How much does a B300 cost per hour? 29 providers, one table.
Every published on-demand and spot price we could find for the NVIDIA B300 (Blackwell Ultra), recorded 15 September 2026 with the source linked on each row, sorted from cheapest to most expensive, in US dollars per GPU-hour. We do not rent B300s ourselves: the comparison at the bottom is with the B200 we do run.
- Cheapest on-demand
- $3.83/hVerda, published rate
- Median on-demand
- $7.40/hAcross 29 providers
- Cheapest spot seen
- $2.99/hLyceum, interruptible
- B200 here
- $1.69/h180 GB · spot · on-demand $2.49
Between $3.83 and $17.80 an hour, median $7.40.
The spread is wide: the top of the table costs 5 times the bottom for the same chip. What changes between rows is the contract around it — node size, commitment, bundled networking and support.
- On-demand, cheapest
- $3.83 per GPU-hourVerda · $2,795.90 for a 730-hour month
- On-demand, median
- $7.40 per GPU-hourHalf of the 29 providers charge more · $5,402.00 a month
- Spot or interruptible
- $2.99 per GPU-hour at LyceumCapacity the provider can take back; notice periods differ
- An 8-GPU node
- $59.20 per hour at the median$30.64 at the cheapest published rate · 2.3 TB of GPU memory
B300 at a glance
Blackwell Ultra- GPU memory288 GB HBM3e
- Memory bandwidthup to 8 TB/s
- NVLink, per GPU1.8 TB/s
- Memory per 8-GPU node2.3 TB
- B200, for comparison180 GB · 7,700 GB/s
Figures from NVIDIA's HGX reference architecture. Same bandwidth as the B200: the B300 buys memory, not speed on a model that already fits.
B300, on-demand price per GPU-hour, 29 providers.
Lowest published rate we recorded for each provider, cheapest first. Spot is shown when the provider publishes one. The source is linked on every row; a provider with a comparison page here is linked to it.
| # | Provider | On-demandper GPU-hour | Spot | 730 h | Source |
|---|---|---|---|---|---|
| 1 | Verdarecorded 15 Sep 2026 | $3.83 | $3.79 | $2,795.90 | shadeform.com |
| 2 | Nebiusrecorded 15 Sep 2026 | $4.30 | $4.30 | $3,139.00 | nebius.com |
| 3 | Bentausrecorded 15 Sep 2026 | $4.35 | not published | $3,175.50 | getdeploying.com |
| 4 | CoreWeaverecorded 15 Sep 2026 | $4.48 | $4.48 | $3,270.40 | coreweave.com |
| 5 | TheAI Cloudrecorded 15 Sep 2026 | $4.90 | not published | $3,577.00 | computeprices.com |
| 6 | Prime Intellectrecorded 15 Sep 2026 | $4.99 | not published | $3,642.70 | primeintellect.ai |
| 7 | 1Legionrecorded 15 Sep 2026 | $5.00 | not published | $3,650.00 | computeprices.com |
| 8 | Civorecorded 15 Sep 2026 | $5.49 | not published | $4,007.70 | getdeploying.com |
| 9 | GPU Outletrecorded 15 Sep 2026 | $5.90 | not published | $4,307.00 | computeprices.com |
| 10 | GPU.airecorded 15 Sep 2026 | $6.33 | not published | $4,620.90 | getdeploying.com |
| 11 | Massed Computerecorded 15 Sep 2026 | $6.60 | not published | $4,818.00 | vm.massedcompute.com |
| 12 | Gcorerecorded 15 Sep 2026 | $6.93 | not published | $5,058.90 | getdeploying.com |
| 13 | RunPodrecorded 15 Sep 2026 | $6.94 | not published | $5,066.20 | runpod.io |
| 14 | Modalrecorded 15 Sep 2026 | $7.10 | not published | $5,182.27 | computeprices.com |
| 15 | Hyperstackrecorded 15 Sep 2026 | $7.40 | not published | $5,402.00 | hyperstack.cloud |
| 16 | Aquanoderecorded 15 Sep 2026 | $7.50 | not published | $5,475.00 | computeprices.com |
| 17 | Envergerecorded 15 Sep 2026 | $7.50 | not published | $5,475.00 | getdeploying.com |
| 18 | Lyceumrecorded 15 Sep 2026 | $7.99 | $2.99 | $5,832.70 | getdeploying.com |
| 19 | Latitude.shrecorded 15 Sep 2026 | $8.00 | not published | $5,840.00 | latitude.sh |
| 20 | Liumrecorded 15 Sep 2026 | $8.00 | not published | $5,840.00 | getdeploying.com |
| 21 | Sestercerecorded 15 Sep 2026 | $8.48 | not published | $6,190.40 | getdeploying.com |
| 22 | Fal.airecorded 15 Sep 2026 | $8.50 | not published | $6,205.00 | getdeploying.com |
| 23 | Scalewayrecorded 15 Sep 2026 | $8.66 | not published | $6,321.80 | shadeform.com |
| 24 | Vast.airecorded 15 Sep 2026 | $9.38 | not published | $6,847.40 | computeprices.com |
| 25 | Spheronrecorded 15 Sep 2026 | $10.21 | not published | $7,453.30 | computeprices.com |
| 26 | DigitalOceanrecorded 15 Sep 2026 | $11.19 | $11.19 | $8,168.70 | getdeploying.com |
| 27 | Oracle Cloudrecorded 15 Sep 2026 | $15.00 | not published | $10,950.00 | cloud-gpus.com |
| 28 | EmpirioLabs AIrecorded 15 Sep 2026 | $16.00 | not published | $11,680.00 | getdeploying.com |
| 29 | AWSrecorded 15 Sep 2026 | $17.80 | $5.59 | $12,994.00 | getdeploying.com |
Published prices move. If a figure is out of date, send us the link and the row is corrected. A provider without a published price is not in the table rather than estimated.
