Direct to chip liquid cooling is a cooling method where coolant flows through a cold plate mounted directly on top of the GPU or CPU die, absorbing heat much more efficiently than air passing over a heatsink, which makes it well suited to high power chips like the B200 or GB200 that can draw close to or above 1000 watts each. The heated coolant is pumped to a coolant distribution unit, or CDU, which transfers that heat to a facility water loop or external chiller, allowing the system to reject far more heat per rack than air cooling could manage at the same rack footprint. Unlike full immersion cooling, direct to chip systems keep the rest of the server, including memory, storage, and networking, air cooled, which makes it a more incremental and widely adopted approach for retrofitting existing data centers. Implementing it requires plumbing infrastructure, leak detection systems, and often facility water loops that many traditional data centers were not originally built with, representing a real infrastructure investment beyond the servers themselves. It has become close to mandatory for the densest Blackwell based systems like the GB200 NVL72, while remaining optional for many H100 and B200 deployments at lower density. Nanobase AI designs and installs direct to chip liquid cooling for clients moving to high density GPU racks.
How the heat actually leaves the chip
Direct-to-chip liquid cooling replaces the air-cooled heatsink and fan assembly sitting on top of a GPU or CPU with a cold plate, a metal block with internal channels through which coolant flows, mounted directly against the die. Heat transfers from the silicon into the coolant far more efficiently than it would into air, because liquid has a much higher heat capacity and thermal conductivity per unit volume. This efficiency is what makes direct-to-chip cooling practical for chips like the B200, which can draw close to or above 1,000 watts each, well beyond what a reasonably sized air heatsink can dissipate without impractical airflow.
The heated coolant does not evaporate or leave the system; it circulates in a closed loop from the cold plate to a coolant distribution unit, or CDU, which acts as a heat exchanger, transferring the absorbed heat to a facility chilled water loop or an external dry cooler before sending the now-cooled liquid back to the racks.
Direct-to-chip versus the alternatives
| Cooling method | What it cools | Typical use case | Infrastructure needed |
|---|---|---|---|
| Air cooling with containment | Entire server via heatsinks and fans | H100-class nodes, moderate density racks | Standard HVAC, hot-aisle/cold-aisle containment |
| Direct-to-chip liquid cooling | GPU and CPU dies via cold plates; rest of server stays air-cooled | B200-class nodes, GB200 NVL72, dense multi-kW racks | CDU, plumbing, leak detection, facility water loop |
| Full immersion cooling | Entire server submerged in dielectric fluid | Very high density or specialized deployments | Sealed tanks, dielectric fluid handling, different rack form factor |
Direct-to-chip is deliberately a partial approach: only the highest-power components get liquid cooling, while memory, storage, networking cards, and power supplies remain air-cooled inside the same chassis. That makes it a more incremental retrofit path than full immersion, which requires a fundamentally different rack and server design.
Why it has become close to mandatory for the densest Blackwell systems
At the power levels of a fully configured 8-GPU HGX B200 node, and especially at the roughly 120 kW draw of a GB200 NVL72 rack, no realistic amount of airflow can remove heat fast enough to keep junction temperatures in a safe operating range. NVIDIA's reference designs for these systems assume direct-to-chip cooling as the default rather than an optional upgrade. Lower-density H100 deployments and smaller B200 configurations can still run on well-provisioned air cooling with hot-aisle containment, so the requirement scales with power density rather than applying uniformly across every current GPU generation.
What a facility needs before adopting it
- A coolant distribution unit sized to the planned rack count and total heat load, not just the initial deployment.
- Plumbing runs from the CDU to each rack, engineered with proper isolation valves so a single leak or maintenance event does not require draining the whole loop.
- Leak detection sensors at rack and row level, since even a small coolant leak near electronics is a serious risk that air-cooled facilities never had to plan for.
- A facility water loop or external dry cooler connection with enough thermal rejection capacity for the added heat load, which may require chiller plant upgrades in older buildings.
- Trained operations staff or a service contract covering coolant quality monitoring, since contaminated or degraded coolant can damage cold plates over time.
Frequently asked questions
Does direct-to-chip cooling replace all fans in the server?
No. Direct-to-chip cooling handles the GPU and CPU dies specifically; memory, storage, networking, and power supplies in the same chassis typically remain air-cooled, so some fans usually remain for those components.
Is direct-to-chip cooling the same as immersion cooling?
No. Direct-to-chip uses cold plates on specific components while the rest of the server stays in normal air, whereas immersion cooling submerges the entire server in a dielectric fluid, requiring a different rack and server design altogether.
Can an existing air-cooled data center be retrofitted for direct-to-chip cooling?
Yes, this is a common retrofit path, but it requires real infrastructure investment: a coolant distribution unit, plumbing, and leak detection, plus enough facility chilled water capacity to reject the added heat. It is a genuine project, not a minor upgrade.
Do all Blackwell GPUs require liquid cooling?
Not all configurations. Lower-power B200 SKUs and reduced GPU-count servers can often still run on well-provisioned air cooling, while full 8-GPU HGX B200 nodes and rack-scale systems like GB200 NVL72 generally require direct-to-chip liquid cooling.
How Nanobase AI helps
Nanobase AI designs and installs direct-to-chip liquid cooling retrofits for clients moving to high-density GPU racks, covering CDU sizing, plumbing layout, and leak detection alongside the compute deployment itself. We assess whether a client's current facility can support liquid cooling as-is or needs upgrades first, then sequence the retrofit around production timelines. Learn more in our H100 vs H200 vs B200 guide or explore GPU infrastructure services.
Ready to discuss your project? Contact Nanobase AI or email hello@bumu.tech.