Ask five providers what an NVIDIA B300 costs, and you’ll get five different answers. Across the market, on-demand B300 GPU pricing runs from about $4.70 to $18.00 per GPU-hour, a spread of nearly 4x for the same chip. That makes it one of the harder line items to budget for when you’re planning AI infrastructure.
This guide explains what’s behind the spread, how the B300 compares with the H100, B200, and GB200, where pricing is likely headed over the next 6 to 12 months, and how to work out what a B300 will cost your workload.
The short version
- On-demand B300 pricing ranges from roughly $4.70 to $18.00 per GPU-hour. Most mid-tier providers land between $7 and $9.
- Provider type and commitment length move the price more than any other factor. Hyperscalers sit at the top of the range, and neoclouds at the bottom.
- The B300 costs about $1 more per GPU-hour than the B200 at providers that list both, and it gives you 50% more memory.
- Expect a gradual easing over the next year rather than a sharp drop. Current pricing is likely close to the ceiling for this generation.
What the B300 is, and why buyers care
The B300 is NVIDIA’s Blackwell Ultra refresh of the B200, and it has been shipping since early 2026. The headline change is memory: 288GB of HBM3e per GPU, up from 192GB. That extra headroom matters for training on very large or long-context models and for memory-bound inference on 70B-plus-parameter models.
The trade-off is power. Each B300 draws up to 1,400W depending on configuration (up from 1,000W for the B200), and the highest-power configurations require direct liquid cooling. Facilities need to be built for it, which limits where the B300 can be deployed and shows up in providers’ charges.
What does a B300 cost right now?
Provider type | B300 price | Notes |
|---|---|---|
Neoclouds & marketplaces (specialist GPU providers) | ~$4.70–$9.20 per GPU-hour | Lowest on-demand rates |
Hyperscalers (AWS, Azure, Google Cloud) | ~$12–$18 per GPU-hour | Egress and other fees may add to cost |
Wholesale/committed-volume | ~$7.00–$12.00 per GPU-hour | Negotiated directly, not publicly listed |
Market median (on-demand) | ~$7.60–$7.90 per GPU-hour | Across 12+ providers |
Buying: DGX B300 (8 GPUs) | $300,000–$350,000 | ~$37,500–$44,000 per GPU, before facility, networking, and installation |
B300 pricing by provider type. Prices as of October 2026.
The median is the number to anchor on. If a quote is far above ~$7.90, find out what’s behind the premium: support, surrounding infrastructure, or just brand. If it’s far below, check the fine print on availability, interruptions, and region.
Why does the same chip have such a wide price range?
Five things do most of the work.
- Supply. Blackwell Ultra production is still ramping, and a meaningful share of early B300 capacity is locked into long-term enterprise contracts. That leaves less for on-demand rental.
- Demand. Training and inference for frontier and near-frontier models still outpace available capacity industry-wide, so pricing power stays with providers.
- Region. B300 capacity is concentrated in the few data center regions that can handle its power and liquid-cooling requirements. Price and availability can vary widely from one geography to the next.
- Deployment model. On-demand, reserved, and spot capacity carry very different prices for identical hardware (more on this below).
- Provider type. Hyperscalers generally charge a premium over neoclouds for the same chip. That premium reflects the surrounding infrastructure, support, and commitment terms instead of the GPU itself.
On-demand, spot, or reserved?
Deployment model | Typical B300 rate (per GPU-hour) | Best fit |
|---|---|---|
On-demand | ~$4.70–$18.00, no commitment | Short-term or variable workloads, evaluation, and testing |
Spot/interruptible | Below on-demand, but capacity can be reclaimed | Fault-tolerant training, batch jobs that can checkpoint and resume |
Wholesale/committed-volume | ~$7.00–$12.00, negotiated by volume | Sustained production workloads at predictable scale |
Long-term reserved (multi-year) | Can fall well below on-demand rates | Large, multi-year training or inference commitments with stable capacity needs |
Two trade-offs are worth knowing about before you commit to any of them.
Spot pricing can undercut on-demand by a wide margin. But if a training run can’t checkpoint cleanly, one reclaimed instance can cost you more than the discount saved.
