Every new GPU generation pushes TDP higher. The industry's messy transition from air to liquid is no longer optional — it's a design constraint for every new build.
The immersion cooling market will hit an estimated $931 million in revenue this year and is projected to reach $4.9 billion by 2033, growing at a 27.1% CAGR. Those numbers, from a new market analysis published March 10, tell a story that anyone who's designed a GPU cluster in the past two years already knows: air cooling is done. Not dying. Done.
The question isn't whether liquid cooling wins. It's which kind, at what cost, and how fast the industry can actually make the transition without breaking everything that's already built.
Let's talk about why this is happening with a specificity that market reports tend to gloss over.
NVIDIA's H100 runs at roughly 700W TDP. The Blackwell B200 pushes that to nearly 1,000W per GPU. The Vera Rubin platform, expected to ship later this year, will push further still. We're trending toward kilowatt-per-chip territory as the baseline for data center accelerators.
Conventional air-cooled systems start struggling above 30-50 kW per rack. A modern GPU rack with eight accelerators easily exceeds that. An NVL72 rack with 72 GPUs in a dense configuration? Air cooling isn't inadequate. It's physically incapable of removing heat at the rates these systems generate.
Schneider Electric's Christopher Leonard made the case plainly in a recent analysis: single-phase direct liquid cooling (DLC) maintained chip-to-coolant temperature differences of roughly 17-20°C at 500W processor loads, while comparable air-cooled systems exceeded 60°C under similar conditions. That's not a marginal difference. That's the difference between sustained GPU performance and thermal throttling that tanks your inference throughput.
And thermal throttling doesn't just reduce performance. It reduces revenue. When GPUs throttle, token generation rates drop, model training stalls, and the per-megawatt economics of your facility crater. Cooling has become, as Schneider puts it, "a revenue protection layer for AI operations."
The cooling industry hasn't consolidated around a single approach, and that's creating real headaches for infrastructure planners. Here's the landscape:
Cold plates mounted directly on CPUs and GPUs, circulating liquid through closed loops. This is the approach Schneider Electric is backing hardest, and it has real advantages: efficient heat capture at the source, operational familiarity for existing data center teams, the ability to access and swap individual components without draining a tank, and compatibility with brownfield retrofits.
The ecosystem is maturing rapidly. Coolant Distribution Units (CDUs), in-rack manifolds, cold plates from companies like Motivair, and rear-door heat exchangers for hybrid configurations are all commercially available. This is the path of least operational disruption.
Servers submerged in dielectric fluid that boils on contact with hot components, condensing in a closed loop. Superior heat removal capacity because the phase change absorbs enormous thermal energy. The market analysis identifies two-phase as the fastest-growing technology segment through 2033, driven by rack densities that even single-phase DLC struggles with at the extreme end.
The problems are practical: specialized tanks, custom server designs, fluid management complexity, and the ongoing PFAS regulatory situation in Europe that threatens the supply of fluorinated dielectric fluids. Bio-based alternatives are emerging but aren't yet proven at scale.
Servers submerged in fluid that stays liquid. Simpler than two-phase, but still requires purpose-built infrastructure. Currently the dominant revenue segment in immersion cooling due to lower cost and easier integration. But it faces the same deployment challenges: you need tanks, you need fluid management, and you need to design servers for submersion from the start.
Still the installed base. Still what most data centers run. And still what many operators are trying to stretch, using computational fluid dynamics and hot/cold aisle optimization to squeeze another generation of life out of air-based designs. It's a losing battle, but the capital cost of conversion means air cooling won't disappear from existing facilities overnight.
This is where the market reports get optimistic and reality gets complicated.
Building a new facility with liquid cooling from the ground up? Increasingly straightforward. The design patterns exist, the components are available, and every major infrastructure vendor has a product line. Greenfield AI factories are being designed liquid-first.
Retrofitting an existing air-cooled data center? Different story entirely. You're looking at new rack designs to accommodate cold plates, manifolds, or immersion tanks. Plumbing infrastructure: CDUs, piping, leak detection systems. Fluid procurement and management, because dielectric fluids aren't cheap and supply chains are still maturing. Modified server hardware, since not all servers are designed for liquid cooling. And operational retraining, because your facilities team knows air cooling and liquid cooling requires different skills, different monitoring, different failure modes.
The market analysis notes that "high initial infrastructure costs and integration challenges" remain the primary barrier to adoption. Total cost of ownership improves over time (immersion cooling can reduce cooling energy consumption by up to 40% compared with traditional systems) but the upfront capital expenditure is substantial.
This is where the cooling discussion connects directly to the compute discussion. NVIDIA's Rubin platform isn't just faster. It's hotter. Every generation pushes TDP higher. Every TDP increase makes air cooling less viable and liquid cooling more necessary.
The industry is locked in a feedback loop: AI workloads demand more GPU power. More GPU power generates more heat. More heat requires liquid cooling. Liquid cooling requires infrastructure investment. Infrastructure investment takes time and capital. Meanwhile, the next GPU generation is already announced.
For operators building new AI factories (the hyperscalers, the sovereign AI builds, the gigawatt-scale projects like the NVIDIA-Thinking Machines Lab partnership) liquid cooling is a given. It's in the design from day one. The Rubin NVL72 rack is essentially designed to be liquid-cooled.
For everyone else, enterprises upgrading existing facilities, mid-tier cloud providers, research institutions, the transition is a rolling capital planning challenge that won't be resolved in a single budget cycle.
North America is emerging as the primary investment zone, driven by gigawatt-class AI data center developments. Server OEMs are partnering directly with fluid providers, specialty chemical manufacturers are racing to develop PFAS-free bio-based dielectric fluids, and modular cooling systems are reshaping supply chain economics.
The regulatory environment matters too. European mandates restricting fluorinated compounds are pushing the industry toward sustainable coolant alternatives. That's a positive long-term development, but in the near term it adds uncertainty to procurement decisions for anyone considering two-phase immersion systems that rely on fluorinated fluids.
The bottom line: cooling infrastructure is no longer a facilities concern that gets handled after the compute architecture is designed. It's a co-design constraint, as fundamental to system architecture as the GPU selection itself. Any organization planning a GPU deployment in 2026 or 2027 that hasn't made a cooling technology decision is already behind.
