The Infrastructure War That Happens Inside the Pipe
The dominant narratives of AI infrastructure in 2026 are stories about scale: the size of capital commitments, the megawatt count of new campuses, the gigawatt ambitions of data centre pipelines extending across continents. These are real and consequential stories. But running parallel to the visible infrastructure race, and largely invisible to the analysts, journalists, and executives following it, is a different kind of contest one fought not at the level of the data centre campus or the GPU cluster but inside the pipes, cold plates, manifolds, and tanks that carry thermal energy away from the chips that generate it. The coolant chemistry war is a battle over fluid standards, proprietary thermal architectures, supply chain control, and regulatory chemistry that will determine as much about the long-term economics and interoperability of AI infrastructure as any of the more prominent debates about power and compute.
The proximate cause of this war is physics. NVIDIA’s B200 GPU, shipping at scale through 2025 and 2026, consumes up to 1,000 watts under air cooling and 1,200 watts when liquid cooled a thermal design power that air-cooling systems, which reached their practical engineering ceiling at roughly 40 to 60 kilowatts per rack, cannot manage at the densities that AI deployments require. Rack power densities in hyperscale AI facilities are now routinely running between 80 and 140 kilowatts, with some GB200 NVL72 deployments approaching higher densities still. Liquid cooling has moved from an option to a necessity, and the entire data centre industry from silicon designers to facility operators, from chemical formulators to cooling hardware manufacturers is reorganising around the thermal requirements of a chip generation that air simply cannot cool.
Liquid cooling reached 22% adoption in new data centre builds during 2025, and every chip exceeding 700 watts of thermal design power now requires liquid cooling as a practical operational necessity, according to Introl’s custom silicon analysis. The market for data centre cooling is projected to reach $248 billion by 2035. The immersion cooling fluids segment alone is expected to grow from $2.79 billion in 2026 to $4.49 billion by 2031. These are large numbers generating competitive pressure at every point in the supply chain, and the decisions being made now about which coolant chemistries will be standardised, which architectures will become proprietary, and which organisations will control the service layer of a liquid-cooled data centre estate are decisions that will shape the economics and operational independence of the industry for a decade.
The Open Compute Project’s Standardisation Bet
PG25 and the Case for Universal Coolant Chemistry
The Open Compute Project has been the technology industry’s most sustained institutional attempt to prevent the data centre supply chain from fragmenting into a collection of proprietary, vertically integrated silos. Founded by Facebook now Meta in 2011, OCP has published open specifications for servers, racks, networking equipment, and power systems that have collectively reduced the cost and complexity of hyperscale infrastructure by enabling multi-vendor competition for standardised components. When liquid cooling began its ascent from niche technology to operational necessity, OCP moved to extend the same standardisation philosophy to the thermal management layer. The result was the PG25 standard a specification for propylene glycol-based heat transfer fluids in direct-to-chip cold plate systems that defines wetted-material compatibility, manifold and tubing requirements, operating temperature and pressure ranges, filtration requirements, and safety practices with the explicit goal of enabling multi-vendor interoperability across IT refresh cycles.
Propylene glycol chosen at a 25% concentration is not an arbitrary selection. Glycol-water blends serve three simultaneous functions in direct-to-chip cooling loops: they depress the freezing point of the working fluid below zero Celsius for cold-climate installations, they inhibit corrosion across the diverse metallic surfaces copper, aluminium, stainless steel, brass, nickel — that a cooling circuit contacts, and they carry biocide packages that prevent microbial growth in systems that circulate warm fluid continuously. Water alone has superior heat transfer properties, with volumetric heat capacity nearly 3,500 times that of air, but it provides no freeze protection and accelerates galvanic corrosion between dissimilar metals in mixed-material circuits. Propylene glycol, unlike ethylene glycol, is non-toxic, making it the preferred formulation for environments where incidental contact with workers or exposure through facility drainage requires a lower-hazard profile. OCP’s guidelines specify that PG25 coolant should maintain a flow rate of approximately 1.5 litres per minute per kilowatt with a target temperature rise of 10 degrees Celsius across the cold plate engineering parameters that define a common operating envelope across which different vendors’ hardware can interoperate without bespoke fluid qualification.
