The NVIDIA B300 is the Blackwell Ultra data center GPU: NVIDIA's datasheet lists 270 GB of HBM3e per GPU on the HGX B300 board, and its HGX page lists 2.1 TB for the board's 8 GPUs. DeployGPU rents the B300 by the hour as a VM whenever one is in stock, which is only part of the time, so this page shows the live table and the way to get a B300 GPU when the table is empty.
No B300 VM is available to launch right now (checked 2026-11-13 06:40 UTC). Check the marketplace for new stock.
B300 VMs you can launch now
The table lists every B300 VM with 8 GPUs that is in stock, cheapest first, and refreshes every 15 minutes. B300 stock appears for a while and then goes, so an empty table is not an error. When it is empty, Auto Provision lets you place an order that waits for a machine.
DeployGPU only lists providers it can launch machines on. For B300 prices across the whole market, compare all providers on GpuPerHour. For a reference price over time, see the B300 rental index (GPUIndexes).
The NVIDIA B300 against the B200 and H200
Blackwell Ultra is the name NVIDIA gives the B300's architecture, and its datasheet describes the memory as 1.5 times larger than its predecessor's. In a B300 vs B200 comparison, the main changes are memory, 270 GB against 180 GB per GPU, dense FP4, 108 against 72 PFLOPS per 8-GPU board, and power. The FP4 and FP8 rows are per 8-GPU board and the rest are per GPU.
| Spec | B300 (HGX B300) | B200 (HGX B200) | H200 SXM |
|---|---|---|---|
| Architecture | Blackwell Ultra | Blackwell | Hopper |
| Memory per GPU | 270 GB HBM3e | 180 GB HBM3e | 141 GB HBM3e |
| Memory in 8 GPUs | 2.1 TB | 1.4 TB | 1.1 TB |
| Memory bandwidth per GPU | 7.7 TB/s | 7.7 TB/s | 4.8 TB/s |
| Tensor Core formats listed | FP4, FP8/FP6, INT8, FP16/BF16, TF32, FP64 | FP4, FP8/FP6, INT8, FP16/BF16, TF32, FP64 | FP8, INT8, FP16/BF16, TF32, FP64 |
| FP4 per board, sparse and dense | 144 and 108 PFLOPS | 144 and 72 PFLOPS | Not listed |
| FP8 per board, with sparsity | 72 PFLOPS | 72 PFLOPS | 32 PFLOPS |
| NVLink per GPU | 1.8 TB/s, fifth generation | 1.8 TB/s, fifth generation | 900 GB/s, fourth generation |
| Power per GPU | Up to 1,100 W | Up to 1,000 W | Up to 700 W |
Specifications from NVIDIA's HGX page, Blackwell Ultra datasheet, Blackwell datasheet, HGX H100 and H200 datasheet and H200 page, checked 12 November 2026. The Blackwell Ultra datasheet gives 270 GB per GPU while the HGX page gives 2.1 TB for 8 GPUs, and this site uses the 2.1 TB, 2,100 GB, for every 8x B300 memory calculation. The GPU column in the live table repeats the offers feed's per-GPU figure, which can differ from NVIDIA's 270 GB. For B200 bandwidth, the Blackwell datasheet gives 7.7 TB/s per GPU and the DGX B200 page gives 64 TB/s across 8 GPUs, which is 8 TB/s each. The HGX page also lists 3 POPS of INT8 and 10 TFLOPS of FP64 for the B300 board against 72 POPS and 296 TFLOPS for the B200 board, so read the datasheet before you pick a B300 for INT8 or double-precision work.
Which workloads need a B300
A B300 earns its place when a checkpoint is too big for every other DeployGPU VM. By the 92% rule used across this site, an 8x B200 VM has 1,324.8 GB of usable memory and an 8x B300 VM has 1,932 GB, both calculated, and Kimi K3's checkpoint sits between them.
| Checkpoint | Size | 8x H200 (1,128 GB) | 8x B200 (1,440 GB) | 8x B300 (2,100 GB) |
|---|---|---|---|---|
| Kimi K3, MXFP4 | 1,560.94 GB | Does not fit | Does not fit | Fits, about 371 GB left |
| Kimi K3.1-Pro, MXFP4 | 1,561.29 GB | Does not fit | Does not fit | Fits, about 371 GB left |
| DeepSeek V4-Pro-0813 | 892.74 GB | Fits, about 145 GB left | Fits, about 432 GB left | Fits, about 1,039 GB left |
Sizes are the summed safetensors files on the Hugging Face cards for Kimi K3, Kimi K3.1-Pro and DeepSeek V4-Pro-0813, checked 12 November 2026, and the space left is what cache and runtime share. DeepSeek V4-Pro shows the second reason to pick a B300: the same weights leave about 1,039 GB for cache instead of 432 GB. The Kimi K3 page has the serving commands and the frontier models page compares the other open models that need a full 8-GPU VM. When your weights and cache fit in less, the B200 page and the H200 page list those VMs.
