B300 pricing ranges from $6.94 to $16.146 per GPU per hour right now, across five providers quoting the exact same NVIDIA chip: Blackwell Ultra, 288GB of HBM3e, the specs we broke down in full in our B300 guide. That roughly 2.3x spread isn't noise in the market, and it isn't one vendor gouging another's customers. It tracks how much commitment each number actually buys, from a prepaid block you can't cancel to a marketplace listing you can walk away from this afternoon. This post lines up B300 pricing across AWS, Oracle, Nebius, Verda, and Runpod exactly as each one publishes it, on-demand and reserved, and shows where Spheron's own live rate lands against all five.
TL;DR: What an HGX B300 Node Costs Right Now, by Provider
B300 on-demand pricing runs from $6.94/GPU-hr on Runpod to $15.00/GPU-hr on Oracle as of 11 Nov 2026, and AWS has no true on-demand rate at all, only a prepaid Capacity Block.
| Metric (as of 11 Nov 2026) | Figure |
|---|---|
| Cheapest on-demand | Runpod, $6.94/GPU-hr on Community Cloud |
| Most expensive on-demand | Oracle, $15.00/GPU-hr |
| No true on-demand rate | AWS, $16.146/accelerator-hour, prepaid Capacity Block only |
| Cheapest spot/preemptible | Nebius, $0.99/GPU-hr |
| Spheron on-demand | $12.36/hr, no contract required, as of 13 Nov 2026 |
On-Demand B300 Pricing: AWS, Oracle, Nebius, Verda, Runpod Compared
Five providers publish a B300 rate right now, but only four of them are selling what most buyers mean by "on-demand": a rate you can start and stop by the hour with no prepay. Here's what each one's rate actually buys, cheapest to most expensive per GPU:
| Provider | On-Demand Rate | What You're Actually Buying |
|---|---|---|
| Runpod | $6.94/GPU-hr (Community Cloud) or $8.99/GPU-hr (Secure Cloud) | Marketplace listing, billed per second, no minimum |
| Nebius | $9.50/GPU-hr | HGX B300 node, pay-as-you-go, no prepay |
| Verda | $11.93/GPU-hr | SXM6 node, on-demand, no commitment |
| Spheron | $12.36/hr | Aggregated on-demand across 5+ providers, 20-minute minimum |
| Oracle | $15.00/GPU-hr ($120.00/hr for the 8-GPU BM.GPU.B300.8 shape) | Bare-metal on-demand, pay-as-you-go |
| AWS | $16.146/accelerator-hour ($129.168/hr for the 8-GPU instance) | Prepaid, non-cancelable Capacity Block; no standalone catalog on-demand rate |
On a straight per-GPU, on-demand basis, Runpod's Community Cloud figure is the cheapest number in this table.
Runpod splits its B300 listings into two tiers, and both price per GPU per hour with billing per second and no minimum runtime: $6.94/GPU-hr on Community Cloud, where capacity comes from vetted third-party hosts, and $8.99/GPU-hr on Secure Cloud, Runpod's own data centers. Nebius's own pricing page lists HGX B300 on-demand at $9.50/GPU-hr. Verda's rate card lists B300 SXM6 at $11.93/hr on-demand per GPU, with a confidential-computing variant running slightly higher at $12.17/hr; for how Verda's rates have moved on its other GPUs, see our Verda GPU pricing breakdown. For how Nebius compares to Spheron beyond just B300, see our Spheron vs Nebius comparison.
Oracle's BM.GPU.B300.8 bare-metal shape lists at $120.00/hr for the full 8-GPU node, or $15.00/GPU-hr, in its us-ashburn-1 region. That's a genuine pay-as-you-go rate with no prepay required, the same pattern Oracle runs across its catalog; see our Oracle Cloud GPU pricing guide for how that holds for H100 and H200.