At OCP’s 2025 Global Summit, the practical commercial momentum behind PG25 was on visible display. Hardware firms across Taiwan, Europe, and North America embraced the newly ratified PG25 coolant material compatibility standards as a design anchor. Shell highlighted its Direct Liquid Cooling Fluid S3, developed explicitly for multi-vendor reliability and full warranty compliance. Castrol ON announced in May 2026 that its PG25 direct liquid cooling fluid had been recognised under OCP’s OCP Inspired programme, with two additional immersion cooling fluids expected to follow. Dober’s COOLWAVE DC-25 heat transfer fluid carries OCP Inspired designation. Proviron’s Proviflow DLC 25 is positioned for compatibility with existing glycol installations including legacy Zitrec and Antifrogen formulations. The ecosystem building around PG25 represents something important: the emergence of a commodity coolant market in which multiple competing suppliers can qualify products against published specifications, breaking the dependency on single-source proprietary fluids that characterised earlier generations of specialised cooling applications.
The OCP-ASHRAE alliance formed in October 2025 took the standardisation effort a layer deeper. ASHRAE’s TC 9.9 technical committee the body that has defined thermal guidelines for data centre facilities for decades and OCP formalised a collaboration spanning the full cooling chain from facility-level water systems through rack-level coolant distribution units to chip-level cold plates. ASHRAE’s 2024 technical bulletin on liquid cooling resiliency explicitly recommends separating facility water systems from technology cooling systems via a coolant distribution unit, a design pattern that OCP’s specifications operationalise through reference designs for CDU interfaces, manifold connections, and fluid loop isolation. This separation is not merely a convenience — it is a risk management architecture that prevents facility water chemistry from contaminating the technology cooling loop and prevents coolant leaks in the IT space from propagating into the facility’s primary water infrastructure.
The Proprietary Counter-Movement: When Hyperscalers Design Their Own Thermal Systems
Microsoft’s Fairwater and the Silicon-Cooling Co-Design Imperative
The OCP standardisation effort operates within a market that its largest and most influential members are simultaneously pulling in the opposite direction. Microsoft, Google, Amazon, and Meta are not simply deploying commodity cooling infrastructure — they are designing proprietary thermal systems co-engineered with their custom silicon, creating cooling architectures whose performance advantages depend on integration with specific chip geometries and which cannot be replicated by any off-the-shelf fluid or hardware combination. The logic is internally coherent and commercially rational: when a company designs its own AI accelerator from the circuit level, the thermal interface between the chip and the cooling system is an engineering variable that can be optimised rather than a constraint that must be accepted. Leaving that optimisation to a commodity cold plate and a generic propylene glycol formulation means accepting a thermal performance penalty that a co-designed system can eliminate.
Microsoft’s Fairwater system, developed alongside the Maia 100 AI accelerator, is the clearest public example of this co-design philosophy. Fairwater uses direct-to-chip liquid cooling with cold plates designed specifically for the Maia silicon’s thermal map — the spatial distribution of heat generation across the chip surface that determines where cooling must be most intense and where moderate thermal management suffices. By designing the cold plate channels to match the Maia chip’s specific hot spots rather than optimising for a generic chip geometry, Microsoft achieves lower junction temperatures, higher sustained clock speeds, and reduced thermal throttling compared to a generic cold plate solution. The system also supports NVIDIA GPU racks, which gives it commercial flexibility without undermining the core co-design advantage for Microsoft’s own silicon. Microsoft went further still in September 2025 by announcing a collaboration with Swiss startup Corintis on microchannel cooling technology — etching cooling channels directly onto the back of the silicon die in patterns resembling leaf veins, inspired by natural heat distribution geometries, to achieve heat removal rates that conventional cold plate designs cannot approach. TechInsights analyst Manish Rawat described the engineering challenge precisely: fabricating micron-scale channels increases process complexity, may raise yield loss due to wafer fragility, and requires ultra-reliable sealing where even minor leaks could degrade chip performance.
