Nebius Pricing 2026: H100 to B300 Rates Rise Up to 21%

H100 SXM5 climbs from $3.85/hr to $4.50/hr, and B300 goes from $7.85/hr to $9.50/hr, the steepest jump in the new rate card. If you're running production workloads on Nebius today, here's exactly what changes in Nebius pricing GPU by GPU, why it's the second hike this year, and what it actually costs to keep running there versus moving.

For the fuller platform comparison behind these numbers, see our Spheron vs Nebius breakdown, and if CoreWeave is also on your shortlist, CoreWeave vs Nebius H100 and H200 pricing puts both rate cards side by side.

TL;DR: Nebius Pricing 2026, H100 to B300 Rates Rise Up to 21%

• Price hike: Nebius raises on-demand GPU rates up to 21% effective October 1, 2026: H100 SXM5 goes from $3.85/hr to $4.50/hr and B300 from $7.85/hr to $9.50/hr, confirmed as of 21 Sep 2026.
• H200 and B200: H200 SXM5 rises 20% to $5.40/hr and B200 rises 19% to $8.50/hr on the same date.
• Second hike: This is Nebius's second GPU rate-card increase in three months, stacking on a mid-2026 adjustment already baked into the pre-October rates above.
• Market reaction: Nebius shares jumped 11.1% in pre-market trading, from $209.37 to $232.50, after the price-increase news broke.
• Alternative: Spheron lists H100 on-demand at $2.64/hr as of 23 Sep 2026, with no lock-in and per-minute billing. Compare live GPU pricing.

The New Nebius Rate Card: Before and After, GPU by GPU

Nebius's own pricing page, checked 21 September 2026 (pre-hike), still listed the old rates: H100 at $3.85/hr, H200 at $4.50/hr, and B200 at $7.15/hr, matching the figures reported ahead of the October change exactly. B300 on-demand showed "contact required" with no listed self-serve rate, which is consistent with Blackwell Ultra capacity being tight enough that Nebius isn't publishing a checkout price for it yet.

H100 SXM5: $3.85 to $4.50/hr (+17%)

H100 is still Nebius's most heavily used SKU, and the increase there is the smallest percentage of the four, though it's the GPU most teams have the biggest running fleets on, so it's also where the absolute monthly dollar impact adds up fastest at scale. For the wider H100 buy-vs-rent picture across providers, see our NVIDIA H100 price guide.

H200 SXM5: $4.50 to $5.40/hr (+20%)

H200's 20% increase is the second-largest in the round. Teams running memory-bound inference on H200's 141GB HBM3e for larger context windows are the ones who'll feel this the most, since H200 workloads tend to run longer and more continuously than short H100 fine-tuning jobs.

B200: $7.15 to $8.50/hr (+19%)

B200 moves from $7.15/hr to $8.50/hr. For the fuller cross-provider B200 rate picture, including how this new Nebius number lines up against CoreWeave and AWS, see our NVIDIA B200 cloud pricing guide.

B300: $7.85 to $9.50/hr (+21%)

B300 gets the steepest increase of the four GPUs Nebius repriced, rising from $7.85/hr to $9.50/hr, about a 21% jump. It's also the newest hardware in the lineup, and the one where Nebius's own pricing page doesn't publish a self-serve on-demand rate at all, which tells you where the actual supply pressure is concentrated.

Why This Is Nebius's Second Hike in Months

The Mid-2026 Increase and the Cumulative Move on B300

The October 1 change isn't Nebius's first adjustment of the year. Coverage of the announcement frames it as the second GPU rate-card increase in roughly three months, meaning the $7.85/hr pre-October B300 rate already reflected an earlier move before this latest 21% jump stacked on top of it. Two increases landing within a single quarter is the detail worth budgeting around: a rate card you priced out over the summer is not the rate card you're paying today, and it won't necessarily be the one you're paying at the end of the year either.

What's Driving It: Tight Blackwell Supply and Price Testing

Nebius had reportedly already been running price testing for scarce Blackwell capacity ahead of the formal rate-card change, achieving transaction prices roughly 15% higher than its previous peak. That's the market signal underneath the announced numbers: demand for B200 and B300 capacity has been outrunning what Nebius can allocate, and the company tested how much of that gap the market would actually pay before writing it into the published rate card.

The reaction from investors reinforces that reading. Nebius shares jumped 11.1% in pre-market trading, from $209.37 to $232.50, following the price-increase news. A GPU cloud raising prices and having its stock jump on the news is a specific signal: investors are reading it as pricing power in a supply-constrained market, not as a company passing along costs it can't otherwise cover. Nebius has also been raising its own infrastructure capex guidance more than once this year to build out that Blackwell capacity, which points the same direction: money is going into more supply, but not fast enough to keep the rate card flat in the meantime.

