The NVIDIA GB300 NVL72 is one liquid-cooled rack with 72 Blackwell Ultra GPUs and 36 Grace CPUs in a single NVLink domain. It carries 20 TB of HBM3e at up to 576 TB/s and delivers 1,080 PFLOPS of dense FP4, 1.5x the GB200 NVL72. Each GPU has about 288 GB of memory and 8 TB/s of bandwidth. NVIDIA sells it by the rack with no public price. The same GPU rents alone as the B300 for $7.10 to $8.99 per hour.
NVIDIA GB300 Specs
NVIDIA publishes GB300 numbers for the whole GB300 NVL72 rack. The per-GPU column divides the rack totals by 72. NVIDIA quotes Tensor Core figures with sparsity unless marked, so the FP4 row shows both.
| Spec | GB300 NVL72 rack | Per Blackwell Ultra GPU |
|---|---|---|
| Configuration | 72 Blackwell Ultra GPUs, 36 Grace CPUs | 1 GPU |
| GPU memory | 20 TB HBM3e | ~279 GB (288 GB per Azure) |
| GPU memory bandwidth | Up to 576 TB/s | 8 TB/s |
| NVLink bandwidth | 130 TB/s | 1.8 TB/s |
| FP4 Tensor Core (sparse / dense) | 1,440 / 1,080 PFLOPS | 20 / 15 PFLOPS |
| FP8/FP6 Tensor Core | 720 PFLOPS | 10 PFLOPS |
| FP16/BF16 Tensor Core | 360 PFLOPS | 5 PFLOPS |
| CPU memory | 17 TB LPDDR5X at 14 TB/s | n/a |
| CPU cores | 2,592 Arm Neoverse V2 | n/a |
| FP64 | 100 TFLOPS | ~1.4 TFLOPS |
GB300 vs GB200 vs B200
GB200 and GB300 share the rack layout. Each pairs Grace CPUs with Blackwell-family GPUs, 36 CPUs and 72 GPUs per NVL72, linked at 130 TB/s. The GB300 swaps in Blackwell Ultra GPUs with more HBM and more FP4 throughput, and doubles the per-GPU network link. The B200 is the air- or liquid-cooled eight-GPU x86 option, covered on the NVIDIA B200 page.
| Spec | GB300 NVL72 | GB200 NVL72 | B200 (per GPU) |
|---|---|---|---|
| GPU | Blackwell Ultra | Blackwell | Blackwell |
| HBM per GPU | 288 GB | 192 GB | 180 GB |
| Dense FP4 per rack | 1,080 PFLOPS | 720 PFLOPS | 9 PFLOPS per GPU |
| NVLink domain | 72 GPUs, 130 TB/s | 72 GPUs, 130 TB/s | 8 GPUs, 1.8 TB/s each |
| Back-end network per GPU | 800 Gbps | 400 Gbps | Provider-dependent |
| Attention performance | 2x GB200 (NVIDIA) | Baseline | n/a |
Microsoft measured about 1.1 million tokens per second of LLM inference per ND GB300 v6 rack, about 27% above ND GB200 v6. That is less than the 1.5x FP4 gain because decode speed follows memory bandwidth, and per-GPU bandwidth stays near 8 TB/s on both generations.
GB300 Price
NVIDIA publishes no list price for a GB300 NVL72 rack or a DGX GB300. On October 31, 2026, CoreWeave lists GB300 NVL72 as contact sales. Its public rates for the neighboring systems set the range.
| System | Provider | Price | Per GPU-hour |
|---|---|---|---|
| GB300 NVL72 (4 GPUs per instance) | CoreWeave | Contact sales | n/a |
| GB200 NVL72 (4 GPUs per instance) | CoreWeave | $42.00/hr on demand | $10.50 |
| HGX B300 (8 GPUs) | CoreWeave | $35.84/hr spot | $4.48 |
| HGX B200 (8 GPUs) | CoreWeave | $68.80/hr on demand | $8.60 |
| B300 (1 GPU) | Modal | $0.001972/s | $7.10 |
| B300 288 GB (1 GPU) | RunPod Secure Cloud | $8.99/hr | $8.99 |
| B300 serverless worker | RunPod Serverless | $11.99/hr | $11.99 |
At Modal's B300 rate, 72 GPUs for one hour cost about $511. A rack you reserve for a year trades that flexibility for a contract, and CoreWeave's GB200 rate of $10.50 per GPU-hour shows the premium rack-scale NVLink carries over single cards.
GB300 NVL72 vs HGX B300
Both use Blackwell Ultra GPUs. The difference is how many GPUs share one NVLink domain. HGX B300 is a baseboard of eight GPUs inside an x86 server. GB300 NVL72 links 72 GPUs through NVLink switches, so tensor and expert parallelism can span the whole rack without dropping to the network.