Get a B300 when none is in stock
Auto Provision is a dashboard page for GPUs that sell out. You pick the B300, the providers you accept from Verda, Hyperstack and QuantaCloud, and a maximum price per hour for the whole machine, fees included, and DeployGPU deploys when a machine appears at or below that price. Placing an order and waiting cost nothing, but your balance must cover one hour at the machine's price when it appears, or that machine is skipped and the order keeps waiting. It deploys a plain GPU VM, with no app template and no volume, and it works from the dashboard only, not the API.
Open Auto Provision in the dashboard to see the 30-day price range and how often a machine has come into stock before you place an order. Stock is not guaranteed, because an order deploys only if a matching machine appears while it is open.
What a B300 costs
NVIDIA's HGX, DGX B300 and GB300 NVL72 pages list no price and send buyers to partners or sales, checked 12 November 2026, so there is no NVIDIA B300 price to quote for buying one. The rental price is the Per hour column in the live table whenever a VM is in stock, including any platform fee. For an Auto Provision order you set the most you will pay per hour.
Software for Blackwell
PyTorch's 2.7 release notes say Blackwell support arrived with prebuilt CUDA 12.8 wheels, so use a build for CUDA 12.8 or newer. NVIDIA's CUDA GPU list puts the B300 at compute capability 10.3, and PyTorch's release table names the Blackwell architectures 10.0 and 12.0 for release 2.15. Run the matmul check in the PyTorch install guide before you trust a wheel.
One VM is 8 GPUs at most
B300 VMs sit on SXM boards, and NVIDIA lists fifth-generation NVLink at 1.8 TB/s per GPU for the HGX B300. One VM is the largest unit DeployGPU launches, with no multi-node cluster and no self-serve InfiniBand, and the multi-GPU page explains how GPUs inside one node talk. For more than 8 B300s, a fixed term or reserved capacity, request a cluster quote and a person replies by email. The highest VRAM GPUs page ranks the B300 against every other card.
How long until it is ready
A VM usually reaches Running in 1 to 3 minutes, and the weights take longer. Disks are fixed per offer on most providers, so check the disk size in the vCPU / RAM / Disk column before you download 1,560.94 GB. Only a launch that reaches Running is charged, and a failed launch refunds the reserved hour automatically. Launching reserves the first hour from your credit balance, the unused seconds go back when you stop, and credits are bought by card with a 5% + $0.30 processing fee on each top-up. Stopping deletes the disk, so copy results off first, and the pricing page has the full breakdown.
Frequently asked questions
How much memory does the NVIDIA B300 have?
NVIDIA's Blackwell Ultra datasheet lists 270 GB of HBM3e per GPU on the HGX B300 board, and NVIDIA's HGX page lists 2.1 TB for the 8 GPUs. This site uses the 2.1 TB figure when it checks whether a model fits an 8x B300 VM.
How do the B300 and B200 specs differ?
The B300 is Blackwell Ultra with 270 GB of HBM3e per GPU against 180 GB for the B200, and 108 PFLOPS of dense FP4 per 8-GPU board against 72 PFLOPS. Both list 1.8 TB/s of NVLink per GPU, and the B300 board lists much lower INT8 and FP64 figures.
Why is there no B300 in the table?
B300 stock appears for a while and then goes, so an empty table is not an error. An Auto Provision order deploys a machine when one appears at or below the maximum price you set.
How much does a B300 cost per hour?
The live table shows the hourly price of every B300 VM in stock, including any platform fee. NVIDIA's B300 pages list no purchase price, and an Auto Provision order lets you set the most you will pay per hour for the whole machine.
Which models need a B300?
Kimi K3 and Kimi K3.1-Pro have checkpoints of 1,560.94 GB and 1,561.29 GB, and an 8x B300 VM is the only VM DeployGPU rents that holds them by the site's 92% rule. By the same rule an 8x B200 VM has about 1,325 GB of usable memory.
Does PyTorch run on a B300?
PyTorch added Blackwell support in version 2.7 with CUDA 12.8 wheels, and NVIDIA lists the B300 as compute capability 10.3. PyTorch's release table names Blackwell 10.0 and 12.0 and does not name 10.3, so run the matmul check on the VM before a long job.