AWS doesn't have a line item here in the normal sense. AWS's own Capacity Blocks pricing page lists P6-B300 at $16.146 per accelerator-hour in US West (Oregon), US East (N. Virginia), the Atlanta Local Zone, Seoul, Hyderabad, Jakarta, and London, which works out to $129.168/hr for the full 8-GPU p6-b300.48xlarge instance. There is no separate, standalone on-demand p6-b300.48xlarge catalog rate you can book by the hour outside a Capacity Block. Every site that quotes an AWS B300 "on-demand" number alongside the other four here is filling that gap with an estimate AWS itself doesn't publish.
Reserved and Spot Discounts Against the On-Demand Baseline
Two of the five providers discount their B300 rate for a term commitment, two more discount it for reclaimable spot or preemptible capacity, and AWS's number is already a commitment with no discount tier underneath it.
- Oracle reserved: a 1-year commitment runs $80.40/hr ($10.05/GPU-hr), about a third off on-demand; a 3-year commitment runs $60.00/hr ($7.50/GPU-hr), half off on-demand.
- Nebius spot: preemptible capacity starts from $0.99/GPU-hr, well below its $9.50/GPU-hr on-demand rate.
- Verda spot: $5.965/GPU-hr, roughly half its $11.93/GPU-hr on-demand rate (the confidential-computing variant runs $6.085/hr spot against $12.17/hr on-demand).
- Spheron spot: $5.84/hr, against $12.36/hr on-demand.
- AWS: no spot tier. The Capacity Block rate already is the reservation, paid for a fixed window, and it keeps climbing: Spheron's own tracking of AWS Capacity Block pricing found P6-B300 at $14.04/accelerator-hour in July 2026, up roughly 20% from an earlier January hike of about 15%. The $16.146 rate above is a further increase on top of that July figure.
None of the smaller spot and preemptible numbers are free money. Spot capacity can be reclaimed by the provider without notice, and that isn't a theoretical caveat: in our own spot reliability testing in October 2026, every 8-GPU spot type we used was reclaimed at least once on 14 Oct 2026. If you're running a job that can't tolerate a mid-run interruption, price the on-demand or reserved row instead, or build in automatic recovery before you commit to the discount.
Pricing fluctuates based on GPU availability. Spheron rates above are live as of 13 Nov 2026; other providers reflect their most recent published rates and may have changed. Check current GPU pricing → for live rates.
Why B300 Pricing Spreads So Wide: Allocation Depth and Contract Terms
The gap between Runpod's $6.94/GPU-hr and AWS's $16.146/accelerator-hour isn't five providers pricing the same thing differently. It's five different products that happen to run on the same chip, and the price reflects exactly what each one commits to.
AWS's Capacity Block is a guaranteed, prepaid, non-cancelable allocation, bought for a fixed future window, usually weeks ahead of when you need it. That's built for a training run that needs certainty on a specific date, not a job you can shift if capacity is tight next week, so it carries no spot option and the highest rate on this list, one that has risen through three separate price moves in 2026.
Oracle's bare-metal shape sits a step down: you still get a dedicated physical machine with no shared tenancy, but no prepay is required to get the on-demand rate. The deep discount only shows up if you're willing to commit years ahead, which is why its 3-year rate ($7.50/GPU-hr) undercuts its own on-demand rate by half.
Nebius and Verda are cloud-native GPU providers competing directly on a published, self-serve on-demand rate, with genuinely flexible preemptible or spot scheduling underneath it because their infrastructure is built for that kind of workload churn. Runpod is a literal marketplace: Community Cloud nodes are owned and operated by third-party hosts, which is cheaper and carries a weaker uptime guarantee than Secure Cloud, Runpod's own data centers.
Spheron sits in the middle of that spread for a structural reason, not a pricing choice: it aggregates on-demand and spot B300 capacity across 5+ providers rather than running one data center's inventory, so its quoted rate blends several of these allocation depths into one number instead of representing just one of them.
