Nebius is the cheaper, more flexible GPU cloud for most rentals, while CoreWeave is built for large distributed training with top-tier reliability. The right pick depends on whether you need a single GPU or a full 8-GPU cluster.
Both CoreWeave and Nebius rent NVIDIA GPUs for AI workloads. They differ on price, minimum node size, billing granularity, and the commitments their best rates require.
Takeaways
- Nebius is cheaper per GPU: its H100 runs $3.85/GPU-hr on-demand versus CoreWeave's $6.16/GPU-hr.
- CoreWeave sells H100 and H200 only in 8-GPU nodes, while Nebius rents a single GPU on-demand.
- Nebius bills per-second and CoreWeave bills hourly, which matters for short, iterative jobs.
CoreWeave vs Nebius Pricing at a Glance
Nebius leads on price while CoreWeave leads on scale and reliability. All figures are on-demand per GPU-hr unless noted, current as of September 2026.
| Metric | CoreWeave | Nebius |
|---|---|---|
| H100 on-demand1 | $6.16/GPU-hr (8-GPU node only) | $3.85/GPU-hr (1 or 8-GPU preset) |
| H200 on-demand | $6.31/GPU-hr (8-GPU node only) | $4.50/GPU-hr (1 or 8-GPU preset) |
| H100 spot / preemptible | $2.46/GPU-hr | $2.15/GPU-hr |
| H200 spot / preemptible | $2.62/GPU-hr | $2.45/GPU-hr |
| Minimum node size | 8 GPUs | 1 GPU |
| Committed-use discount | Up to 60% (sales only) | Up to 35% (sales only) |
| GPU catalog breadth | 11 models | 8 models |
| Billing granularity | Hourly | Per-second |
| 1 Last Reviewed on September 24, 2026. GPU pricing changes frequently, so confirm current rates at the source before budgeting. | ||
CoreWeave wins on reliability and multi-GPU scale. Nebius wins on price and single-GPU granularity. CoreWeave bills hourly and Nebius bills per-second, and neither gives its best rate without a sales conversation.
CoreWeave vs Nebius H100 and H200 On-Demand Pricing
CoreWeave's 8-GPU Node Floor
CoreWeave sells H100 and H200 in 8-GPU nodes on its public pricing tier. HGX H100 lists at about $49.24/node-hr, or $6.16/GPU-hr, and HGX H200 lands near $6.31/GPU-hr on the same 8-GPU basis. For these two cards, no smaller unit exists on that tier.
The 8-GPU design fits distributed training that needs all eight GPUs joined by NVLink (NVIDIA's high-speed GPU interconnect). It fits poorly for single-model inference or fine-tuning under 70B parameters, where teams pay for seven idle GPUs to reach the one they need.
Nebius Per-GPU Pricing
Nebius prices the same hardware per GPU and provisions one at a time. HGX H100 SXM lists at $3.85/GPU-hr and HGX H200 SXM at $4.50/GPU-hr, whether you launch a 1-GPU or 8-GPU preset.
Single-GPU granularity is decisive for inference or single-node fine-tuning. You pay Nebius's rate at the GPU count your workload uses, instead of rounding up to a node.
Spot and Preemptible Pricing Compared
The price gap narrows sharply off on-demand. CoreWeave's H100 spot runs about $2.46/GPU-hr against Nebius's $2.15/GPU-hr preemptible rate, roughly 14% apart. On H200, CoreWeave spot at about $2.62/GPU-hr sits about 7% above Nebius preemptible at $2.45/GPU-hr.
Both discounted tiers can be reclaimed without notice, so they suit only workloads that checkpoint often and restart elsewhere. Neither provider changes its access model on the cheaper tier, so CoreWeave spot H100s and H200s still ship only in 8-GPU units.
Contracts, Minimums, and How Fast You Can Start
Both providers put their lowest rates behind a sales call. CoreWeave advertises up to 60% off on-demand for committed use, and its weighted-average contract length has risen to about five years by the company's own account. Long-dated enterprise commitments define CoreWeave's model, which leaves little short-term flexibility in public pricing.
Nebius advertises up to 35% off through multi-month reservations, a lower ceiling on a shorter commitment. Neither provider publishes committed-rate tiers, so the discounted number still requires a sales conversation.
Time-to-first-GPU is the quieter cost. CoreWeave has stated it is largely sold out of near-term capacity, and Nebius adds quota and onboarding steps past its self-serve presets. Both paths can involve waiting, and waiting costs money even when it never appears on an invoice.
Free trials are thin on both sides. CoreWeave offers no self-serve free tier or open credits. Nebius provides only a nominal trial credit of about $1, valid for 30 days. Real evaluation on either provider means paying from your first serious run.
