NVIDIA Grace Blackwell Ultra

NVIDIA GB300 GPU cloud

Reserve NVIDIA GB300 NVL72 capacity on VESSL Cloud: 72 Blackwell Ultra GPUs and 36 Grace CPUs per rack, with the GPUs on one 130 TB/s NVLink domain for cluster-scale training and high-concurrency reasoning inference.

Reserve capacity See pricing & availability

NVIDIA GB300 NVL72

GPU memory: up to 288GB HBM3e
Memory bandwidth: 8 TB/s

Technical specifications

Architecture: Blackwell
GPU memory: up to 288GB HBM3e
Memory bandwidth: 8 TB/s
NVLink: 1.8 TB/s
FP8 (Tensor): 10 PFLOPS
FP4 (Tensor): 20 PFLOPS
Max TDP: 1,400W
GPUs per rack: 72 + 36 Grace CPUs (NVL72)

*Peak performance with sparsity, per NVIDIA official specs. Final specs may vary by node configuration.

Pricing & availability

NVIDIA GB300 NVL72Reserved only

Talk to sales

What's the GB300 best for?

Cluster-scale training

A full NVL72 rack behaves like one giant accelerator: roughly 20 TB of HBM3e and a 130 TB/s NVLink domain keep trillion-parameter runs resident without hopping across slower networks.

Rack-scale reasoning inference

The unified NVLink domain spreads massive KV caches across 72 GPUs, sustaining high-concurrency agentic and long-context serving.

CPU-heavy AI pipelines

Each Grace CPU connects to its GPUs over NVLink-C2C, so data preprocessing and inference-time orchestration feed accelerators without PCIe bottlenecks.

Compare NVIDIA data-center GPUs

	A100 Ampere	H100 Hopper	H200 Hopper	B200 Blackwell	B300 Blackwell	GB300 You're viewing
Architecture	Ampere	Hopper	Hopper	Blackwell	Blackwell	Blackwell
GPU memory	80GB HBM2e	80GB HBM3	141GB HBM3e	192GB HBM3e	up to 288GB HBM3e	up to 288GB HBM3e
Memory bandwidth	2.0 TB/s	3.35 TB/s	4.8 TB/s	8 TB/s	8 TB/s	8 TB/s
FP8 (Tensor)	Not supported	3,958 TFLOPS	3,958 TFLOPS	9 PFLOPS	9 PFLOPS	10 PFLOPS
Access	from $1.77/hr	from $3.57/hr	Reserved only	Reserved only	Reserved only	Reserved only
Best for	Cost-efficient fine-tuning & inference	Cost-efficient training & inference	Long-context & large-model inference	Frontier-scale training (FP4)	Largest models & reasoning inference	Cluster-scale training & inference

Why AI teams run GPUs on VESSL Cloud

From A100 to Blackwell Ultra

Run A100 and H100 on demand. Reserve H200 and Blackwell GPUs for larger runs, so each workload gets the GPU it needs.

Issues caught early

Every VM Cluster node is health-checked continuously, with its status refreshed about once a minute. Workspaces and Jobs alert you when a run crosses a resource threshold or hits a runtime issue.

From one node to a cluster

Train and serve on up to 8 GPUs on one node, and move to a multi-node VM Cluster when one node is not enough.

Enterprise-ready

SOC 2 Type II compliance, with dedicated support for production AI.

Frequently asked questions

How do I get access to NVIDIA GB300 GPUs?

GB300 capacity is reserved by the NVL72 rack and allocated on request. Talk to our team for current availability and terms.

What is the GB300 NVL72?

A rack-scale system with 72 Blackwell Ultra GPUs and 36 Grace CPUs. The 72 GPUs share one 130 TB/s NVLink domain, so the whole rack behaves like one large accelerator.

What's the difference between the B300 and GB300?

Both use the Blackwell Ultra GPU. The B300 ships as 8-GPU HGX nodes. The GB300 is an NVL72 rack built for cluster-scale workloads: 72 GPUs on one 130 TB/s NVLink domain, plus 36 Grace CPUs.