Reservations open
GB300
NVIDIA NVL72
- GPUs
- 72x Blackwell Ultra per rack
- Memory
- 288 GB HBM3e per GPU
- Fabric
- NVLink 5 / Quantum-X800
Reserve GPU Capacity
Managed training and inference workloads with serverless or dedicated compute. From 64 GPUs to multi-thousand-GPU islands, configured around your workload.
64 to 10k+
GPUs
1 mo+
Terms
24 / 7
Support
Capacity request
The machines that matter, in stock.
Current NVIDIA training platforms from our own racks and a vetted provider network, with status updated as capacity comes online.
Rack-scale / NVL72 systems
Available now
GB200
NVIDIA NVL72
- GPUs
- 72x Blackwell per rack
- Memory
- 192 GB HBM3e per GPU
- Fabric
- NVLink 5 / Quantum-X800
Node-scale / HGX systems
B200
NVIDIA HGX
Available now
- GPUs
- 8x B200 per node
- Memory
- 180 GB HBM3e per GPU
- Fabric
- 400G Quantum-2 InfiniBand
H200
NVIDIA HGX
Available now
- GPUs
- 8x H200 per node
- Memory
- 141 GB HBM3e per GPU
- Fabric
- 400G Quantum-2 InfiniBand
H100
NVIDIA HGX
Available now
- GPUs
- 8x H100 per node
- Memory
- 80 GB HBM3 per GPU
- Fabric
- 3.2 Tb/s InfiniBand per node
Need another node count, storage ratio, or generation? Share the workload and we will configure the cluster around it.
First in line for every new generation.
NVIDIA is an investor and our closest engineering partner. In a market where frontier GPUs are allocated before they ship, that partnership is why the fleet above reads available, and it carries through how every cluster is designed, built, and supported.
01
Early allocation
Each new platform generation is allocated to us as it ships, so reservations are fulfilled on silicon timelines, not the resale market.
02
Reference architecture
Every cluster is built and burned in to NVIDIA reference designs and validated before handover.
03
Direct engineering support
When an issue needs NVIDIA's attention, it reaches NVIDIA engineering through the partnership, not a vendor queue.
Customer zero was our own frontier lab.
Before this platform ran an external workload, it trained Reka's multimodal frontier models end to end. Link flaps, straggler nodes, checkpoint stalls, and silent data corruption were found and fixed on our runs, not yours.
10,000+
GPUs in one training fabric
99.2%
measured training goodput
24 / 7
operator coverage
Built for training. Tuned for goodput.
The machine is the floor. What you are buying is everything that keeps it busy.
Bare metal
No hypervisor between your job and the silicon. You get the whole machine, and every FLOP you pay for.
Non-blocking fabric
Full-bisection, rail-optimized InfiniBand built for collective operations at cluster scale.
Parallel storage
High-throughput storage sits beside the compute so checkpoints do not gate step time.
Scheduler ready
Slurm or Kubernetes configured, images loaded, and a reference job completed before handoff.
Goodput SLAs
Hot spares, uptime commitments, and reporting on the hours your workload actually trained.
24/7 training engineers
Around-the-clock coverage from engineers who operate frontier training clusters themselves.
Reservation to first job, in four steps.
Scope
We size compute, fabric, storage, and regions around your workload.
Reserve
Lock the capacity against a firm go-live date, from one month to multi-year.
Provision
We configure your scheduler, images, networking, access, and observability.
Operate
Start on a burned-in cluster with root access and live engineering support.
What infrastructure teams ask first.
Is this bare metal or virtualized?+
Bare metal. Your jobs run directly on the machines, with root access and the full fabric available to your workload.
How quickly can a cluster be live?+
Standard HGX blocks can be ready in days. Custom scheduler, image, storage, and networking layouts typically take a few weeks after signature.
What reservation terms are available?+
One month to multi-year. Reserve for a single training run or lock a longer allocation with a defined go-live date.
What is included in the price?+
Compute, power, cooling, storage, fabric, and support are included in the GPU-hour rate, with no ingress or egress fees.
What happens when nodes fail?+
Hot spares remain in the same fabric, failed nodes are swapped quickly, and downtime is credited.
Who operates the infrastructure?+
Reka Cloud does, 24/7. The same engineering discipline used to train frontier multimodal models is applied to your cluster.
Put a go-live date on the cluster.
Share the workload and timeline. We will return with concrete capacity, topology, and commercial options.
Reserve capacity