Google’s approach to cooling co-design follows the same principle at a different architectural scale. Google has deployed liquid cooling across more than 2,000 TPU Pods, making it one of the largest operational liquid-cooled AI deployments in production anywhere in the world. The Ironwood TPU Google’s seventh-generation tensor processing unit, unveiled at Cloud Next 2025 is designed from the ground up for liquid cooling, with thermal design points that assume direct-to-chip cooling as the baseline operating condition rather than as an enhancement. Google’s Optical Circuit Switching interconnect architecture, which allows Ironwood chips to operate in unified clusters of nearly 10,000 units, generates thermal management requirements at the cluster level not just the chip level that a generic cooling specification cannot address without performance compromise. Amazon Web Services developed its own custom liquid cooling solution in approximately eleven months, achieving material reduction in mechanical energy consumption during peak cooling through an in-house coolant distribution unit designed to maximise flow and minimise energy use at the system level. Meta’s Air-Assisted Liquid Cooling systems with rear-door heat exchangers currently support rack densities up to 40 kilowatts, with the company’s roadmap explicitly planning a transition to facility water cooling as thermal loads increase with successive GPU generations.
The Interoperability Problem That Co-Design Creates
The commercial logic of thermal co-design is sound when applied within a single hyperscaler’s vertically integrated infrastructure stack. It creates a set of serious interoperability problems when it encounters the multi-vendor reality of the broader data centre market. A colocation operator hosting compute from multiple hyperscalers or a large enterprise running a mix of NVIDIA GPUs, Google TPU-class accelerators, and Microsoft Maia-based hardware faces a cooling estate in which each component may have been designed around a different fluid temperature, flow rate, pressure drop, connector standard, and material compatibility profile. The OCP PG25 standard defines a common envelope within which compliant hardware should operate without fluid interaction problems. A co-designed proprietary system optimised for a specific silicon geometry may operate outside that envelope at higher temperatures, different flow velocities, or with surface chemistry that interacts differently with the glycol inhibitor package without formally violating OCP specifications while practically making mixed-deployment fluid management significantly more complex.
The connector and manifold layer compounds the problem. Foxconn Interconnect Technology’s demonstration of UQD quick-disconnect connectors at OCP 2025 was explicitly framed as an attempt to standardise the wet-zone hardware interfaces that must connect cooling infrastructure across different vendors’ racks. These standards matter because a data centre hall with six different connector standards across six different rack types requires six different maintenance workflows, six different spare parts inventories, and six different leak detection calibration profiles. Fluid contamination the introduction of incompatible chemistry through cross-contamination of cooling loops is a failure mode with consequences that extend well beyond the immediate rack. Corrosion inhibitor packages tuned for propylene glycol at 25% concentration may behave differently when a marginal volume of a competing fluid chemistry enters the loop through a shared manifold or an inadvertent fluid transfer during maintenance. The OCP’s Cooling Environments Project is working to standardise wet-zone components across vendors precisely because the alternative a patchwork of proprietary connectors, fluid specifications, and material compatibility regimes across a single data hall creates operational complexity that increases downtime risk, maintenance cost, and the probability of failures that trace back not to hardware defects but to fluid chemistry interactions that no single vendor’s documentation anticipated.
nVent, as a participant in Google’s Project Deschutes 5.0, has articulated what responsible standardisation within a co-design context looks like in practice: providing cooling solutions that meet the highest standards for high-density AI compute while maintaining compatibility with OCP reference designs through a modular, scalable technology cooling system that eliminates complexity without requiring operators to choose between standardisation and performance. This approach co-designed for specific silicon performance while maintaining OCP interface compatibility represents the industry’s attempt to hold the tension between proprietary optimisation and open interoperability without fully resolving it in either direction.
