NVIDIA GB300 NVL72
Seventy-two Blackwell Ultra GPUs in one rack and one NVLink domain. The platform of the first hall.
- 72 GPUs per rack
- 20 TB GPU memory
- Liquid cooled rack
- Available Q1 2027
Configurations
Pick the footprint that fits. Every one is sold on the same contract.
- 72 GPUsOne rack
- One NVLink domain. The smallest unit, sold on its own.
- 144 to 792 GPUsTwo to eleven racks
- Joined by the InfiniBand fabric into one cluster.
- 864 GPUsTwelve racks
- The whole first hall, one training cluster.
Specifications
- GPUs
- 72 Blackwell Ultra per rack
- GPU memory
- 20 TB, 288 GB HBM3e per GPU
- NVLink
- 1.8 TB/s per GPU, 130 TB/s domain
- CPUs
- 36 Grace, 2,592 Arm cores
- System memory
- 17 TB LPDDR5X
- Networking
- ConnectX-8, 800 Gb/s per GPU; BlueField-3
- Power
- 132 to 140 kW rated, 155 kW peak
- Cooling
- Liquid cooled rack
- Fabric at Fulmis
- InfiniBand, Quantum-X800, rail optimised
- Software
- Slurm and managed Kubernetes
- Location
- A data centre in the Santiago Metropolitan Region
- Available
- Q1 2027
Performance
Per rack, as NVIDIA publishes it.
- FP4 Tensor, dense
- 1,080 PFLOPS
- FP8 / FP6 Tensor
- 720 PFLOPS
- FP16 / BF16 Tensor
- 360 PFLOPS
- TF32 Tensor
- 180 PFLOPS
- GPU memory
- 20 TB HBM3e, up to 576 TB/s
- Fast memory, GPU and CPU
- 37 TB
- NVLink bandwidth
- 130 TB/s
- Scale-out network
- ConnectX-8, 800 Gb/s per GPU
- Rack power
- 132 to 140 kW rated, 155 kW peak; NVIDIA's reference architecture says up to 142 kW
NVIDIA's published figures. Tensor Core numbers carry sparsity unless the row says dense; H100, H200 and RTX PRO figures are per GPU because NVIDIA publishes no node total. Power figures are the ones we reserve and protect to, with NVIDIA's own where it publishes one.
How we deploy it
GB300 NVL72 is the rack-scale Blackwell Ultra system: eighteen compute trays and nine NVLink switch trays joined into one 72-GPU NVLink domain. The first Fulmis hall is twelve of them in one Santiago data centre, joined by an InfiniBand Quantum-X800 fabric into a single training cluster.
A rack is the smallest unit. One rack is one NVLink domain and it is sold on its own; twelve is the whole hall. Each position is reserved at 140 kW and protected to 155 kW.
Slurm and managed Kubernetes run on top. Storage is a parallel file system sized to the workload. Acceptance is a burn-in against the vendor baseline, signed before the term starts.
Fulmis places the purchase order, or takes delivery of the fleet you already own. Either way we bring it into Chile, commission it against the vendor baseline and operate it for the term. You reserve the capacity, or your own hardware; the workload and the data stay yours.
Terms
- Term
- Three to five years
- Price
- Fixed for the term, in the term sheet
- Commitment
- Take-or-pay
- Prepayment
- None to 25% at signature
- Smallest block
- One rack, 72 GPUs
- Start
- Q1 2027, first slot
Other platforms
| Platform | GPUs | GPU memory | NVLink | Power | Cooling | Smallest block | Available |
|---|---|---|---|---|---|---|---|
| GB200 NVL72Rack scale | 72per rack | 13.4 TB186 GB per GPU | 130 TB/s1.8 TB/s per GPU | 120 to 132 kWper rack | Liquid | 1 rack, 72 GPUs | Q1 2027 |
| HGX B3008-GPU node | 8per node | 2.3 TB288 GB per GPU | 14.4 TB/s1.8 TB/s per GPU | About 14 kWper node | Air or liquid | 1 node, 8 GPUs | December 2026 |
| HGX B2008-GPU node | 8per node | 1.4 TB180 GB per GPU | 14.4 TB/s1.8 TB/s per GPU | About 14.3 kWper node | Air or liquid | 1 node, 8 GPUs | December 2026 |
| HGX H2008-GPU node | 8per node | 1.1 TB141 GB per GPU | 7.2 TB/s900 GB/s per GPU | About 10.2 kWper node | Air | 1 node, 8 GPUs | Q1 2027 |
| HGX H1008-GPU node | 8per node | 640 GB80 GB per GPU | 7.2 TB/s900 GB/s per GPU | About 10.2 kWper node | Air | 1 node, 8 GPUs | Q1 2027 |
| RTX PRO 6000 Blackwell8-GPU node | 8per node | 768 GB96 GB per GPU | PCIe Gen 5no NVLink | 5.4 to 6.7 kWper node | Air | 1 node, 8 GPUs | Q1 2027 |
| Vera RubinRack scale | 72per rack | HBM4not yet published | Not published | Not published | Liquid | 1 rack | On request |