Nebius is a well-funded neocloud with serious GPU infrastructure in Europe. Their H100 on-demand runs $3.85/hr and H200 at $4.50/hr, with preemptible instances available at $2.15/hr and $2.45/hr respectively. The structure matters: per-hour billing, a quota process for large allocations, and H200 on-demand pricing that Spheron's $3.70/hr undercuts by nearly 18%. This post covers the actual per-hour rates, how committed pricing works, what the hidden costs are, and how Nebius stacks up against Spheron, RunPod, and CoreWeave so you can make a direct cost comparison before committing. For broader context on where Nebius fits in the competitive landscape, see our Spheron vs Nebius breakdown and the full Nebius alternatives comparison.
TL;DR: Nebius vs Spheron H100 and H200 (Jun 2026)
| GPU | Nebius On-Demand | Nebius Preemptible | Spheron On-Demand | Spheron Spot |
|---|---|---|---|---|
| H100 SXM (80GB) | $3.85/hr | $2.15/hr | $3.92/hr (SXM5) / $2.01/hr (PCIe) | $2.91/hr (SXM5) |
| H200 SXM (141GB) | $4.50/hr | $2.45/hr | $3.70/hr | $3.31/hr |
| Billing granularity | Per hour | Per hour | Per minute | Per minute |
| Quota required | Yes (at scale) | Yes (at scale) | No | No |
| Egress fees | $0.015/GiB (Standard) | $0.015/GiB (Standard) | Included | Included |
Nebius H100 SXM on-demand ($3.85/hr) is marginally cheaper than Spheron SXM5 ($3.92/hr). Nebius H200 on-demand ($4.50/hr) is significantly more expensive than Spheron H200 on-demand ($3.70/hr). Nebius preemptible ($2.15 H100, $2.45 H200) undercuts Spheron SXM5 spot for fault-tolerant SXM workloads. Spheron H100 PCIe at $2.01/hr is among the cheapest on-demand H100 options in this comparison, undercutting Nebius preemptible ($2.15/hr) with no interruption risk.
Pricing fluctuates based on GPU availability. The prices above are based on 29 Jun 2026 and may have changed. Check current GPU pricing → for live rates.
Nebius H100 Pricing Per Hour
Nebius offers H100 as HGX H100 NVLink SXM configurations. There is no H100 PCIe option on self-serve, and all H100 instances are sold with InfiniBand networking included.
| Configuration | On-Demand $/hr | Preemptible $/hr | vCPUs | RAM | InfiniBand |
|---|---|---|---|---|---|
| HGX H100 NVLink SXM (80GB) | $3.85 | $2.15 | 16 | 200 GB | NDR 400 Gb/s per link |
Each 8-GPU HGX H100 host delivers 3.2 Tbit/s aggregate InfiniBand throughput, which makes multi-node training viable at meaningful scale. However, Nebius only sells H100 SXM, not PCIe. Teams that need a single GPU for inference or don't need NVLink bandwidth pay the full $3.85/hr on-demand, or can use preemptible instances at $2.15/hr with interruption risk.
By comparison, Spheron offers H100 PCIe at $2.01/hr on-demand and H100 SXM5 at $3.92/hr on-demand. Nebius preemptible H100 at $2.15/hr is cheaper than Spheron's H100 SXM5 spot at $2.91/hr for fault-tolerant SXM workloads. However, Spheron H100 PCIe at $2.01/hr on-demand beats Nebius preemptible without any interruption risk. For the technical specs behind the H100 lineup, see the full H100 SXM5 specs breakdown covering memory bandwidth, NVLink topology, and FP8 throughput figures.
Teams looking to compare bare-metal H100 options across both providers should check Spheron's H100 SXM5 page for current availability and per-minute pricing alongside Spheron's PCIe tier.
Pricing fluctuates based on GPU availability. The prices above are based on 29 Jun 2026 and may have changed. Check current GPU pricing → for live rates.
Nebius H200 Pricing Per Hour
Nebius's H200 offering uses HGX H200 SXM5 hardware with 141GB HBM3e. The memory bandwidth jump from H100 to H200 (4.8 TB/s vs 3.35 TB/s) is meaningful for large model inference.
| Configuration | On-Demand $/hr | Preemptible $/hr | vCPUs | RAM |
|---|---|---|---|---|
| HGX H200 SXM5 (141GB HBM3e) | $4.50 | $2.45 | 16 | 200 GB |
Nebius H200 at $4.50/hr on-demand is significantly more expensive than Spheron's H200 SXM5 at $3.70/hr on-demand, a 17.7% premium. Nebius preemptible H200 at $2.45/hr undercuts Spheron H200 SXM5 spot at $3.31/hr for fault-tolerant training, though Spheron's per-minute billing means short H200 jobs avoid full-hour billing waste on Spheron's side.
