Technology
Immersion cooling for AI GPUs: how it works and why it matters
How single-phase immersion cooling works, why dense AI GPUs need it, and what it changes for energy efficiency, density and operations.
Immersion cooling technology
Single-phase immersion cooling, built for AI GPUs.
Inside the BAYWAW AI Container, 624 NVIDIA B300 GPUs run fully submerged in a non-conductive liquid. The result is an immersion cooling data center with a PUE of 1.05, sealed from dust and humidity, that fits 1.2 MW into a standard 40ft container.
- PUE
- 1.05
- Tanks (3 + 3)
- 6
- GPUs per tank
- 104
Redundant coolant loops and CDUs
01The principle
What is single-phase immersion cooling?
Single-phase immersion cooling replaces air with liquid. Servers are lowered into a bath of dielectric fluid, a liquid that does not conduct electricity, so every component, from GPU to memory to power stage, is in direct contact with the coolant.
“Single-phase” means the fluid stays liquid from start to finish. It never boils: heat is carried away by circulation, through a heat exchanger, and the cooled liquid flows back into the bath.
Liquid absorbs and carries heat far more effectively than air. That removes server fans, hot and cold aisles and much of the room-level air handling that weighs on the energy overhead of a conventional data center.
- Dielectric
- Non-conductive and safe for electronics, so hardware runs fully submerged.
- Single-phase
- No boiling, no evaporation. The coolant stays liquid throughout the loop.
- Circulation
- Pumps move the warm liquid to a heat exchanger and return it cooled.
Anatomy of an immersion tank
Coolant distribution manifold
Feeds cooled liquid evenly along the base of the tank.
Server blades
Seated on liquid-cooled plates and fully submerged, thirteen per tank.
Dielectric liquid bath
Absorbs heat directly from every component it touches.
Return manifold
Collects the warmed liquid and returns it to the coolant distribution unit (CDU).
02How it works
Follow the heat, from chip to sky.
Two loops that never mix. A dielectric loop carries heat from the servers to the coolant distribution units; a separate facility water loop carries it to roof-mounted coolers, which release it into the ambient air.
Heat enters the liquid
GPUs, CPUs, memory and power components release their heat directly into the dielectric liquid around them. There is no air gap between component and coolant.
Warm liquid is collected
The return manifold draws the warmed liquid from the top of the bath and sends it to the coolant distribution unit (CDU).
The CDU exchanges the heat
Inside the CDU, a heat exchanger transfers the heat from the dielectric loop to the facility water loop. The two fluids never touch.
Cooled liquid returns
The cooled dielectric liquid flows back through the coolant distribution manifold, which spreads it evenly beneath the server blades.
Water carries heat to the roof
The facility water loop carries the heat out of the IT space to the heat-rejection units mounted on the container roof.
Heat is released to the air
The roof-mounted coolers release the heat into the ambient air and send cooled water back to the CDUs, closing the loop.
Redundant by design
Coolant loops and CDUs are redundant, and the whole thermal chain is monitored around the clock as part of the managed service.
03Why AI needs it
AI GPUs have outgrown air.
Blackwell-class GPUs such as the NVIDIA B300 concentrate far more heat per chip and per rack than previous generations. Liquid immersion cooling for GPUs removes that heat at the source.
Air cooling scales poorly with that curve. Moving enough air through a dense GPU server takes larger heatsinks, faster fans and colder supply air. At rack level, airflow and the space needed to deliver it become the limit, not the compute.
Immersed in a dielectric bath, eight-GPU servers sit side by side in a tank and hand their heat directly to the liquid. That is how the BAYWAW AI Container holds 624 GPUs within a single 40ft footprint.
More heat per chip
Each accelerator generation packs more power into a similar package, raising the heat flux the cooling system has to remove.
More heat per rack
AI clusters keep GPUs physically close to shorten interconnects, concentrating heat into a small footprint.
Air reaches its limits
Air carries little heat per unit of volume. Past a certain density, no practical amount of airflow keeps up.
- NVIDIA B300 GPUs
- 624
- 8-GPU servers
- 78
- GPUs per tank
- 104
- Facility power
- 1.2MW
04Benefits
Less overhead, more compute, steadier hardware.
Immersion is not only about efficiency. It changes how densely you can build, how steadily the hardware runs and where you can deploy.
Power usage effectiveness
Lower is better
BAYWAW AI Container1.05
Industry average≈ 1.5–1.6
A PUE of 1.05
Almost every watt the facility draws reaches the compute. Industry-average PUE, by comparison, has hovered around 1.5–1.6 for years.
624 GPUs in 40 feet
Because liquid removes heat at the source, six tanks hold 78 servers and 624 NVIDIA B300 GPUs inside a single ISO container.
No server fans
Without fans, servers have fewer moving parts that can fail, and the hardware runs in a quieter, more stable thermal environment.
Sealed from dust and humidity
Components sit beneath the liquid surface, isolated from airborne dust, sand and humidity that wear down air-cooled hardware.
Ready for harsh climates
Built to withstand transport and harsh environments, the container keeps its electronics in a stable liquid bath, from desert sites to industrial yards.
Redundant cooling chain
Coolant loops and CDUs are redundant, so the cooling of the compute does not hinge on a single unit.
05Air vs. immersion
Air cooling and immersion, side by side.
Air cooling remains a sound, familiar choice for general-purpose IT. The balance shifts as density rises. Here is a fair comparison for dense AI GPU clusters.
