NVIDIA Blackwell · Pre-order

Blackwell B200 — reserve your slot.

HGX B200 brings 192GB of HBM3e per GPU, FP4 Transformer Engine v2, and 5th-gen NVLink at 1.8 TB/s. Reserve 2026 allocation today — Blackwell supply is gated and prioritized by commitment.

FP4 Transformer Engine v2

Native 4-bit floating point with auto-scaling — up to 15x H100 inference on large transformer models.

192 GB HBM3e per GPU

1.5 TB HBM3e in an 8-GPU node. Run 405B-class models in higher precision with full KV cache headroom.

5th-gen NVLink @ 1.8 TB/s

Double H100 NVLink bandwidth. Tensor-parallel scaling stays linear well past 8 GPUs in a single domain.

Specifications

GPU (HGX)
8 × NVIDIA B200 SXM
192GB HBM3e · 8 TB/s each
CPU
2 × Intel Xeon 6 / Grace optional
PCIe 5.0/CXL
Memory
Up to 4 TB DDR5
Configurable per workload
Storage
Up to 122 TB NVMe
Gen5 scratch tier
Fabric
8 × NDR/XDR 400–800G IB
Quantum-X800 ready
Power
10 kW+ per node
Liquid-cooling option
Best fit: frontier model training

If you're training 500B+ parameter models, Blackwell is the platform — and getting allocation in 2026 requires reserving now.

Best fit: high-throughput FP4 inference

Quantization to FP4 with no accuracy loss on most LLMs gives 4x serving density vs H100 FP8.

Forward-compatible stack

Your H100 CUDA code runs on B200. Migrate when allocation arrives — no architectural rewrite.

Frontier training: 500B+ parameter models

Blackwell was designed for the next generation of foundation models — multi-trillion-parameter dense transformers, sparse MoE architectures with hundreds of experts, and long-horizon RL post-training. 5th-gen NVLink at 1.8 TB/s keeps tensor-parallel groups linear well past a single 8-GPU domain, and NVL72 rack-scale options extend that to 72 GPUs in one coherent memory pool.

Trillion-parameter dense models

1.5 TB of HBM3e per 8-GPU node fits 1T+ parameter models in FP8 without offload. FP4 doubles that effective capacity again for inference.

Sparse MoE at scale

Wider expert routing benefits directly from doubled NVLink bandwidth. All-to-all collectives that bottlenecked on H100 disappear on Blackwell.

RLHF, DPO, long-horizon RL

Reward-model + policy + reference all resident in HBM. PPO/GRPO rollouts no longer dominated by parameter shuffling.

Transformer Engine v2

Auto-scaling FP4/FP6/FP8 with no accuracy loss on most architectures — typical 2.5x training throughput over H100 on transformer workloads.

High-throughput agentic and FP4 inference

Production AI in 2026 is dominated by agentic loops, reasoning models that emit thousands of tokens per request, and multimodal serving. FP4 with Transformer Engine v2 gives Blackwell up to 15x H100 inference throughput on large transformers — turning $/Mtok economics on its head.

Reasoning-model serving

o1/R1-style models with long chain-of-thought benefit most from FP4 throughput. Serve 4x more concurrent reasoning sessions per node vs H100 FP8.

Million-token contexts

192GB HBM3e per GPU keeps massive KV caches resident. Document-scale and codebase-scale agents stay in memory across turns.

Multimodal at production scale

Vision-language, video understanding, and speech models share HBM with text decoders on a single node — no cross-GPU activation transfers.

vLLM, TRT-LLM, SGLang ready

Forward-compatible with your H100 inference stack. Migrate when allocation lands; no architectural rewrite required.

Privacy, security, and sovereign AI

Blackwell allocation is scarce and customers training on it are usually working with their most sensitive data — proprietary research, regulated industries, or sovereign-AI mandates. We rent single-tenant bare-metal Blackwell so the only software touching that HBM is yours.

Single-tenant bare metal

No hypervisor, no neighbor workloads. Confidential Computing modes on Blackwell available for attested training and inference.

Customer-held keys

Full-disk encryption with keys you generate and rotate. NIST 800-88 drive sanitization between tenants. SOC 2 and HIPAA-ready deployments.

On-prem & sovereign deployments

Identical HGX B200 / NVL72 builds deployable in your facility or jurisdiction — useful for EU AI Act, ITAR, FedRAMP, and sovereign-AI programs.

Private fabric by default

Dedicated VLANs, no public egress unless requested, and Direct Connect / ExpressRoute peering for training data and model weights.

Frequently asked

When will B200 be available to rent?+

We're taking reservations now for HGX B200 systems shipping in 2026. Early allocation is prioritized by term length and use case — talk to sales about your timeline.

How much faster is B200 than H100?+

NVIDIA quotes ~2.5x training throughput and ~15x inference throughput on FP4 for large transformer models. Real-world speedups depend heavily on whether your stack uses FP4/FP6 and the new Transformer Engine v2.

Should I wait for B200 or rent H100/H200 now?+

If you're shipping models in the next 6 months, rent Hopper now and migrate later — the software stacks (CUDA, NCCL, vLLM, TensorRT-LLM) port forward cleanly. If you're 12+ months out, reserving B200 allocation today is reasonable.

Lock in 2026 Blackwell allocation.

Slots are filling quickly. Tell us your scale target — we'll secure the right configuration.

Reserve B200