion compatibility resolve into commercial products that hyperscale procurement teams can rely on. Direct-to-chip will remain dominant through the current GPU generation because the supply chain is mature, the server compatibility is established, and the hyperscaler validation evidence is extensive. Single-phase immersion will gain share among new AI-native campuses and neocloud deployments where density requirements and greenfield facility design justify the operational model change. Two-phase immersion’s commercial trajectory depends most critically on the PFAS alternative chemistry resolving into volume production from Chemours and its competitors a condition that has not been met as of mid-2026 but that the thermal density trajectory of AI accelerator hardware makes commercially urgent in a way that no earlier application context did.
The GPU accelerator manufacturers are, more than any data centre operator or cooling vendor, the force that will determine when and how this architecture competition resolves. Each successive GPU generation that pushes per-chip TDP higher narrows the operating envelope within which direct-to-chip maintains its density ceiling without crossing into territory where immersion’s performance advantage becomes economically decisive. The Rubin Ultra platform’s anticipated TDP levels will test whether single-phase immersion can absorb the load or whether two-phase becomes unavoidable for the highest-density AI training clusters. American hyperscalers are building the campuses that will house those accelerators right now, in Prince William and Fauquier counties, in Des Moines and Phoenix and the Columbia River corridor — and the cooling infrastructure decisions embedded in those facility designs will either accommodate the architecture evolution that accelerator roadmaps will require or will force costly mid-lifecycle retrofits in facilities that will still be operating when the next GPU generation arrives. The vendors and operators that understand this constraint most clearly are those building the most flexibility into their facility designs today, and that flexibility, more than any specific technology bet, is the asset that the American AI infrastructure industry most needs.
