Air cooling has hit a thermal wall. As GPU thermal design power (TDP) climbs past 700W toward 1,000W and beyond, and rack densities push from roughly 20 kW toward 50 to 150 kW and higher, moving that much heat with fans and cold air stops being practical. Air simply cannot carry heat away fast enough to keep a densely packed AI rack inside safe thermal limits.

‍

Direct-to-chip (D2C) liquid cooling has become the practical default for high-density artificial intelligence (AI) and high-performance computing (HPC). It brings coolant to the exact spot where heat is generated, so operators can run current-generation accelerators at full clocks without hot spots or oversized air handling. For teams building or buying high-density AI and HPC colocation, the question is no longer whether to adopt direct-to-chip cooling. It is how to deploy it well. This article answers the questions we field most often.

‍

‍

What Is Direct-to-Chip Cooling and How Does It Work?

‍

Direct-to-chip (D2C) cooling is a liquid cooling technique where a cold plate is mounted directly on the chip or processor. Coolant circulates through the plate, absorbing heat at the source and carrying it away far more efficiently than traditional air cooling or rear-door heat exchangers. Unlike immersion or in-row cooling, D2C focuses cooling capacity precisely where it is needed most: on the CPU, GPU, or other heat-generating components.

‍

The full loop is a chain, and each link matters:

‍

1. Cold plate: a metal block bolted to the processor die, with internal channels the coolant flows through.
2. Coolant: the fluid that absorbs heat at the plate and transports it out of the server.
3. Circulation loop: sealed tubing and quick-disconnects that move fluid between the plate and the distribution unit.
4. Coolant Distribution Unit (CDU): the pump-and-heat-exchanger stage that isolates the server loop from facility water and controls flow, pressure, and temperature.  
5. Facility heat rejection: dry coolers, cooling towers, or chillers that dump the collected heat outside the building.

‍

Example: trace the heat path for a single 1,000W GPU. The die transfers heat into the cold plate; coolant in the plate warms and flows through the circulation loop to the CDU; the CDU passes that heat across a heat exchanger into the facility water loop while keeping the two fluids separate; the facility loop carries the heat to outdoor rejection equipment. The cooled coolant returns to the plate and the cycle repeats. Nothing about that path depends on pushing large volumes of cold air, which is the whole point.

‍

‍

Why Are AI and HPC Workloads Driving the Need for Better Cooling?

‍

Al and HPC systems generate far more heat than traditional enterprise workloads. Three drivers stack up:

‍

1. High-density GPU clusters packed into a single rack.
2. Tightly packed CPU arrays feeding those GPUs.
3. Massive parallel computation that keeps every device near full utilization for hours or days.

‍

The silicon makes this concrete. NVIDIA H100 parts draw roughly 700W, and AMD Instinct M1300 accelerators sit in similar territory. Current-generation NVIDIA B200 and B300 (Blackwell and Blackwell Ultra) and GB200-class parts reach roughly 1,000W to 1,400W TDP per GPU. Pack enough of those into a cabinet and rack density moves from about 20 kW into the 50 to 150 kW range and higher. Legacy air-cooled systems cannot hold safe thermal thresholds at that density without hot spots or brute-force airflow. Direct-to-chip cooling lets operators deploy them at scale without thermal throttling or overbuilding air handling.

‍

‍

What coolant does direct-to-chip cooling use, and does it use water?

‍

It depends on the design. Single-phase D2C systems typically circulate treated water mixed with glycol (a propylene or ethylene glycol blend) for corrosion control and freeze protection. So yes, many D2C systems do use water, but it is conditioned, closed-loop water, not tap water splashing near live electronics.

‍

Other designs use a dielectric fluid, a non-conductive coolant that will not short components even on direct contact. That is the core tradeoff: water and glycol are thermally excellent but conductive, while dielectric fluids are electrically safe but harder to source and maintain.

‍

Either way, the loop is sealed. Reputable deployments run closed circulation with leak detection at the rack and row, plus quick-disconnects designed to seal on separation. The "liquid near servers" fear is reasonable, and the operational answer is containment and telemetry, not hope.

‍

‍

Single-phase vs. two-phase direct-to-chip cooling

‍

Direct-to-chip comes in two thermodynamic modes, and the difference is whether the coolant changes state.

