GPUs, on your terms.
When shared capacity is not enough: single-tenant NVIDIA B200, B300, and GB300 NVL72 clusters on a non-blocking InfiniBand fabric, with availability commitments written into your agreement. When the capacity does not exist yet, we build it.
Rent
Capacity for a training run, a quarter, or a launch. Scale up for the job and hand it back when it’s done.
Reserve
Dedicated clusters held for you, with availability and support terms written into the contract.
Buy
Current-generation hardware sourced for you, then run by us or handed over to your team.
Tell us the scale and the timeline, and we’ll come back with how and how fast. Reserve GPUs →
Current-generation NVIDIA systems
| Attribute | B200 | B300 | GB300 NVL72 |
|---|---|---|---|
| GPU memory | 192 GB HBM3e | 288 GB HBM3e | 288 GB HBM3e · ~20 TB/rack |
| Configuration | 8× HGX per node | 8× HGX per node | 72 GPUs + 36 Grace CPUs per rack |
| Scale-out fabric | 800G InfiniBand XDR | 800G InfiniBand XDR | 800G InfiniBand XDR |
| In-rack interconnect | NVLink | NVLink | 5th-gen NVLink · 130 TB/s |
| Power | ~40 kW / rack | ~45 kW / rack | ~132–140 kW / rack |
| Deploy unit | 16 to 10,000+ nodes | 16 to 10,000+ nodes | 1 to 16+ racks |
| Best for | Large-scale training and inference | Memory-heavy training, long context | Frontier-scale training |
GPU cluster or a new data center?
Choose by whether the capacity already exists. Buildouts →
| Attribute | GPU Clusters | Data Centers |
|---|---|---|
| What it is | A managed cluster deployed into ready capacity | New capacity engineered to your specification |
| Best when | You need GPUs running soon | The region, facility, or density does not exist yet |
| Model | Turnkey, operated by Aurora | Bespoke, engineered end to end |
| Timeline | Weeks | A full buildout, committed at scoping |
| Start with | Reserve GPUs | Scope a buildout |
Tell us the scale. We’ll tell you how fast.
Rent for a run, reserve a cluster, or buy the hardware outright. One conversation covers all three.