New — commissioning
HGX B200
The HGX B200 is the current Blackwell training node: eight GPUs, 180 GB of HBM3e each, and an 800G Quantum-X800 XDR fabric that keeps a large all-reduce from becoming the bottleneck.
Specification
- GPU memory
- 180 GB HBM3e · 7.7 TB/s
- Intra-node
- NVLink 5 · 1.8 TB/s per GPU
- Inter-node
- Quantum-X800 XDR 800G
- Node config
- 8 GPU · 30 TB NVMe · 2 TB RAM
- Best for
- Frontier training, FP4 inference
B200 pods run Quantum-X800 InfiniBand XDR at 800G per link, 1:1 on every tier. SHARP in-network reduction is available, and inter-node traffic is not billed.
Why this machine
B200 is what you reserve when the run is large enough that fabric and goodput decide the schedule, not raw FLOPs. NVLink 5 moves 1.8 TB/s per GPU inside the node, and Quantum-X800 XDR carries 800G per link between them — rail-optimized, non-blocking, and dedicated to your tenancy rather than blended with anyone else’s east-west traffic.
Because there is no hypervisor, the FP4 and FP8 paths are yours at full rate and the p99 has no neighbor VM in it. On identical Hopper hardware we measure 5–7% higher sustained MFU than virtualized instances; the same argument applies here, and the 72-hour burn-in report tells you what your specific pod achieved before you sign for it.
What it is for
Frontier and foundation training
Reserved pods from 512 to 10,000+ GPUs on dedicated XDR fabric with in-rack spares. On a six-week run, the difference between 97% and 91% goodput is the entire margin.
FP4 and FP8 inference at scale
Blackwell’s low-precision throughput without a virtualization layer skimming it, and without jitter from co-tenants on the same host.
Large-scale fine-tuning
Whole nodes on short reserved blocks, WEKA-backed datasets that persist between runs, per-minute metering inside the reservation.
HGX B200 questions
How much does an HGX B200 node cost?
From $39 per 8-GPU node-hour on a yearly reservation, or from $56 on a monthly term. That is the whole machine — 8 GPUs, 30 TB of NVMe and 2 TB of RAM — not a per-GPU slice.
How does B200 compare to B300 here?
B300 carries 288 GB of HBM3e per GPU against the B200’s 180 GB, at 8 TB/s versus 7.7 TB/s, and lists from $45 rather than $39 per node-hour. If your bottleneck is memory capacity per GPU, B300 is the answer; if it is fabric and cost per FLOP, B200 is.
Is the B200 capacity available now?
B200 pods are commissioning. New capacity is commissioned in six-week blocks with the 72-hour burn-in included, so a delivery date comes with the quote.
Can I bring my own scheduler?
Yes. You get root on the machine — custom kernels, your own drivers, NIC firmware and RDMA tuning. Slurm and Kubernetes control planes are managed at $0 if you want them, and the BIOS is never locked.