Longer commitments bring the effective hourly rate down substantially. The price is flexibility: you can’t easily switch providers or scale down if your workload changes.
How does the B300 compare to the H100, B200, and GB200?
GPU | Architecture | Typical on-demand range (per GPU-hour) | Memory |
|---|---|---|---|
H100 | Hopper | ~$1.40–$12.30 (commonly $1.50–$4 on neoclouds) | 80GB HBM3 |
B200 | Blackwell | ~$2.70–$27.00 (commonly $6–$9 on mid-tier providers) | 192GB HBM3e |
B300 | Blackwell Ultra | ~$4.70–$18.00 (commonly $7–$9 on mid-tier providers) | 288GB HBM3e |
GB200 NVL72 | Blackwell (rack-scale) | ~$8.00–$27.00, billed per GPU across a 72-GPU rack | 192GB HBM3e per GPU |
Two comparisons matter most for procurement.
- B300 vs. B200. At providers that list both, the B300 typically runs about $1 more per GPU-hour. For 50% more memory, that’s a modest premium.
- B300 vs. H100. H100 pricing has dropped sharply since 2023 as Blackwell-generation hardware absorbed new demand. Analysts expect the B300 to follow that path eventually, once supply catches up and Rubin arrives.
The GB200 NVL72 is a different kind of purchase. Instead of being sold as individual instances, it’s sold as rack-scale capacity, so its per-GPU rate isn’t directly comparable to a standalone B300 rental unless you account for the full-rack commitment.
Where is B300 pricing headed over the next 6–12 months?
Early in a hardware generation’s life, demand outruns supply and prices stay firm. The B200 followed a similar pattern after launch, though less pronounced. Two things should ease that pressure from here:
- More manufacturing capacity. As Blackwell Ultra output ramps, more providers can list B300 on-demand, which should pull down the top of the range.
- Rubin. NVIDIA’s next-generation architecture is beginning to reach cloud providers in the second half of 2026. Newer generations tend to shift frontier training and inference demand toward the newest chips, which relieves pricing pressure on the previous generation.
We wouldn’t expect the B300 to fall as fast as the H100 did. Its memory advantage keeps it relevant for a narrower but still substantial set of memory-bound workloads, even after Rubin ships. If you’re buying now, plan for gradual compression rather than a cliff. Treat any specific forecast, including this one, as an informed estimate rather than a promise.
How to work out what a B300 will actually cost you
A single hourly rate rarely tells the whole story. Four questions help turn a price list into a decision:
- What’s your workload’s memory profile? If your model fits comfortably on an H200 or B200, the B300 premium may not be worth it. It earns its cost on memory-bound workloads and those that need the largest single-GPU memory pool.
- What utilization will you realistically hit? Divide the hourly rate by your utilization to get the cost per useful GPU-hour. As an illustration, $5.00/hr at 50% utilization works out to $10.00 per useful hour, while $7.50/hr at 90% utilization works out to about $8.33. The cheaper sticker price can be the more expensive option.
- How predictable is the next 6–12 months? Stable, sustained demand favors reserved or committed pricing. Variable or exploratory work is usually better on-demand.
- How many providers are you comparing? With a spread this wide, checking one or two clouds means you’re looking at an incomplete picture of the market.
Common budgeting mistakes
- Comparing only on-demand sticker prices, and ignoring egress and other fees.
- Signing a multi-year commitment before the workload has stabilized.
- Comparing a GB200 per-GPU rate directly with a single B300 rental.
- Treating the lowest listed price as the real price without confirming the region, availability, and interruption terms.
Wrapping up
B300 pricing isn’t a single number, and the gap between the cheapest and most expensive option is large enough to change a project’s budget. The most reliable way to navigate that gap is to compare the entire market rather than a single provider’s rate card.
That’s what Shadeform is built for. Teams get one API, one vendor, and one support team across 30+ vetted, SOC 2 Type II-compliant cloud providers in 100+ regions, at market rate with no markup.
See current B300 availability and pricing on Shadeform, deploy in as little as an hour, or reach out if you need B300 or GB300 capacity reserved.