The 3M Withdrawal and Its Industry-Wide Consequences
When the Supply Chain’s Foundation Exits
The coolant chemistry war acquired a new and complicating dimension in December 2022, when 3M announced that it would completely exit all PFAS and fluorochemical manufacturing by the end of 2025. The decision was driven by an avalanche of environmental litigation, billions of dollars in potential liabilities over drinking water contamination attributed to PFAS compounds, and tightening regulatory pressure from both the US Environmental Protection Agency and European regulatory bodies. For the data centre cooling industry, the consequences were immediate and structural. Three fluids that had made two-phase immersion cooling commercially viable Novec 7100, Novec 649, and Fluorinert FC-72 — were gone. One corporate decision wiped out the entire established supply chain for the fluid chemistry that two-phase immersion cooling systems had been designed around. 3M ended its Novec supply on December 31, 2025, eliminating between 30% and 40% of global fluorinated cooling fluid capacity in a single event.
Two-phase immersion cooling is the most thermally capable of the available liquid cooling approaches it achieves power usage effectiveness below 1.1, eliminates chiller requirements entirely, and handles thermal design powers exceeding 2,500 watts per chip in advanced configurations. Its commercial adoption trajectory had been accelerating precisely as 3M withdrew the fluid chemistry it depended on. The European Chemicals Agency’s evaluation of a PFAS restriction proposal covering more than 10,000 substances submitted by five EU member states under REACH added regulatory uncertainty beyond the 3M withdrawal itself, with updated proposals published in August 2025 expanding exemptions for certain heat transfer applications but with final ECHA opinions not expected until end of 2026 and European Commission restriction legislation anticipated in early 2027. The US AIM Act creates a parallel regulatory pressure trajectory pushing suppliers toward PFAS-free formulations on a timeline that, while longer than the European one, is moving in the same direction.
The 3M exit created two parallel dynamics in the cooling fluid market. The first was an immediate search for replacement supply for fluorinated fluids, met by Syensqo’s Galden PFPE product line positioned as a direct functional replacement for Novec and Fluorinert across semiconductor, EV, and data centre applications along with other fluorinated alternatives from suppliers that had not been subject to 3M’s litigation exposure. The second was a structural acceleration of the shift toward alternative fluid chemistries synthetic hydrocarbons, hydrocarbon oils, bio-ester formulations that do not carry PFAS classification and can be produced and qualified under stable regulatory conditions. Synthetic hydrocarbons led the immersion cooling fluids market with over 37% market share in 2025 after Intel certified Shell and ExxonMobil formulations for use with its hardware. The immersion cooling fluids market is forecast to grow from roughly $190 million in 2025 to over $840 million by 2032 a growth trajectory driven by both the underlying AI compute expansion and the fluid chemistry substitution dynamic that the 3M exit accelerated.
Ecolab’s $4.75 Billion Move: The Chemistry Company That Bought Its Way Into Hardware
What the CoolIT Acquisition Actually Signals
On March 20, 2026, Ecolab announced an agreement to acquire CoolIT Systems from KKR for approximately $4.75 billion in cash the largest acquisition in Ecolab’s history, priced at 29 times next-twelve-month EBITDA, and a transaction that SEC filings confirm is expected to close in the third quarter of 2026. CoolIT designs and manufactures coolant distribution units, cold plates, liquid-to-liquid heat exchangers, and direct-to-chip cooling technologies used by major hyperscale and colocation operators. The company generates approximately $550 million in annual sales, operates with the high-margin profile of a specialist hardware supplier to an infrastructure market with few credible alternatives, and holds deep pre-existing relationships with the world’s major hyperscalers relationships built through years of co-engineering cooling systems for specific silicon thermal requirements. For Ecolab, a company that had previously defined its data centre presence through water treatment chemistry, corrosion inhibitors, and digital monitoring services, the CoolIT acquisition is a structural transformation: it converts Ecolab from a chemistry supplier to the cooling system into the manufacturer of the cooling system itself.
