As AI Chips Near 4,000 Watts, Liquid Cooling Becomes the Backbone of Data Center Design

The AI buildout is often described as a race for chips and power. Increasingly, it is a race to manage heat. Every watt a chip consumes becomes heat that has to be removed, and as AI hardware grows more powerful, how data centers handle that heat is becoming one of the defining infrastructure questions of the decade.

That message came through clearly at last month's Open Compute Project Global Summit in San Jose. A new OCP thermal paper argued that the industry tends to adopt new cooling methods only after old ones fail, and pays for waiting. Its conclusion was blunt.

Energy demand is accelerating

The numbers explain the urgency. The International Energy Agency's Key Questions on Energy and AI report found that data center electricity demand rose 17% in 2025, far ahead of the 3% growth in global power demand. The agency expects data center electricity use to double by 2030, with AI-specific demand tripling over the same period. AI-focused data centers grew faster still, with their electricity use climbing 50% last year.

Racks are heading toward the megawatt range

That energy is being concentrated into smaller spaces. Dell'Oro Group's summit recap noted that NVIDIA's MGX 1.2 specification supports racks up to 250 kW, while the OCP community sees a clear path to 1 MW racks, even if they arrive later than earlier roadmaps suggested. NVIDIA's Ian Buck told the summit that the first 800-volt DC AI factory scales from today's 100-kilowatt racks toward 1 megawatt.

From optional upgrade to structural requirement

Market research points in the same direction. Dell'Oro Group forecasts the data center liquid cooling market will reach roughly $7 billion in manufacturer revenue by 2029, as rising accelerator power and rack densities push air cooling beyond practical limits. The firm also projects that leading-edge GPUs will exceed 4,000 W of thermal design power by 2029.

That 4,000-watt mark is where cooling strategies begin to diverge. Single-phase direct liquid cooling handles most liquid-cooled capacity today, but Dell'Oro expects two-phase direct liquid cooling to accelerate as chip-level power and heat flux exceed what single-phase systems can practically manage. For operators, that makes the next few years a planning window: infrastructure decisions made now will determine how easily facilities absorb the next generation of chips.

The numbers shaping the shift

• 17%: growth in global data center electricity demand in 2025, against 3% for all electricity (IEA)
• 50%: growth in electricity use at AI-focused data centers in 2025 (IEA)
• 250 kW: rack capacity supported by NVIDIA's MGX 1.2 specification, with a path to 1 MW (Dell'Oro, OCP)
• 4,000 W: projected thermal design power of leading-edge GPUs by 2029 (Dell'Oro)
• $7 billion: projected size of the data center liquid cooling market by 2029 (Dell'Oro)

Two-phase cooling prepares for the next threshold

Two-phase cooling works by letting a refrigerant boil at the chip surface, carrying heat away as vapor. One company building for that threshold is Austin-based Accelsius, whose NeuCool platform uses two-phase, direct-to-chip cooling. The company reports NeuCool can cool more than 4,500W per socket, enough to handle future GPUs and CPUs without forcing operators into costly facility redesigns. That places its capacity above the level Dell'Oro expects leading GPUs to reach by the end of the decade.

Accelsius is also scaling with rack density. In October, the company upgraded its Austin lab to multi-megawatt power to validate its platform at the rack densities and facility water temperatures next-generation AI infrastructure is expected to require. It also announced a 375kW-capable unit as the next step beyond its 250kW MR250 multi-rack system.

The approach addresses efficiency and risk, two concerns that grow with density. Independent benchmarking commissioned by Accelsius found its system could match single-phase performance at 45°C facility water while using 54°C inlet water, and the company says each degree of warmer water equals roughly 4% in annual energy savings. Its non-conductive dielectric refrigerant causes no damage to electronics if a leak occurs, which the company positions as a way to reduce insurance exposure for high-value AI hardware.

Adoption is extending to the workforce as well. EdgeConneX is deploying the Accelsius IR150 at its global training center, where operators, technicians and partners will get hands-on experience deploying and servicing two-phase cooling for AI workloads. Accelsius CEO John Hewitt told Facilities Dive the deployment signals that operators are poised to transition to the technology.

Planning ahead of the curve

For data center leaders, the lesson is about timing as much as technology. Energy demand is surging, rack densities are climbing toward a megawatt, and chip power is approaching 4,000 watts. The operators that validate next-generation cooling now, before their hardware demands it, will be best positioned to scale with AI rather than retrofit after the fact.