GPU Catalog and Availability
CoreWeave lists the broader catalog: 11 GPU models versus Nebius's 8. CoreWeave spans GB300 NVL72 and GB200 NVL72 down to GH200, L40, L40S, and A100. Nebius runs a leaner, current-generation lineup topping out at HGX B300 and GB200/GB300 NVL72 (quote-only), focused on H100, H200, B200, L40S, and RTX PRO 6000.
The breadth gap matters mainly for mixed fleets. CoreWeave's inclusion of A100, L40, and GH200 as self-serve line items suits training on Hopper while serving inference on older, cheaper cards. Nebius covers current-generation Hopper and Blackwell with less menu to navigate.
Which One Fits Your Workload: CoreWeave, Nebius, or Neither
Match your workload to one of three buckets before comparing sticker prices. The right unit of purchase saves more than a lower per-GPU rate.
| Choose this | If your situation looks like |
|---|---|
| CoreWeave | You run distributed training that needs a full 8-GPU NVLink domain, you can commit to a multi-year contract, and a top reliability rating matters to procurement more than the per-GPU price. |
| Nebius | You want H100 or H200 access at single-GPU granularity without an enterprise sales cycle, a European-based provider suits your data-residency needs, and your workload is inference or fine-tuning rather than a full-node training run. |
| Neither | You want one or two GPUs, per-minute billing, no quota process, and no sales call, all available the moment you need them. That is a different tier of provider. |
The single-GPU bucket is larger than it looks. Most individual developers, small ML teams, and anyone iterating daily do not need an 8-GPU cluster or a five-year contract. They need one fast GPU, cheaply, without procurement.
A Developer-First Alternative: Thunder Compute
Thunder Compute rents a single GPU on-demand with per-minute billing and no commitment. An H100 PCIe instance costs $3.20/hr and an A100 80GB costs $1.09/hr, both with persistent storage and no contract.
| GPU | Thunder Compute | Nebius (on-demand) | CoreWeave (on-demand) |
|---|---|---|---|
| A100 80GB | $1.09/hr (1 GPU) | Current-gen focus | $2.70/GPU-hr (GPU component, 8-GPU node) |
| H100 80GB | $3.20/hr PCIe (1 GPU) | $3.85/GPU-hr SXM (1 or 8-GPU) | $6.16/GPU-hr SXM (8-GPU node only) |
| Minimum | 1 GPU | 1 GPU | 8 GPUs |
| Billing | Per minute | Per-second | Hourly |
| Contract | None | Multi-month for best rates | ~5-year average for committed rates |
Thunder Compute can launch instances from a VS Code or Cursor extension, so you go from editor to live GPU in under a minute, with no quota approval. One-click templates for local LLMs and image generation tools like Stable Diffusion and ComfyUI run common workloads without setup.
See Thunder Compute's full guide to the cheapest cloud GPU providers.
Last Thoughts on CoreWeave vs Nebius
CoreWeave and Nebius serve different ends of the same market: CoreWeave for reliability-rated, multi-year, 8-GPU training clusters, and Nebius for cheaper, per-GPU, self-serve inference and fine-tuning. Nebius wins on price for most workloads that fit its granularity, while CoreWeave wins for full-cluster training on a long commitment. If you need a single GPU today, billed by the minute with no contract or quota, Thunder Compute is worth checking before you commit to either.
CoreWeave vs Nebius Comparison FAQ
Is CoreWeave or Nebius Cheaper for H100?
Nebius is cheaper, at $3.85/GPU-hr on-demand versus CoreWeave's $6.16/GPU-hr. CoreWeave sells H100 only in 8-GPU nodes, so that rate applies only if you need all eight. Nebius also offers a single-GPU H100 preset.
Does CoreWeave Rent Single H100 or H200 GPUs?
No. CoreWeave's public pricing lists HGX H100 and H200 only as 8-GPU nodes. Nebius sells both 1-GPU and 8-GPU presets, so a one-GPU job is a normal checkout on Nebius and unavailable on CoreWeave's public tier.
What Is the Cheapest Way to Rent a Single H100?
Providers that sell per-GPU beat those that sell per-node. Nebius offers a single H100 at $3.85/hr, and Thunder Compute offers a single H100 PCIe at $3.20/hr with per-minute billing. CoreWeave's 8-GPU floor is the most expensive route.
Do CoreWeave or Nebius Require a Contract?
Both offer on-demand access without a contract, but their best rates require commitments. CoreWeave's committed pricing runs on multi-year contracts averaging about five years. Nebius uses shorter multi-month reservations.
Do CoreWeave or Nebius Offer a Free Trial?
Not meaningfully. CoreWeave offers no self-serve free tier or open credits. Nebius provides a nominal trial credit of $1, valid for 30 days. Real evaluation on either means paying from the first run.