Ecolab chairman and CEO Christophe Beck described the strategic intent with unusual precision in the acquisition announcement. By combining CoolIT’s thermal engineering with Ecolab’s expertise in water chemistry, fluid management, digital monitoring, and global service, the company is creating what Beck called “an end-to-end fluid management and cooling platform for AI data centers.” The Cooling-as-a-Service model that Ecolab has been developing — in which the company manages the full thermal lifecycle of a data centre cooling system rather than simply supplying chemicals for periodic treatment now has a hardware anchor that makes the service model operationally meaningful. A Cooling-as-a-Service offering that includes CDU design, cold plate engineering, fluid chemistry, corrosion monitoring, microbial control, and real-time performance analytics is a fundamentally different commercial proposition than a fluid supply contract or a hardware sale. It creates a service relationship in which the customer’s operational continuity depends on continuous engagement with a single provider across the full thermal management stack.
The Futurum Group’s analysis of the acquisition identified the critical commercial question this model creates: whether a company rooted in industrial services can match the engineering velocity and customer intimacy that data centre cooling demands in a market defined by rapid product cycles and hyperscaler expectations that evolve faster than most industrial service organisations are designed to accommodate. CoolIT’s alignment with NVIDIA’s liquid cooling roadmap its CDUs and cold plates are functional requirements for the latest generation of NVIDIA accelerators including the B200 and the next-generation Vera Rubin architecture — gives the combined Ecolab-CoolIT entity a position in the AI infrastructure supply chain that is not dependent on any single customer relationship but is structurally embedded in the hardware requirements of the most widely deployed AI chip architecture in the world. The acquisition doubles Ecolab’s Global High-Tech market opportunity from $5 billion to $10 billion, with that market growing at strong double-digits annually according to Ecolab’s own investor communications.
The Supply Chain Power This Creates
The Ecolab-CoolIT combination lands in the coolant chemistry war at the precise moment when both dimensions of the market fluid chemistry and cooling hardware are simultaneously in flux. On the fluid side, the 3M exit has disrupted the established supply chain for immersion cooling fluids and is pushing the entire market toward alternative chemistry. The companies best positioned to capture that transition are those that can offer qualified, tested, production-ready fluid formulations with stable regulatory profiles exactly the kind of industrial chemistry competency that Ecolab has spent a century developing. On the hardware side, the shift from commodity air cooling to application-specific liquid cooling architectures is creating premium margin opportunities for the suppliers that can co-engineer thermal systems with hyperscaler silicon teams and deliver them at the reliability and scale that mission-critical AI infrastructure requires exactly what CoolIT brings to the combination. Together, the merged entity holds a position across both dimensions of the market simultaneously, creating a commercial moat that neither a pure chemistry company nor a pure hardware manufacturer could construct independently.
The broader competitive map of the cooling market is being drawn rapidly and by well-capitalised players. Carrier, which entered the data centre cooling space through its acquisition of Viessmann’s climate business, is a late entrant but a large one. Vertiv, a long-established data centre infrastructure provider, is scaling its liquid cooling portfolio aggressively. Johnson Controls and Schneider Electric each hold positions in the facility cooling and power management layers that adjoin the liquid cooling stack. The oil majors Shell, ExxonMobil, TotalEnergies have entered the cooling fluid market with qualified single-phase and immersion fluids, drawn by margins and growth rates that their traditional lubricants businesses cannot match. FUCHS, Castrol, and Lubrizol are competing for fluid qualification spots across hyperscaler and colocation operator approved supplier lists. The chemical companies that dominated industrial lubricants are now competing for a market defined not by viscosity grades and API classifications but by dielectric properties, glycol purity specifications, microbial stability, and compatibility matrices with the specific cold plate materials of the world’s largest GPU operators.