Teams interested in H200 SXM5 pricing on Spheron can compare on-demand and spot tiers directly against Nebius's rates, with Spheron's on-demand H200 being the more cost-effective choice for continuous workloads.
Pricing fluctuates based on GPU availability. The prices above are based on 29 Jun 2026 and may have changed. Check current GPU pricing → for live rates.
Nebius Reserved and Committed-Use Pricing
Nebius does not publish a standard committed-use tier table on their pricing page. Reserved pricing requires a sales conversation and multi-month lock-in.
From what Nebius has indicated publicly: large committed allocations can reach approximately $2.00/hr per H100 GPU for extended reservations. Contact Nebius for current committed-use terms, as this is not a self-serve option.
Break-even math for committed pricing: at Nebius's $3.85/hr on-demand H100 rate, if committed pricing brings you to $2.00/hr, you save $1.85/hr per GPU. Running one H100 at 100% for 30 days costs $2,772 on-demand vs $1,440 committed. The monthly savings of $1,332 per GPU needs to offset the lock-in risk. Note that Spheron H100 PCIe at $2.01/hr on-demand with no commitment is already close to the committed price, without any lock-in or multi-month obligation.
The catch: committing to H100 capacity for 6-12 months carries real obsolescence risk when H200 and B200 instances are available at competitive prices now. Teams that committed to H100 contracts in 2024 are paying above-market rates relative to what Spheron lists today.
Spheron takes the opposite approach: per-minute billing, no reserved contracts required, and market-rate pricing that reflects current supply. For a deeper look at when committed GPU pricing makes sense, see the on-demand vs reserved GPU pricing analysis.
Nebius Hidden Costs
Object Storage egress: Nebius charges $0.015/GiB for Standard Object Storage egress to the public internet. Enhanced tier egress is free. If you store model checkpoints or datasets in Standard Object Storage and transfer them regularly, this cost adds up. At $0.015/GiB and 100GB of daily checkpoint traffic, that is $45/month in storage egress alone.
Inter-region networking: Data transferred between Nebius regions may carry additional fees. Verify current rates at nebius.com before planning a multi-region architecture.
Quota friction: Nebius's quota process is not a hard dollar cost, but the time it adds to onboarding is real. Teams that need compute for a deadline this week and get stuck waiting for quota approval effectively lose access to GPUs they need to pay for regardless.
Per-hour billing: Nebius bills in hourly increments. A 47-minute training job costs one full hour at $3.85 per GPU. Run 10 such jobs and you pay for 10 hours while consuming 7.8. At 8 GPUs, that rounding adds roughly $67/day in billing waste for 10 short daily jobs. Spheron's per-minute billing eliminates this overhead.
Committed pricing lock-in: The best Nebius rates require multi-month commitments. If your workload drops or you switch models, you are paying for capacity you are not using.
Nebius vs Spheron vs Other Neoclouds: H100 and H200 Price Comparison
This is the full multi-provider picture. Spheron pricing below is from the live API as of 29 Jun 2026.
| Provider | H100 On-Demand $/hr | H100 Spot $/hr | H200 On-Demand $/hr | H200 Spot $/hr | Spot Available | Billing |
|---|---|---|---|---|---|---|
| Spheron (H100 SXM5) | $3.92 | $2.91 | $3.70 | $3.31 | Yes | Per minute |
| Spheron (H100 PCIe) | $2.01 | N/A | No | Per minute | ||
| Nebius | $3.85 | $2.15 (preemptible) | $4.50 | $2.45 (preemptible) | Yes (preemptible) | Per hour |
| RunPod (Secure Cloud) | ~$3.29 | Available | ~$4.50 | Available | Yes | Per minute |
| Lambda Labs | $2.49-$3.44 | None | N/A | N/A | No | Per hour |
| Hyperstack | ~$1.90 (PCIe) | N/A | $3.50 | N/A | No | Per hour |
| CoreWeave | ~$6.16 | N/A | ~$6.31 | N/A | No | Per hour |
For the full 15+ provider pricing table including AWS, GCP, Azure, and OCI rates, see the GPU cloud pricing comparison 2026.
