Blackwell Ultra · data center GPU
Rent a dedicated NVIDIA B300 server for $3,790 a month
The B300 is NVIDIA's Blackwell Ultra data center GPU: 288 GB of HBM3e on a single card, 60% more than the B200, at 8 TB/s. That is enough for a 70-billion-parameter model to run at 16-bit precision on one GPU with a 128k-token context to spare, and for a 400B-class model to fit on a single 8-GPU node. Here it is a dedicated bare-metal server with its own CPU cores, RAM and NVMe, at one flat monthly price.
- Per GPU, per month$3,790$5.19/h effective · fixed for your term
- Same GPU, hourly on-demand$5,760market median $7.89/h × 730 h
- You keep−34%$1,970 a month, every month
Specification
NVIDIA B300 server specifications
One dedicated machine, sized around the GPU. Nothing is shared, nothing is metered, and every line below is included in the monthly price.
Why teams rent the B300 here
- 288 GB HBM3e per GPU. The largest memory of any NVIDIA GPU: fewer GPUs per model, larger batches, longer contexts.
- Blackwell Ultra Transformer Engine. FP4 and FP8 inference with the accuracy management of the second-generation Transformer Engine.
- NVLink 5 in the 8-GPU node. 1.8 TB/s per GPU between the eight B300s of a node, and 3.2 Tb/s InfiniBand out of it.
- One flat invoice. $3,790 a month covers the GPU, the host, the NVMe, 20 TB of transfer and 24/7 support — the same number every month, fixed for your term.
Below about 480 hours of use a month, an hourly provider is the cheaper way to run this GPU. Above it — and a machine that trains, serves or renders is above it — the monthly rate wins, and the gap is the $1,970 shown above.
Pricing
B300 price per month, by term
Per GPU, in USD, excluding VAT. Longer terms cost less; every term includes the same dedicated host, network and support.
| Term | Per GPU, per month | Effective hourly | vs on-demand | Order |
|---|---|---|---|---|
| MonthlyRolling month-to-month. Cancel with 30 days' notice. | $3,790 | $5.19/h | −34% | Deploy |
| 3 months5% off the monthly rate for a 3-month term. | $3,601 | $4.93/h | −37% | Deploy |
| 6 months10% off the monthly rate for a 6-month term. | $3,411 | $4.67/h | −41% | Deploy |
| 12 months15% off the monthly rate for a 12-month term. | $3,222 | $4.41/h | −44% | Deploy |
| 8× B300 nodeNVLink between the GPUs · NVLink 5 · 3.2 Tb/s InfiniBand · 224 vCPU · 4 TB RAM · 30 TB NVMe | $27,900 | $3,488/GPU | −8% | Cluster details |
Launch pricing: these rates apply to orders confirmed before Oct 31, 2026 and stay fixed for your whole term. From Nov 1, new B300 orders are billed at the list price of $4,359 a month. On-demand comparison: market-median published hourly rate for the same GPU ($7.89/h) over 730 hours. Block storage $25/TB/month and extra IPv4 addresses $4/month are optional add-ons.
Market comparison
B300 rental price compared
Every provider in our audit that publishes an on-demand rate for the B300, read on 19 September 2026 from their own pricing page and multiplied by 730 hours — what the same GPU costs kept for a month.
| Provider | Published rate, B300 | One month, 24/7 | vs our $3,790/mo |
|---|---|---|---|
 GPU Cloud HQDedicated, billed monthly |
$3,790 per month, flat | $3,790 | Our reference |
 Verda (DataCrunch)On-demand, hourly |
$7.88/h on demand · $3.94 spot | $5,752 | −34% |
 NebiusOn-demand, hourly |
$7.85/h on demand · $4.30 preemptible | $5,731 | −34% |
| Market median30–40 providers, GetDeploying index | $7.89/h on demand | $5,760 | −34% |
Rates are each provider's standard on-demand tier for the B300 — not spot, preemptible or community hardware — as printed on their pricing page on 19 September 2026; each row links to the full comparison with its source. Where a provider is cheaper than us for a full month, the row says so. Method and caveats: GPU cloud alternatives.
