Power
Two fully independent paths from utility intake to rack, each able to carry the entire load. UPS and standby generation bridge any interruption.
Engineered for extreme density.
AI accelerators draw more power and shed more heat than any workload before them. Our facilities are designed around them from day one — not retrofitted.
(01)The rack
Anatomy of an
AI rack.
A single accelerator rack can now draw as much power as an entire row did a few years ago. Everything around it — power, cooling, network — is sized to that reality.
- 1Redundant 2N power feeds
- 2800G fabric uplinks
- 3Direct-to-chip cold plates
- 4Supply & return manifolds
- 5Coolant distribution unit
- 6High-density GPU trays
130kWPer-rack design envelope
Direct-to-chip liquid cooling built for next-generation accelerator racks.
≤1.3PUEDesign efficiency target
Liquid cooling loops, economiser hours and heat-reuse readiness.
800GNetwork fabric
Non-blocking InfiniBand and RoCE fabrics for large-scale training.
2NpowerCritical power architecture
Fully redundant power paths from utility intake to rack.
99.99%Availability design target
Concurrently maintainable design — maintenance without downtime.
24/7NOCOperations
On-site engineers and a remote network operations centre.
Figures are design targets and may vary by facility and deployment.
(02)Facility design
Four systems,
one design.
Power, cooling, network and security are engineered together, so no single system becomes the limit on the others.
Cooling
Direct-to-chip liquid cooling removes most heat at the source. Redundant coolant distribution units and air systems handle the remainder.
Network
Non-blocking fabrics of up to 800G between accelerators, and diverse, carrier-neutral fibre entry to the outside world.
Security
Layered physical security from perimeter to rack, with on-site staff and continuous monitoring at every layer.
(03)Grid to GPU
Every watt in,
every watt out.
Electricity arrives over two independent paths, does its work in the accelerator, and leaves as heat through a closed liquid loop.
Power path
- Utility intakeDual feeds
- TransformersHigh to low voltage
- UPS & generators2N redundancy
- BuswayA + B to every rack
- GPU rack≤130 kW
Cooling loop
- Cold platesAt the chip
- ManifoldsSupply & return
- CDUN+1 redundancy
- Facility waterClosed loop
- Heat rejectionReuse-ready
(04)Efficiency
Efficient
by design.
Power usage effectiveness (PUE) measures how much energy a facility spends on everything other than computing. Liquid cooling lets us design well below the industry average.
Industry average1.54
Our design target≤1.30
Theoretical ideal1.00
IT loadCooling & power overhead
01
Liquid first
Water carries heat far more effectively than air. Cooling at the chip cuts fan and chiller energy across the hall.
02
Heat-reuse ready
Warm return loops are designed so waste heat can be recovered wherever there is demand for it.
03
Cleaner power
A procurement roadmap that favours lower-carbon electricity as it becomes available.