Nebius does offer preemptible pricing ($2.15/hr H100, $2.45/hr H200), which undercuts Spheron's SXM5 spot rates for fault-tolerant SXM workloads. Spheron H100 PCIe at $2.01/hr on-demand is among the lowest on-demand H100 prices in this table, undercuts Nebius preemptible ($2.15/hr) with no interruption risk, and sits close to Hyperstack's ~$1.90/hr PCIe listing, which is also competitive. Spheron's per-minute billing and no-quota model separate it from Nebius for teams that need fast access and short-job efficiency.
Pricing fluctuates based on GPU availability. The prices above are based on 29 Jun 2026 and may have changed. Check current GPU pricing → for live rates.
When Nebius Makes Sense vs an Aggregated Marketplace
Nebius wins on:
- EU GDPR compliance: Nebius has operated EU-region GPU infrastructure since their founding and has formal GDPR-compliant data processing agreements. For teams with a legal mandate for EU-only data processing, this is a genuine structural advantage.
- Slurm HPC clusters: Nebius offers managed Soperator for Slurm-based cluster configuration. Teams running HPC job scheduling on H100 fleets at scale get this out of the box.
- Data residency in specific regions: Nebius operates EU-north1, EU-north2, EU-west1, and US-central1 regions. The EU-north and EU-west options give EU-focused teams geographic specificity that a global multi-provider marketplace may not match today.
- Preemptible pricing for SXM hardware: Nebius H100 preemptible at $2.15/hr and H200 preemptible at $2.45/hr undercut Spheron's SXM5 spot rates. For fault-tolerant training that specifically needs SXM-class interconnect bandwidth, Nebius preemptible is the cheaper route.
- Fixed, predictable on-demand rates: One rate, no marketplace variation, no provider-level differences to navigate.
Spheron wins on:
- H100 PCIe pricing: $2.01/hr on-demand vs Nebius's $3.85/hr on-demand and even Nebius's $2.15/hr preemptible. For teams that don't need SXM interconnect bandwidth, Spheron PCIe saves 48% over Nebius on-demand with no commitment and no interruption risk.
- H200 on-demand pricing: Spheron H200 SXM5 at $3.70/hr is 17.7% cheaper than Nebius H200 at $4.50/hr on-demand. This is the most significant on-demand price gap between the two providers.
- No quota process: Deploy immediately after signup, no sales call required, GPU running in under 5 minutes.
- Per-minute billing: Short jobs don't waste a full hour of billing. Teams running many short training runs save meaningful money versus per-hour billing.
- Broader hardware catalog: A100 SXM4 at $1.69/hr on-demand, RTX 5090, GH200, B200, and more. Nebius's self-serve catalog covers H100, H200, L40S, and B200 only.
Worked TCO examples:
30-day inference workload, 4x H100 SXM, on-demand:
- Nebius (on-demand): $3.85 × 4 × 720 = $11,088/month
- Nebius (preemptible): $2.15 × 4 × 720 = $6,192/month
- Spheron (SXM5 on-demand): $3.92 × 4 × 720 = $11,290/month
- Spheron (PCIe on-demand): $2.01 × 4 × 720 = $5,789/month
- Spheron (SXM5 spot): $2.91 × 4 × 720 = $8,381/month
Spheron PCIe on-demand is the cheapest option for continuous inference at $5,789/month, undercutting even Nebius preemptible at $6,192/month and with no interruption risk. Spheron SXM5 on-demand is slightly more expensive than Nebius on-demand.
7-day training run, 8x H100 SXM, with checkpointing enabled:
- Nebius (on-demand): $3.85 × 8 × 168 = $5,174
- Nebius (preemptible): $2.15 × 8 × 168 = $2,890
- Spheron (SXM5 spot): $2.91 × 8 × 168 = $3,911
- Spheron (PCIe on-demand): $2.01 × 8 × 168 = $2,701
Spheron PCIe on-demand at $2,701 is the cheapest option overall, including versus Nebius preemptible at $2,890. For SXM-class hardware with checkpointing, Nebius preemptible ($2,890) is cheaper than Spheron SXM5 spot ($3,911). The per-minute billing advantage on Spheron means shorter training windows reduce the gap further.
How to Cut Per-Hour GPU Spend on Spheron
Use spot instances for fault-tolerant training. H100 SXM5 spot at $2.91/hr and H200 SXM5 spot at $3.31/hr are your cheapest SXM-class entry points on Spheron. Enable checkpoint-and-resume in your training loop (saving state every 15-30 minutes) and spot interruptions become a minor inconvenience rather than a project risk. Note that Nebius preemptible rates for SXM hardware are lower, so compare both if your workload requires NVLink interconnect.