Workloads
What people run on a B300
Frontier training & long-context inference — and the three jobs below are where a dedicated B300 at a flat monthly price earns its keep.
-
Long-context inference
Serve 70B-class models at 16-bit, or 400B-class models at FP8 on one node, with the KV cache for 128k-token prompts kept entirely in HBM.
-
Frontier fine-tuning and post-training
Full-parameter fine-tunes, RLHF and DPO on models that no longer fit an H100 node — without sharding the optimizer state across racks.
-
Mixture-of-experts serving
Keep every expert of a large MoE resident in memory instead of paging them, and stop paying for the idle GPUs a smaller card forces on you.
Is the B300 the right GPU for you?
Choose the B300 when memory per GPU is the constraint: models above 70B parameters, contexts above 32k tokens, or mixture-of-experts architectures whose experts must stay resident. If your model fits in 180 GB, the B200 delivers the same Blackwell generation for less; if it fits in 80 GB, an H100 SXM costs a third as much.
B300 rental: questions answered
How much does it cost to rent a B300 server?
$3,790 a month, flat, for a dedicated NVIDIA B300 server with 32 vCPU · 512 GB RAM · 4 TB NVMe and a 25 Gbps uplink — $5.19 per hour effective over 730 hours. A 3-, 6- or 12-month term takes 5%, 10% or 15% off: $3,601, $3,411 or $3,222 a month. The market-median on-demand rate for the same GPU is $7.89 per hour, about $5,760 for a full month.
Is the B300 in stock right now?
Yes. 5 of the current batch of 8 were unallocated at the last stock update (18 September 2026). Provisioning is automated: your server is imaged, secured with your SSH key and handed over in under 10 minutes after your payment confirms. The next batch is due Oct 20.
What is included with a B300 server?
Everything on the pricing table: the GPU, 32 vCPU · 512 GB RAM · 4 TB NVMe, a 25 Gbps uplink with 20 TB of outbound transfer, a dedicated public IPv4 and a /64 IPv6 block, DDoS protection, root access and Ubuntu 24.04 with the NVIDIA driver, CUDA 12 and Docker — or PyTorch, vLLM, Kubernetes images at checkout. Hardware replacement within 4 hours and a 99.9% uptime SLA are part of the contract.
How is the B300 different from the B200?
Both are Blackwell-generation data center GPUs with 8 TB/s of HBM3e bandwidth. The B300 (Blackwell Ultra) carries 288 GB of memory against 180 GB on the B200 and steps up FP4 inference throughput. If your models fit in 180 GB, the B200 is the better buy; the B300 is for the models, contexts and expert mixtures that do not.
Can I rent several B300s or a full 8-GPU node?
Yes: order 1, 2, 4 or 8 GPUs at once, subject to the units left in the batch, on the same monthly terms. The 8× B300 node — NVLink between the GPUs, 3.2 Tb/s InfiniBand, NVLink 5 · 3.2 Tb/s InfiniBand · 224 vCPU · 4 TB RAM · 30 TB NVMe — is priced separately at $27,900 a month on the clusters page.
Do I need KYC or a card to rent it?
No. An email address is all we ask for. You fund a prepaid balance in Bitcoin, Ether or USDT (ERC-20 or TRC-20); the first invoice settles from it and monthly renewals charge to it automatically. No ID document, no card, no bank transfer.
Other data center GPUs by the month
Same dedicated hosts, same monthly model, same 10-minute provisioning. All 10 GPUs are on the pricing page.
Deploy a B300 in under 10 minutes
$3,790 a month, fixed for your term. Fund your balance in BTC, ETH or USDT and the server is imaged, secured with your key and handed over automatically.