What GB300 Changes for LLM Inference
Two numbers matter for serving. More HBM per GPU (288 GB) means fewer GPUs per replica and more room for KV cache, which raises the batch size a replica can hold. Higher attention throughput helps long-context requests, where attention cost grows with sequence length. Decode tokens per second per user still track memory bandwidth, which is about 8 TB/s per GPU on B200, GB200, and GB300 alike.
For a one-trillion-parameter mixture-of-experts model in FP4, weights take roughly 500 GB, so two B300 GPUs hold them and six more are left for KV cache in one HGX node. Rack-scale NVL72 pays off when you shard experts across dozens of GPUs to serve many concurrent users. Measured B200 and B300 serving numbers for open models are on the dedicated inference benchmarks page.
Rent, Reserve, or Dedicated Endpoint
Rent single B300s by the second for evaluation and bursty jobs. Reserve GB300 NVL72 capacity through a cloud contract when one workload needs the full NVLink domain for months. To serve an open model in production without running the stack, Morph runs reserved B200 and B300 capacity behind an OpenAI-compatible API at $9.98 per B200-hour. See dedicated inference for models and regions.
FAQ
What is the NVIDIA GB300?
GB300 is NVIDIA's Grace Blackwell Ultra platform. The GB300 NVL72 rack holds 72 Blackwell Ultra GPUs and 36 Grace CPUs in one NVLink domain with 130 TB/s of NVLink bandwidth, 20 TB of GPU memory at up to 576 TB/s, and 1,080 PFLOPS of dense FP4. NVIDIA also sells it as DGX GB300, and Microsoft rents it as the ND GB300 v6 virtual machine with four GPUs per VM.
What are the GB300 NVL72 specs?
From NVIDIA's spec sheet: 72 Blackwell Ultra GPUs and 36 Grace CPUs with 2,592 Arm Neoverse V2 cores, 130 TB/s of NVLink bandwidth, 20 TB of GPU memory at up to 576 TB/s, 17 TB of LPDDR5X at 14 TB/s, 37 TB of fast memory in total, 1,440 PFLOPS of sparse FP4 (1,080 dense), 720 PFLOPS of FP8/FP6, 360 PFLOPS of FP16/BF16, and 100 TFLOPS of FP64.
How much memory does a GB300 GPU have?
Microsoft's ND GB300 v6 page lists 288 GB of HBM3E per Blackwell Ultra GPU. CoreWeave lists 279 GB of VRAM per GPU, which matches NVIDIA's rack total of 20 TB divided by 72. Bandwidth works out to 8 TB/s per GPU. A B200 has 180 GB and a GB200 GPU has 192 GB.
GB300 vs GB200: what changed?
Dense FP4 rises 1.5x per rack, from 720 to 1,080 PFLOPS. HBM per GPU rises from 192 GB to 288 GB. Back-end network bandwidth doubles from 400 to 800 Gbps per GPU. NVIDIA says attention performance doubles. NVLink bandwidth stays at 130 TB/s per rack and the 72-GPU, 36-CPU layout is unchanged. Microsoft reports about 1.1 million tokens per second per ND GB300 v6 rack, about 27% above ND GB200 v6.
How much does a GB300 cost?
NVIDIA publishes no list price for a GB300 NVL72 rack, and CoreWeave lists it as contact sales on October 31, 2026. For comparison, CoreWeave rents a four-GPU GB200 NVL72 instance at $42.00 per hour ($10.50 per GPU) and an eight-GPU HGX B300 at $35.84 per hour spot. A single B300, the same Blackwell Ultra GPU outside the rack, rents for $7.10 per GPU-hour on Modal and $8.99 on RunPod Secure Cloud.
Do I need a GB300 NVL72 or is a B300 enough?
The NVL72 matters when one model or one training job must span more than eight GPUs at full NVLink speed, such as trillion-parameter mixture-of-experts serving with wide expert parallelism. If the model and its KV cache fit in eight GPUs, or 2.3 TB at 288 GB per GPU, an HGX B300 node or single B300 GPUs cost less and are available by the hour.
Related Resources
The fastest endpoints are private deployments
Morph's top speeds come from dedicated deployments, not shared public endpoints: speculators trained on your traffic, caching tuned to your workload, and volume discounts over public per-token rates. Over 500 billion tokens per day run this way.
Serve Open Models on Reserved Blackwell Capacity
Morph runs dedicated endpoints on B200 and B300 GPUs behind one OpenAI-compatible API, billed at $9.98 per B200-hour. No rack to reserve or serving stack to tune.
Sources
- NVIDIA: GB300 NVL72 specifications, read October 31, 2026
- NVIDIA: DGX GB300 (1.5x dense FP4 and 2x attention vs DGX GB200)
- Microsoft Learn: ND GB300 v6 series (288 GB per GPU, 800 Gbps, ~1.1M tokens/s per rack)
- CoreWeave pricing (GB300 NVL72, GB200 NVL72, HGX B300, HGX B200), read October 31, 2026
- Modal pricing (B300 and B200 per-second rates), read October 31, 2026
- RunPod pricing (B300 288 GB Secure Cloud and Serverless), read October 31, 2026