Liquid where it
matters most.
Cold-plate direct-to-chip cooling delivers the thermal headroom of liquid systems with the operational familiarity of traditional rack architectures. The pragmatic path to 80 kW racks.
The architecture
Liquid the chips need. Air for everything else.
Cold-plate DTC routes coolant directly to processor cold plates, capturing 70–80% of rack heat at source while leaving the rest of the rack — drives, NICs, ancillary boards — air-cooled. The result is liquid-grade thermal performance with the familiar operational model of an air-cooled hall: same rack form factor, same hot-aisle/cold-aisle layout, same servicing approach.
- Captures heat at the highest-energy components, where it actually matters.
- Compatible with most modern server designs without specialized chassis.
- Faster to deploy than full immersion — useful for compressed schedules.
- Operationally familiar to teams running existing air-cooled estates.
Technical specifications
What you're committing to.
Production specifications across active DTC deployments in our Riyadh facility.
- Cold-plate technology
- Cu cold plates, treated water/glycol secondary loop
- Rack density
- Up to 80 kW
- Annualized PUE
- 1.10–1.15
- Heat capture at chip
- 70–80%
- Coolant distribution units
- In-row CDUs, N+1
- Compatible hardware
- NVIDIA H100, H200, custom DTC SKUs
- Deployment lead time
- 6–10 weeks per pod, post-design freeze
- Certifications
- Uptime Tier III · ASHRAE-aligned
How it works
The DTC cooling stack.
Cold plates on chip
Copper cold plates clamp directly to GPUs and CPUs. Microchannels carry coolant through the highest-energy components.
Manifold + secondary loop
Quick-disconnect manifold routes treated coolant to each rack. Pressure-balanced flow ensures uniform thermal envelope.
In-row CDU
Coolant distribution unit transfers heat from the secondary loop to the chilled-water primary. N+1 configuration.
Hybrid air assist
Modest in-row air handling captures ancillary heat from drives and NICs that don’t justify cold plates.
Live-serviceable
Quick-disconnects let technicians swap servers without draining the loop. Standard rack maintenance windows.
Telemetry
Coolant flow, return temperature, and pressure sensors feed Finblade for predictive analytics and alerting.
Use cases
Where DTC is the right answer.
GPU inference at scale
High-throughput H100/H200 inference fleets where 40–80 kW racks deliver the right TCO without immersion overhead.
Learn moreMid-density training pods
Training workloads up to ~80 kW per rack where operational familiarity matters as much as raw thermal headroom.
Learn moreEnterprise AI retrofits
Enterprise teams retrofitting existing halls to host higher-density GPU workloads without a full immersion rebuild.
Learn more“DTC is the answer for most enterprise teams entering the AI era. It gets you to 80 kW with engineering that your operations team already understands.”
Capacity is being committed daily.
Reserve yours.
Tour the facility. Talk to engineering. Get a precision quote. We respond to qualified inquiries within one business day.
Olaya District, Riyadh · Open for tours by appointment