Use H100 PCIe for single-GPU inference and fine-tuning. A100 SXM4 at $1.69/hr on-demand covers most fine-tuning and smaller inference workloads. H100 PCIe at $2.01/hr on-demand is the right tier when you need H100 performance without SXM networking overhead. If your model fits in 80GB and you don't need H100 throughput, the A100 is significantly cheaper.
Exploit per-minute billing for short jobs. Short evaluation runs, hyperparameter sweeps, and dataset preprocessing jobs that take 10-30 minutes cost 10-30 minutes on Spheron. On Nebius or Lambda Labs, they cost a full hour. At 20 short jobs per day on 2 GPUs, that difference compounds quickly.
Deployment guides and SSH setup documentation are available at docs.spheron.ai for connecting to your first instance and configuring distributed training.
Spheron H100 and H200 Live Pricing (Jun 2026)
Current Spheron rates fetched from the Spheron pricing API on 29 Jun 2026:
| GPU | On-Demand $/hr | Spot $/hr | Notes |
|---|---|---|---|
| H100 SXM5 | $3.92 | $2.91 | Per-minute billing |
| H100 PCIe | $2.01 | N/A | Lowest H100 entry tier |
| H200 SXM5 | $3.70 | $3.31 | Per-minute billing |
| A100 SXM4 (80GB) | $1.69 | N/A | Per-minute billing |
No commitment required for any of the above. No quota process. SSH root access on bare-metal hardware within minutes of provisioning.
Pricing fluctuates based on GPU availability. The prices above are based on 29 Jun 2026 and may have changed. Check current GPU pricing → for live rates.
FAQ
How much does Nebius charge for H100 GPUs per hour in 2026?
Nebius charges approximately $3.85/hr per GPU for HGX H100 NVLink SXM (80GB) on-demand as of 2026. Preemptible instances are available at $2.15/hr. Each instance includes 16 vCPUs and 200GB RAM. Committed multi-month pricing can reduce rates further, but requires a sales process. Spheron's H100 PCIe starts at $2.01/hr and H100 SXM5 spot pricing runs $2.91/hr, both with per-minute billing and no commitment required.
What is Nebius H200 pricing per hour?
Nebius HGX H200 (141GB HBM3e) runs approximately $4.50/hr on-demand per GPU as of 2026. Preemptible H200 instances are available at $2.45/hr. Committed rates require contacting Nebius sales for current terms. Spheron's H200 SXM5 is $3.70/hr on-demand and $3.31/hr spot, with per-minute billing. Spheron's H200 on-demand is 17.7% cheaper than Nebius on-demand.
Does Nebius offer spot or preemptible GPU instances?
Yes. Nebius offers preemptible GPU instances at $2.15/hr for H100 SXM and $2.45/hr for H200 SXM as of 2026. These rates undercut Spheron's SXM5 spot rates ($2.91/hr H100, $3.31/hr H200). However, Nebius preemptible is billed per hour, while Spheron spot is billed per minute. Spheron H100 PCIe at $2.01/hr on-demand is cheaper than Nebius preemptible and carries no interruption risk.
What are Nebius's hidden GPU costs beyond the hourly compute rate?
Watch for: Object Storage Standard egress at $0.015/GiB, storage volumes charged separately from compute, per-hour billing that rounds up partial jobs, and minimum commitment periods for reserved pricing. Data transfer between Nebius regions may carry fees. Spheron includes egress and bills per minute with no minimum commitment.
How does Nebius H100 pricing compare to Spheron?
For H100 SXM on-demand, Nebius ($3.85/hr) is marginally cheaper than Spheron SXM5 ($3.92/hr). For preemptible/spot: Nebius H100 preemptible at $2.15/hr undercuts Spheron SXM5 spot at $2.91/hr for fault-tolerant workloads needing SXM interconnect. However, Spheron H100 PCIe at $2.01/hr on-demand is cheaper than Nebius preemptible with no interruption risk. For H200, Spheron's $3.70/hr on-demand is 17.7% cheaper than Nebius's $4.50/hr.
Nebius is a serious option for EU-focused teams with data residency requirements or HPC cluster needs. Their H100 NVLink on-demand rate at $3.85/hr is marginally competitive for SXM hardware, their InfiniBand infrastructure supports multi-node training at scale, and their preemptible pricing ($2.15/hr H100, $2.45/hr H200) undercuts Spheron's SXM5 spot rates for fault-tolerant SXM workloads. Where Nebius falls short: H200 on-demand at $4.50/hr is 17.7% more expensive than Spheron's $3.70/hr, per-hour billing penalizes short jobs, and a quota process slows access for teams that need to move quickly. Spheron's H100 PCIe at $2.01/hr undercuts Nebius preemptible for teams that don't need SXM interconnect, with no interruption risk.