‍

1. Single-phase: the coolant stays liquid the entire trip, absorbing heat as a temperature rise. Mature, well understood, and serviceable with widely available parts.
2. Two-phase: the coolant boils at the cold plate and condenses downstream, using the latent heat of vaporization to move more heat per unit of flow. Higher density headroom, but more complex.
3. The tradeoff: single-phase wins on ecosystem maturity and field serviceability; two-phase wins on raw heat-transfer capacity for the densest future silicon, at the cost of tighter fluid handling and a smaller supplier base.

‍

Our view: single-phase D2C is the right default for most enterprise Al today. It handles current 1,000W-plus GPUs, integrates with standard rack and server form factors, and does not ask operators to bet production workloads on a thinner supply chain. Two-phase will earn its place as densities keep climbing, but "more headroom" is not free, and precision beats novelty when uptime is the job.

‍

‍

Is direct-to-chip cooling more sustainable than air cooling?

‍

Yes, measurably. D2C liquid cooling can reduce cooling energy consumption by up to roughly 40% compared with air. Because liquids move heat far more efficiently than air, fans and Computer Room Air Conditioning (CRAC) units run at lower duty or drop out entirely. That translates to real operational gains:

‍

1. Lower Power Usage Effectiveness (PUE), since less energy goes to moving air.
2. A smaller cooling infrastructure footprint per unit of compute.
3. Less water use than evaporative cooling in some closed-loop designs.

‍

D2C also enables waste-heat reuse, a growing practice in efficient data center design. Because the captured heat leaves the rack in a concentrated liquid stream rather than diffuse warm air, facilities can repurpose it for nearby buildings or district heating instead of simply rejecting it.

‍

‍

Direct-to-chip vs. immersion cooling: which fits your workload?

‍

Both are liquid cooling, but they solve the problem differently. D2C targets the hottest components with cold plates while the rest of the server stays air-adjacent. Immersion submerges whole boards in dielectric fluid to capture heat across the entire system. The choice comes down to honest tradeoffs:

‍

1. Serviceability: D2C works with standard rack and server form factors, so hardware swaps stay familiar. Immersion requires pulling dripping boards from a tank.
2. Ecosystem maturity: D2C has broader vendor and integrator support today; immersion has a narrower field.
3. Density ceiling: immersion captures more total heat per system, which matters at the extreme high end; D2C covers the vast majority of current Al racks.
4. Retrofit fit: D2C is generally easier to introduce into existing high-density rooms than a tank-based redesign.

‍

Over a 3 to 5 year total cost of ownership (TCO) horizon, D2C tends to earn its keep through operational results rather than any single line item:

‍

1. Fewer hardware failures from thermal stress.
2. Longer equipment lifespan under stable temperatures.
3. Smaller space and power requirements per workload.

‍

‍

‍

‍

How WhiteFiber deploys direct-to-chip cooling

‍

WhiteFiber is a GPU cloud provider, and we engineer cooling as one layer of a matched full-stack system, not a bolt-on. Power, high-density data center architecture, direct liquid cooling, networking fabric, and storage are designed together, so GPUs are never starved for data or throttled by heat. That integration is the difference between a rack that benchmarks well and a cluster that sustains utilization month after month.

‍

In practice, that means liquid-ready racks provisioned for 150 kW and above, sized for current 1,000W-plus accelerators rather than retrofitted after the fact. We pair that with operational discipline: closed-loop containment, leak detection, and transparency into cooling and power telemetry so operators can see thermal and utilization behavior instead of guessing at it.

‍

Our conclusion is straightforward: for high-density Al, direct-to-chip cooling is a present-day necessity, not a future upgrade. The workloads are here, the silicon runs past 1,000W, and the racks are already past the point air can handle. Teams that treat cooling as core infrastructure, designed in from the start, are the ones that keep their GPUs busy.

‍

‍

‍

‍

FAQ

Is liquid cooling the same as direct-to-chip cooling?

Not quite. Direct-to-chip is one type of liquid cooling that uses cold plates on the chips; liquid cooling also includes immersion and rear-door heat exchangers.

‍

‍

Does direct-to-chip cooling use water?

Often, yes: many single-phase systems circulate treated water and glycol, while two-phase systems use a dielectric fluid. In both cases the loop is closed, with leak detection to contain risk.

‍

‍

What rack densities require direct-to-chip cooling?

Air cooling starts to struggle past roughly 30 to 50 kW per rack. Direct-to-chip is typically needed for the 50 to 150 kW and higher densities of modern Al clusters.

‍

‍

How long do direct-to-chip cooling systems last?

With disciplined fluid management and routine maintenance, the cold plates and loop hardware run for years and can outlast several GPU refresh cycles.