How many GPUs a large model needs on each, and what the node costs.
Weights plus a 32K-token context, nodes of 1, 2, 4 or 8 GPUs, with 10% of the memory kept free — the same arithmetic as our model sizing pages. B300 cost at the median published rate; B200 at our spot price.
| Model | Memory needed | B300s | B300 node, median | B200s | B200 node here, spot |
|---|---|---|---|---|---|
| DeepSeek V3.2FP8 (native) | 692 GB | 4× | $29.60/h | 8× | $13.52/h |
| GLM-5.3FP8 (native) | 759 GB | 4× | $29.60/h | 8× | $13.52/h |
| Kimi K2.6INT4 (native) | 598 GB | 4× | $29.60/h | 4× | $6.76/h |
| Qwen3.5 397B A17BBF16 | 808 GB | 4× | $29.60/h | 8× | $13.52/h |
| Llama 3.1 405BFP8 | 423 GB | 2× | $14.80/h | 4× | $6.76/h |
| gpt-oss-120bBF16 | 235 GB | 1× | $7.40/h | 2× | $3.38/h |
A B300 costs more per hour than a B200, but a model that needs eight B200s may need only four B300s. Where a row shows the same GPU count on both, the extra memory buys headroom for longer contexts or larger batches, not a smaller node.
B300 pricing, in detail
How much does a B300 cost per hour?
Published on-demand prices for the NVIDIA B300 run from $3.83 to $17.80 per GPU-hour across the 29 providers in this index (September 2026), with a median of $7.40. The lowest spot or interruptible price we found is $2.99 at Lyceum. A full month on one GPU at the median is $5,402.00.
What is the difference between the B300 and the B200?
The B300 (Blackwell Ultra) carries 288 GB of HBM3e per GPU against 180 GB on the B200, with the same 8 TB/s of memory bandwidth and fifth-generation NVLink at 1.8 TB/s per GPU, according to NVIDIA's HGX reference architecture. The extra memory is what you pay for: a large model fits on fewer GPUs, and a long context or a big batch leaves more room for the KV cache. Per-token speed on a model that already fits is governed by the same bandwidth.
Does OnlineGPU rent B300 GPUs?
Not yet. The closest GPU in our catalogue is the B200 SXM, at $1.69 per GPU-hour on spot and $2.49 on-demand, billed per minute. The table below shows when two B200s cost less than one B300 for the same model.
When is a B300 worth it over a B200?
When the model does not fit on the B200 node you would otherwise use. A model that needs between 1.3 TB and 2.07 TB of GPU memory (weights plus cache, with a working margin) fits on one 8-GPU B300 node but not on one B200 node. Below that, compare the two node prices for the GPU count each one needs: the table on this page does it for six large models.
Why do B300 prices vary so much?
Most published B300 rates are for capacity in 8-GPU HGX nodes or rack-scale systems, divided per GPU; some include networking, storage and support, some require a commitment, and a few are list prices for reserved terms. This table compares one thing: the lowest published price of one B300 for one hour, with the source on every row.
More guides: H100 price per hour · H100 vs H200 vs B200 · the B200 we run · GPU requirements of 64 open models
A B200 for $1.69 an hour, billed by the minute.
180 GB of HBM3e and the same 8 TB/s as a B300. Prepaid in crypto, no contract, stop whenever you want.
Billed per minute from the moment the instance is reachable. Minimum credit $30, no subscription.