Nebius H100 NVLink on-demand runs $3.85/hr with preemptible at $2.15/hr and per-hour billing. Spheron's H100 PCIe starts at $2.01/hr on-demand, undercutting Nebius preemptible with no interruption risk, and H200 on-demand at $3.70/hr beats Nebius's $4.50/hr.
On-demand H100 access → | H200 SXM5 pricing → | Get started →
Quick Setup Guide
Go to nebius.com/prices and filter by GPU compute. Confirm whether the rate you see is on-demand or preemptible. For HGX H100 NVLink SXM, the public on-demand rate is approximately $3.85/hr per GPU, with preemptible available at $2.15/hr. For HGX H200, approximately $4.50/hr on-demand and $2.45/hr preemptible. Committed rates require contacting Nebius sales.
Multiply your expected daily GPU-hours by 30 and by Nebius's applicable rate ($3.85/hr on-demand or $2.15/hr preemptible for H100; $4.50/hr on-demand or $2.45/hr preemptible for H200). Add Object Storage Standard egress at $0.015/GiB if you transfer checkpoints. Compare this total against Spheron's per-minute billing at spheron.network/pricing/ to find the break-even point for your workload.
Visit spheron.network/pricing/ and filter by H100 or H200. Check both on-demand and spot rates. Spheron H100 PCIe at $2.01/hr on-demand is cheaper than Nebius H100 preemptible at $2.15/hr, with no interruption risk. Spheron H200 on-demand at $3.70/hr is 17.7% cheaper than Nebius H200 on-demand at $4.50/hr. Spot is suitable for training runs with checkpointing enabled.
Sign in at app.spheron.ai, select H100 or H200 from the GPU catalog, choose on-demand or spot, and deploy. SSH root access is available within minutes. Spheron bills per minute with no minimum commitment and no quota application required.
Frequently Asked Questions
Nebius charges approximately $3.85/hr per GPU for H100 NVLink SXM (HGX H100, 80GB) on-demand as of 2026. Preemptible H100 instances are available at $2.15/hr. Each instance includes 16 vCPUs and 200GB RAM. Committed multi-month pricing can reduce rates further for large allocations negotiated through Nebius sales. Spheron's H100 PCIe starts at $2.01/hr on-demand, and H100 SXM5 spot pricing runs $2.91/hr with per-minute billing and no commitment required.
Nebius HGX H200 (141GB HBM3e) runs approximately $4.50/hr on-demand per GPU as of 2026. Preemptible H200 instances are available at $2.45/hr. Committed-use rates are negotiated via Nebius sales for longer reservations. Spheron's H200 SXM5 starts at $3.70/hr on-demand and $3.31/hr spot, with per-minute billing and no minimum commitment.
Yes. Nebius offers preemptible GPU instances at $2.15/hr for H100 SXM and $2.45/hr for H200 SXM as of 2026. These rates are lower than Spheron's SXM5 spot rates ($2.91/hr for H100 SXM5, $3.31/hr for H200 SXM5). However, Spheron H100 PCIe at $2.01/hr on-demand is cheaper than Nebius preemptible H100 with no interruption risk. Key Spheron advantages over Nebius include per-minute billing (not per-hour), no quota process, and H200 on-demand at $3.70/hr versus Nebius's $4.50/hr.
Beyond Nebius's hourly GPU rate, watch for: Object Storage Standard egress at $0.015/GiB (Enhanced tier egress is free), storage volumes charged separately, and minimum commitment periods for reserved pricing. Data transfer between Nebius regions may carry fees. Nebius offers zero inter-instance networking within a cluster, but verify egress to public internet at nebius.com before budgeting.
For H100 SXM on-demand, Nebius runs $3.85/hr versus Spheron SXM5 at $3.92/hr (Nebius is marginally cheaper). For preemptible/spot: Nebius H100 preemptible is $2.15/hr while Spheron H100 SXM5 spot is $2.91/hr (Nebius is cheaper for preemptible SXM). However, Spheron H100 PCIe at $2.01/hr on-demand undercuts both and carries no interruption risk. For H200, Spheron's $3.70/hr on-demand beats Nebius's $4.50/hr by 17.7%. Spheron also offers per-minute billing and no quota process.