Where Spheron Prices In: Renting B300 Today
Spheron's own live B300 rate sits almost exactly in the middle of this table: higher than Nebius and Runpod's cheapest tier, lower than Oracle's on-demand rate and well below AWS's prepaid Capacity Block. As of 13 Nov 2026, Spheron lists B300 (Blackwell Ultra, 288GB) at $12.36/hr on-demand and $5.84/hr spot per GPU, billed per minute after a 20-minute minimum, with no contract required.
That positioning won't suit every buyer. Teams chasing the single lowest on-demand number should price Nebius's $9.50/GPU-hr or Runpod's $6.94/GPU-hr Community Cloud rate first; Spheron's on-demand rate sits above both. Spheron also doesn't publish an Oracle-style multi-year committed-use schedule, so a team with a genuine multi-year B300 forecast and the balance sheet to prepay should run Oracle's 1-year and 3-year numbers against Spheron's rate before signing anything.
Where it does win against the other four: per-minute billing after a short minimum instead of AWS's and Oracle's per-hour terms, no Capacity-Block-style prepay to unlock a guaranteed on-demand rate, and an on-demand number that beats both AWS's Capacity Block and Oracle's catalog rate while its spot tier still sits above Nebius's and Verda's published floors.
If AWS's prepaid Capacity Block or Oracle's multi-year commitment doesn't fit how your team buys compute, Spheron's B300 rate is live and bookable by the minute with no contract.
Frequently Asked Questions
It depends on the provider and what kind of commitment you're buying. Oracle's BM.GPU.B300.8 bare-metal shape runs $15.00/GPU-hr on-demand ($120.00/hr for the 8-GPU node) in us-ashburn-1. Runpod lists B300 at $6.94/GPU-hr on Community Cloud and $8.99/GPU-hr on Secure Cloud, both on-demand. Nebius and Verda publish $9.50/GPU-hr and $11.93/GPU-hr on-demand. Spheron's own live rate is $12.36/hr on-demand ($5.84/hr spot) as of 13 Nov 2026, aggregated across 5+ providers.
Both exist, but AWS only sells it one way. AWS has no standalone on-demand p6-b300.48xlarge catalog rate: the only way to get B300 on AWS is a prepaid, non-cancelable EC2 Capacity Block at $16.146 per accelerator-hour ($129.168/hr for the 8-GPU instance). Oracle, Nebius, Verda, Runpod, and Spheron all sell true on-demand B300 by the hour or minute with no prepay. Oracle also publishes a discounted multi-year reserved tier, and Nebius, Verda, and Spheron each publish a discounted spot or preemptible tier on top of their on-demand rate.
On a straight on-demand, per-GPU basis, Runpod's Community Cloud rate of $6.94/GPU-hr is the lowest published figure, with Nebius at $9.50/GPU-hr next. The cheapest number on a rate card isn't automatically the cheapest real deployment, though: Community Cloud and preemptible tiers carry weaker uptime guarantees than a bare-metal reservation like Oracle's, so the right comparison depends on how much interruption risk your workload can absorb.
On-demand guarantees the GPU for as long as you pay for it; spot or preemptible capacity is cheaper but can be reclaimed by the provider without notice. Nebius's preemptible B300 capacity starts from $0.99/GPU-hr against $9.50/GPU-hr on-demand, and Verda's spot rate is $5.965/GPU-hr against $11.93/GPU-hr on-demand. Spot reclaim risk is operational, not theoretical: in our own spot reliability testing in October 2026, every 8-GPU spot type we used was reclaimed at least once on 14 Oct 2026.
Yes. Spheron lists NVIDIA B300 (Blackwell Ultra, 288GB) at $12.36/hr on-demand and $5.84/hr spot per GPU as of 13 Nov 2026, billed per minute after a 20-minute minimum, with no contract required and capacity aggregated across 5+ providers rather than a single data center's inventory.