TSMC’s Direct-to-Silicon Roadmap and the Next Frontier of Thermal Integration
When the Coolant Touches the Chip Directly
The coolant chemistry war’s next front is one that neither the OCP’s PG25 standard nor Ecolab’s Cooling-as-a-Service model currently addresses, because it sits at a layer of the chip architecture that has historically been outside the data centre operator’s domain: the chip package itself. TSMC is developing Direct-to-Silicon Liquid Cooling, a technology that etches cooling microchannels directly into the silicon die as part of the semiconductor manufacturing process, allowing coolant to flow in contact with the chip surface rather than through a cold plate mounted above it. The thermal resistance between the heat source and the coolant is reduced dramatically when the fluid contacts the silicon directly rather than conducting through successive layers of thermal interface material, copper cold plate, and fluid boundary layer. TSMC plans to deploy Direct-to-Silicon cooling commercially around 2027, targeting NVIDIA’s Feynman architecture the successor to the Blackwell generation and integrating the technology into multi-chiplet, multi-reticle-sized AI accelerators packaged using TSMC’s CoWoS advanced packaging technology.
Direct-to-silicon cooling introduces a new set of fluid compatibility requirements that push the boundaries of what the PG25 standard addresses. When coolant contacts silicon directly rather than flowing through copper or nickel-plated cold plate channels, the material compatibility requirements change fundamentally the fluid must be chemically inert to silicon, must not deposit particulates at the microchannel scale, must carry corrosion inhibitors appropriate for direct metallic silicon contact, and must maintain these properties across the operational lifetime of a chip whose cooling channels cannot be cleaned, flushed, or replaced in the field without replacing the chip itself. TechInsights analyst Manish Rawat flagged precisely this concern: long-term exposure to coolant, even dielectric, can induce chemical and mechanical stress on silicon structures, necessitating extensive qualification over five to ten year operational timescales that no existing coolant standard has addressed. Microsoft’s collaboration with Corintis on microchannel silicon cooling using organic channel geometries etched at micron scale faces the same qualification challenge, compounded by the additional process complexity and potential yield losses that sub-surface microchannel etching introduces.
The Outcome of the War Is Not Yet Written
Standardisation, Fragmentation, or Consolidation?
The coolant chemistry war does not have a predetermined winner, and the forces pushing toward standardisation and fragmentation are roughly balanced in ways that suggest the market will settle into a segmented structure rather than converging on a single outcome. The OCP’s PG25 standard is gaining genuine commercial traction the ecosystem of OCP Inspired and OCP Accepted fluid products is growing, the ASHRAE partnership adds institutional weight, and the operational logic of multi-vendor interoperability is compelling for the majority of the market that is not building proprietary silicon. For colocation operators, enterprise data centres, and the smaller cloud providers that cannot fund custom silicon and cooling co-design programmes, the PG25 standard represents the difference between a manageable cooling estate and an unmanageable collection of incompatible proprietary systems.
For the hyperscalers building custom silicon and co-designed thermal systems, standardisation is a floor rather than a ceiling a baseline from which proprietary innovation diverges when performance requirements demand it. The Microsoft-Fairwater system, Google’s Ironwood liquid cooling architecture, and Amazon’s custom CDU all build on propylene glycol chemistry while varying the hardware integration, fluid parameters, and service architecture in ways that are optimised for their specific silicon and operational contexts. These are not rejections of standardisation; they are extensions beyond it, operating in the premium performance layer where competitive advantage depends on thermal integration that commodity specifications cannot deliver. The Ecolab-CoolIT combination occupies a strategic position that spans both segments chemistry expertise covering the commodity fluid market and hardware expertise covering the premium co-design market with a service model that generates recurring revenue regardless of which fluid standard ultimately dominates in any given deployment tier.
The question that the coolant chemistry war will answer over the next three to five years is not which fluid wins but who controls the service relationship that determines what fluid runs where, at what chemistry, monitored by which system, and serviced by which organisation. In an AI infrastructure market where a cooling system failure can take a GPU cluster offline within minutes, where the cost of a single unplanned downtime event at a hyperscale facility can exceed the annual revenue of a mid-sized technology company, and where the thermal requirements of successive GPU generations are escalating faster than any cooling specification can comfortably track, the organisation that holds the service relationship across fluid chemistry, hardware design, and digital monitoring has constructed a position whose strategic value exceeds what any component of that position would be worth in isolation. The war inside the pipe is quieter than the war over power and compute. It is not less consequential.