Heat transfer medium
- Air cooling
- Air, moved by server fans and room air handlers.
- Single-phase immersion
- Dielectric liquid in direct contact with every component.
Compute density
- Air cooling
- Limited by how much air can be pushed through each rack.
- Single-phase immersion
- Built for dense GPU servers: 104 GPUs per tank, 624 per container.
Server fans
- Air cooling
- Required in every server, often at high speed under AI loads.
- Single-phase immersion
- None. Pumps in the CDUs circulate the liquid.
Energy overhead
- Air cooling
- Fans, chillers and air handling add substantial overhead; industry-average PUE is around 1.5–1.6.
- Single-phase immersion
- Minimal: PUE 1.05 for the BAYWAW AI Container.
Sensitivity to ambient conditions
- Air cooling
- Dust, humidity and heat in the room reach the electronics; filtration and climate control are essential.
- Single-phase immersion
- Electronics sit in a sealed bath, isolated from dust and humidity.
Noise
- Air cooling
- Loud: high-speed fans in every server, plus room-level cooling.
- Single-phase immersion
- Quiet at the tanks; the main moving parts are pumps and the roof-mounted coolers.
Maintenance
- Air cooling
- Familiar, widely known procedures; filters and fans to service and replace.
- Single-phase immersion
- Blades are lifted out of the bath; procedures and fluid handling call for trained engineers, provided within our managed service.
| Criterion | Air cooling | Single-phase immersion |
|---|---|---|
| Heat transfer medium | Air, moved by server fans and room air handlers. | Dielectric liquid in direct contact with every component. |
| Compute density | Limited by how much air can be pushed through each rack. | Built for dense GPU servers: 104 GPUs per tank, 624 per container. |
| Server fans | Required in every server, often at high speed under AI loads. | None. Pumps in the CDUs circulate the liquid. |
| Energy overhead | Fans, chillers and air handling add substantial overhead; industry-average PUE is around 1.5–1.6. | Minimal: PUE 1.05 for the BAYWAW AI Container. |
| Sensitivity to ambient conditions | Dust, humidity and heat in the room reach the electronics; filtration and climate control are essential. | Electronics sit in a sealed bath, isolated from dust and humidity. |
| Noise | Loud: high-speed fans in every server, plus room-level cooling. | Quiet at the tanks; the main moving parts are pumps and the roof-mounted coolers. |
| Maintenance | Familiar, widely known procedures; filters and fans to service and replace. | Blades are lifted out of the bath; procedures and fluid handling call for trained engineers, provided within our managed service. |
Qualitative comparison for dense AI GPU deployments. Air-cooled performance varies widely with design and climate.
Discuss your site conditions with an engineer06Operations & maintenance
Designed to be serviced in place.
Immersion changes the maintenance routine, not its rigor. The 3 + 3 tank layout leaves a central service aisle, so every tank can be reached from inside the container.
Access from the aisle
Engineers work from the central service aisle, with every tank within reach.
Lift the blade
The server blade is lifted vertically out of the bath and left to drain above the tank before any work begins.
Service or replace
Faulty components are diagnosed and replaced, then the blade returns to its slot and is checked before it rejoins the cluster.
Fluid handled by specialists
The dielectric liquid is handled by trained engineers as part of the managed service, so your team does not need to touch it.
Covered by the managed service
- 24/7 monitoring and alerting
- Rapid diagnosis and component replacement
- On-site engineers for maintenance and incidents
- Warranty management and OEM coordination
07Go deeper
Keep exploring the technology.
Product
Inside the BAYWAW AI Container
Six immersion tanks, 624 NVIDIA B300 GPUs, the rail-optimized network fabric and the power architecture, component by component.
08FAQ
Immersion cooling, answered.
Straight answers to the questions infrastructure teams raise most often about immersion-cooled AI compute.
Is the liquid safe for electronics?
Yes. The servers are immersed in a dielectric liquid, which does not conduct electricity, so components operate fully submerged without short-circuiting. The fluid is selected for compatibility with server hardware and is handled by trained engineers.
What happens during maintenance?
Engineers work from the central service aisle. A server blade is lifted vertically out of the tank, left to drain, serviced or replaced, then returned to its slot and checked. Fluid handling is part of the BAYWAW managed service, so your team does not need to touch the liquid.
Does immersion cooling use water?
No water ever touches the servers: they are cooled by the dielectric liquid. Heat leaves the CDUs through a separate facility water loop to roof-mounted coolers, which release it into the ambient air. The heat-rejection setup, and any water requirement on your site, is confirmed during project engineering.
Does immersion affect hardware warranties?
Warranty is managed for you, end to end. Our warranty management service coordinates with the original equipment manufacturers and handles claims on your behalf. The warranty terms that apply to your configuration are reviewed with you during the proposal stage.
Is immersion cooling reliable?
Immersion removes several common failure sources of air-cooled servers: there are no server fans, and components are shielded from dust, humidity and temperature swings. Coolant loops and CDUs are redundant, and the system is monitored 24/7 with alerting and automated remediation where possible.
Can an immersion-cooled container operate in desert conditions?
Yes. The electronics sit in a sealed liquid bath, away from sand, dust and humidity, and the container is built to withstand transport and harsh environments. We review your site’s climate with you so that heat rejection is engineered for local conditions.
Technical briefing
Review the thermal design with our engineers.
Bring your site, climate and workload. We walk your team through the tank layout, coolant loops and heat rejection, and answer every question in detail.