GPU Cloud Pricing Comparison: Where Nebius's New Rates Sit Against CoreWeave, AWS, and Spheron

Even after the October increase, Nebius's per-GPU rate is still below the two providers that sell the same silicon in fixed multi-GPU bundles rather than by the single GPU.

CoreWeave's published rate card prices the 8-GPU HGX H100 node at $49.24/hr ($6.16/GPU-hr), the HGX H200 node at $50.44/hr ($6.31/GPU-hr), and the HGX B200 node at $68.80/hr ($8.60/GPU-hr), all sold as full nodes with no single-GPU checkout option. AWS's EC2 Capacity Blocks pricing page lists P6-B200 at $12.355 per accelerator-hour and P6-B300 at $14.04, a fixed-window reservation product rather than a flexible on-demand rate; our AWS Capacity Blocks pricing breakdown covers that hike in full. For the wider cross-provider table across 5+ providers and every mainstream GPU tier, see our GPU cloud pricing comparison.

The structural point is what changes here, not just the numbers. Nebius, CoreWeave, and AWS each set one rate card and move it when they choose to. Spheron runs as a marketplace where $2.64/hr and every other live rate reflects multiple providers competing for the same workload on the backend, so a single vendor's capex or supply decision doesn't move the whole rate card overnight. It's a genuinely different pricing mechanism, not just a different number.

Pricing fluctuates based on GPU availability. Spheron rates above are live as of 23 Sep 2026; Nebius, CoreWeave, and AWS figures reflect their most recently published rate cards and may have changed.

What It Costs to Move H100/B300 Workloads Off Nebius

Worked Example: 8x H100 and 8x B300 Node, Before vs After vs Spheron

Take a team running an 8-GPU H100 node continuously for a 30-day month (720 hours) on Nebius, before and after the October 1 change:

• Nebius before (pre-October, $3.85/hr): $3.85 x 8 x 720 = $22,176/month
• Nebius after (post-October, $4.50/hr): $4.50 x 8 x 720 = $25,920/month, a $3,744/month increase for the same hardware and the same workload
• Spheron on-demand ($2.65/hr as of 21 Sep 2026): $2.65 x 8 x 720 = $15,264/month

Run the same math on an 8-GPU B300 node:

• Nebius before (pre-October, $7.85/hr): $7.85 x 8 x 720 = $45,216/month
• Nebius after (post-October, $9.50/hr): $9.50 x 8 x 720 = $54,720/month, a $9,504/month increase, the largest jump of any GPU in this comparison
• Spheron spot ($5.85/hr as of 21 Sep 2026, on-demand not listed): $5.85 x 8 x 720 = $33,696/month

Those Spheron figures are frozen to the rate at the time of writing and will move with marketplace supply, the same way Nebius's own numbers just did. For a full shortlist of alternatives beyond Spheron, 10 Best Nebius Alternatives covers the rest of the field.

What Actually Locks You In (Committed-Use Terms) vs What Doesn't

The one thing that could soften this hike is a committed-use discount, and it's worth knowing exactly what that trades away. Nebius advertises "up to 35% less than on-demand rates by reserving large-scale clusters for multiple months," which would bring the new $4.50/hr H100 rate down closer to $2.93/hr, but only for teams that can commit to a multi-month reservation and go through a sales process to get an actual number.

Spheron's marketplace works the other way: on-demand and spot rates are live with no contractual lock-in, billed per minute after a 20-minute minimum, and you can spin down one GPU instance and spin up a different one at any time with no penalty. That's the real trade-off between the two models. Nebius's committed-use discount is a real number, but it's a bet that your GPU needs won't change for months at a time. If they might, the value of not being locked in is worth pricing into the comparison, not just the headline hourly rate.

Where Nebius still has a genuine edge that's worth being direct about: it owns its EU infrastructure and has a stronger EU data-residency compliance story than a multi-provider marketplace built on aggregated third-party capacity. And Spheron doesn't currently list B300 on-demand at all, only spot, so a team specifically moving an on-demand B300 workload should verify current availability before assuming a like-for-like swap. If neither of those caveats applies to your workload, the per-minute, no-lock-in model is the more direct answer to a rate card that just moved twice in three months. Deployment steps for standing up GPU instances outside Nebius's console are in the Spheron docs.

If Nebius's October rate card just changed your budget, Spheron runs H100, H200, and B200 on-demand with no reservation and per-minute billing, plus B300 on spot, so moving a workload doesn't mean waiting out a contract.

Rent H100 on Spheron → | Spheron B300 instances → | Get started on Spheron →