NVIDIA · Blackwell · 2025
NVIDIA GB300. 288 GB HBM3e, 2500 TFLOPS FP16 tensor.
GB300 NVL72 racks are in stock at $14.30 per GPU-hour on a 3-month term, prepaid in full. 288 GB of HBM3e at 8 TB/s per GPU, for frontier-model pretraining.
GB300 NVL72 rack
In stock$14.30per GPU-hour
- Term
- 3 months
- Payment
- Prepaid in full
- GPUs per rack
- 72
The card
288 GB of HBM3e. At up to 8.0 TB/s.
- Memory
- 288 GB
- HBM3e
- Bandwidth
- up to 8.0 TB/s
- Memory to compute
- FP16 tensor
- 2500
- TFLOPS
- Per node
- 8 GPUs
- NVLink, up to 1400 W each
Workloads
What teams run on the GB300. Sized for this card.
1T Parameter Training1T-param training state (about 16 TB) fits in one NVL72 rack
Reasoning Model TrainingLong-context RL at full bf16
Frontier Multi-ModalImage + video + text + audio combined
Llama 405B InferenceFP8 weights (405 GB) fit across 2 GPUs (576 GB)
Full specs
NVIDIA GB300 specifications.
Memory
- Memory
- 288 GB
- Type
- HBM3e
- Bandwidth
- up to 8.0 TB/s
Compute
- FP32
- 83 TFLOPS
- FP16 tensor
- 2500 TFLOPS
- INT8
- 167 TOPS
- Tensor cores
- up to 640
- CUDA cores
- up to 20,480
Power and form
- TDP
- up to 1400 W
- Form factor
- Superchip
- Architecture
- Blackwell
- Released
- 2025
Node
- GPUs per node
- 8
- Interconnect
- NVLink
- Maker
- NVIDIA
Alternatives
Alternatives to the GB300. Same account, different price and performance.
| GPU | Memory · FP16 tensor | Stands out | Price gap | Price |
|---|---|---|---|---|
| NVIDIA GB300This GPU | 288 GB HBM3e · 2500 TFLOPS | $14.30/GPU-hr | ||
| NVIDIA GB200 | 186 GB HBM3e · 2500 TFLOPS | -102 GB memory | Not in stock right now | |
| NVIDIA B200 | 180 GB HBM3e · 2250 TFLOPS | -108 GB memory | $10.01/hr | |
| NVIDIA B300 | 288 GB HBM3e · 2250 TFLOPS | $17.02/hr | ||
| AMD MI300X | 192 GB HBM3 · 1307.4 TFLOPS | -96 GB memory | $4.85/hr |
Questions
Renting the GB300. What buyers ask first.
3 months. A GB300 NVL72 rack is rented for a 3-month term.
The 3-month term is prepaid in full.
As bare metal: you get the machines, with root access. Bare metal is the default for reserved clusters on Runcrate.
Handover within 72 hours after your payment clears, with the date confirmed in writing at booking.
No. The GB300 is rented by the rack: one GB300 NVL72 rack holds 72 GPUs.
Yes. Enter the number of racks you want in the request form on this page. A GB300 cluster is one or more GB300 NVL72 racks.
The B300 is the same Blackwell Ultra GPU, without the Grace CPU. On Runcrate the GB300 is rented by the GB300 NVL72 rack, 72 GPUs at a time.
Send the request form on this page with your name, work email, company, the number of racks and a start date. No account is needed.
GB300 is NVIDIA's Grace Blackwell Ultra superchip: one Grace CPU with two Blackwell Ultra GPUs, each with 288GB of HBM3e at 8 TB/s and 15 PFLOPS of dense FP4 compute (NVIDIA spec).
GB300 has about 55% more memory per GPU (288 vs 186GB) at the same 8 TB/s of bandwidth, and 1.5x the dense FP4 compute: meaningful for frontier-model pretraining and trillion-parameter inference. For models that fit in 186GB per GPU, GB200 is the better fit.
Other GPUs
Browse every GPU by architecture.
NVIDIA Blackwell
NVIDIA Hopper
NVIDIA Ampere
NVIDIA Ada Lovelace
NVIDIA Volta