An AI cloud region, delivered to your data center.
Nimbo Rack is a pre-integrated GPU rack that we ship, install and operate inside your facility. It joins Nimbo Fleet on day one, serves the same OpenAI-compatible API as the Nimbo cloud, and bursts to Nimbo capacity when it runs full.
Three racks, sized by power and workload
Power and cooling decide what fits in your room. We start every engagement with a site survey of available kW per rack, floor loading and heat rejection.
Branch and regulated sites
- GPUs
- 16× NVIDIA RTX PRO 6000 Blackwell Server Edition, 96 GB each
- Power
- about 12 to 15 kW
- Cooling
- Air
- For
- Inference up to ~120B-parameter models, RAG, document AI
Production inference and fine-tuning
- GPUs
- 32× NVIDIA B200 or B300 in 4 HGX nodes, or AMD MI355X
- Power
- about 55 to 60 kW
- Cooling
- Rear-door heat exchanger or direct liquid
- For
- Large-model serving, fine-tuning, private model hosting
Rack-scale training
- GPUs
- 72× NVIDIA Blackwell Ultra in a GB300 NVL72 rack
- Power
- about 130 to 140 kW
- Cooling
- Direct liquid, facility water required
- For
- Pre-training, large fine-tunes, reasoning-model serving
Power figures are planning estimates from vendor system specifications and vary with configuration. Rack bill of materials and OEM partners to confirm.
Everything between the PDU and the API
Edge
Two 4U GPU servers, a 400G switch, local storage and dual PDUs. Fits a standard 42U rack with room to grow.
Inference
Four liquid-cooled 4U HGX nodes on a rail-optimised InfiniBand or RoCE fabric, with shared storage, an in-rack coolant distribution unit and redundant power.
The rack runs itself. We watch it anyway.
Every rack streams GPU, NVLink, fabric, power and thermal telemetry to Nimbo operations. Models trained on that telemetry flag failing parts before they take a job down.
- Predictive part replacementGPU memory errors, link flaps and fan or pump drift trigger a replacement before failure.
- Power and thermal tuningWorkloads are placed and power-capped to stay inside the kW your room can deliver.
- Remote operations 24×7Firmware, drivers and cluster upgrades applied as audited Fleet workflows, with your approval gates.
- Burst to the cloudWhen the rack is full, policy-approved jobs run on Nimbo capacity under the same API key.
- Data stays on siteInference and training data never leave your facility unless a policy says so.
Buy it, lease it, or pay per GPU-month
Rack as a service
No capex. A monthly fee per installed GPU covers hardware, software and operations, on 36 to 60-month terms.
$ per GPU-month
Lease
Financed through our leasing partners, with Nimbo operations and software on a separate subscription.
partners to confirm
Purchase
You own the hardware. Fleet, Secure and operations are licensed per GPU per month.
$ per GPU-month
Find out what your room can hold.
Send us your available power per rack and cooling type. We will propose a configuration and a delivery date.