At general assembly stage early partner enquiries open

Modular AI data centres · Designed to Tier III

Rack-scale AI,
built as a pod.

Each EVADAUN pod is its own data centre: every power and cooling path has a backup, so it keeps running while any part is serviced. Liquid-cooled for NVIDIA GB300 NVL72 racks, sized from day one for Vera Rubin NVL72.

Roof cooling deck A-side power AI pod, 1 to 8 racks B-side power
Illustration, not to scale

Start with GB300 NVL72 today and move to Vera Rubin NVL72 tomorrow without a redesign. Power, cooling and structure are sized for the next rack generation from day one, and every pod will be tested under load before it ships.

Electrical bay Up to 8 NVL72 racks Network bay CDU bay
Illustration

AI pod

One self-contained pod for one to eight NVIDIA GB300 NVL72 racks, with its own electrical, network and cooling bays. Start with one rack and add more in place.

Electrical and service bay

Switchgear, PDUs, the UPS interface, fire suppression, leak detection and building management.

Network bay

Structured fibre, switching and cable management, ready for the cluster fabric.

CDU and mechanical bay

Coolant distribution with N+1 pumps, filters, valves and manifolds.

Service access

Front and rear access to every rack, so one rack can come out while the others keep running.

Transformer Switchgear Busway to the pod
Illustration

Power modules

Separate A-side and B-side power modules stand at either end of the pod and feed it on two independent paths.

Two independent paths

Either path is designed to carry the pod while the other is serviced.

UPS and energy storage

Ride-through for the compute load, with space to add battery storage on site.

Grid, gas or both

Designed to take grid power, on-site generation or a mix, so a pod can start before the full grid connection lands.

Dry coolers on the roof Warm-water loop CDU bay inside the pod
Illustration

Liquid cooling

Warm-water direct-to-chip cooling takes the heat at the source, and a dry-cooler deck on the pod's own roof rejects it.

Coolant distribution

N+1 pumps run the secondary loop to every rack manifold inside the pod.

On the roof

The cooler deck travels on its own truck and is lifted onto the pod on site, so the whole set needs much less ground.

Warm water

Supply water at 30 to 36 °C lets the roof coolers do most of the work, with water spray on the hottest days and chillers only where sites run hot.

Heat reuse

Connection points for district heating or an industrial neighbour that can use the heat.

Substation Pods Fibre spine Next phase
Illustration

Multi-pod campus

Identical pods join into one cluster over an NVIDIA fabric, one pod at a time.

Cluster fabric

Each pod's leaf switches connect to a shared spine on NVIDIA Quantum-X800 InfiniBand or Spectrum-X Ethernet.

Customer and edge access

Customer, storage, management and data-centre links run on a separate network.

Phased expansion

Start with the first pod and add the next as the grid connection or generation grows.

GPU cloud platform

A planned EVADAUN software stack that turns each pod into a self-service GPU cloud, the same whether a pod runs one rack or eight.

  1. 07Customer experienceSelf-service portal, API and CLI, notebooks, training jobs, inference endpoints, dashboards
  2. 06Platform servicesIdentity and access, secrets, security policy, monitoring, logging, metering and billing
  3. 05Data and storageHigh-performance shared storage, object storage, NVMe cache, model and artifact registries
  4. 04NVIDIA softwareDrivers, CUDA, NCCL, GPU operator, container toolkit, Triton inference server
  5. 03Cluster and schedulingKubernetes and Slurm, queues and fair share, multi-tenancy and quotas
  6. 02Bare metal and automationProvisioning, out-of-band management, infrastructure as code, firmware lifecycle
  7. 01Physical infrastructureGB300 NVL72 pods, NVIDIA fabric, redundant power and liquid cooling

Where a pod fits

Pods make the most sense where a new building would slow you down.

Bare powered land

Grid or gas at the fence and no building yet. The first pod can run while the campus grows around it.

Brownfield sites

Former mills, plants and yards that still have heavy power and gas connections.

Sovereign and edge AI

In-country GPU capacity close to the people who use it, from one rack to a campus.

Design targets

From the September 2026 design. Final values follow detailed engineering and certification.

Racks
1 to 8 NVIDIA GB300 NVL72 racks per pod; start with one and add more in place
GPU load
About 1.14 MW with eight GB300 NVL72 racks, about 142 kW each
Rubin-ready
Power, cooling and structure sized for about 1.9 MW with Vera Rubin NVL72, a 240 kW per rack allowance, using the warmer water Rubin is designed for
Power
2N: separate A-side and B-side power modules, each designed to carry the pod
Cooling
Warm-water direct-to-chip, 30 to 36 °C supply, N+1 pumps; dry coolers on the pod's roof with water spray for hot days, chillers where sites run hot; air handling for residual heat
Network
NVIDIA Quantum-X800 InfiniBand or Spectrum-X Ethernet, with a dedicated network bay
Resilience
Designed to Tier III: concurrently maintainable, with N+N power and cooling paths; fire, leak and building-management monitoring
Deployment
Arrives by road, with the cooler deck on its own truck; plug in power, water and network
Certification
Route to be set during detailed engineering; CE and UKCA marking planned
Start with 2 racks Room for 6 more
Illustration

Start small, scale in place, then scale out

  1. Launch with a lightly populated pod: one or two GB300 NVL72 racks.

  2. Add racks inside the same pod, up to four, with no re-architecting.

  3. Up to eight racks per pod; power, cooling and fabric are already sized for Rubin.

  4. Add pods for larger clusters, more regions or higher availability.

Put AI capacity where the power is.

Tell us about your site, your power and your timeline.

info@evadaun.com