Nebius prices HGX H100 at $3.85/GPU-hr on-demand. CoreWeave prices the same GPU generation at $6.16/GPU-hr, about 60% higher, and only sells it in 8-GPU nodes (coreweave.com/pricing, nebius.com/prices). That single fact drives most of the comparison shopping happening between these two neoclouds right now, but it isn't the whole picture. CoreWeave carries a Platinum ClusterMAX rating and a $99.4 billion contracted backlog; Nebius sells GPUs one at a time and just posted 684% year-over-year revenue growth. This post puts CoreWeave vs Nebius pricing, contract terms, and catalog breadth side by side, then covers where Spheron fits as the no-minimum alternative to both.
If you want the deep-dive on either provider alone, see our CoreWeave H100 and H200 pricing breakdown and Nebius H100 and H200 pricing breakdown, which go further into each provider's hidden costs and billing structure.
CoreWeave vs Nebius at a Glance
| Metric | CoreWeave | Nebius |
|---|---|---|
| H100 on-demand | $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) |
| Typical contract | ~5-year weighted average for new capacity | Multi-month reservations |
| GPU catalog breadth | 11 models | 8 models |
| ClusterMAX 2.0 tier | Platinum (sole member) | Gold |
| Billing granularity | Hourly | Hourly |
CoreWeave wins on scale and reliability rating. Nebius wins on price and access granularity. Neither bills by the minute or skips a sales conversation for its best rates, which is where a third option like Spheron changes the math later in this post.
Pricing fluctuates based on GPU availability. The prices above are based on 4 Aug 2026 and may have changed. Check current GPU pricing → for live rates.
H100 and H200 On-Demand Pricing Head to Head
The headline gap is real and consistent across both GPU generations: CoreWeave's on-demand rate runs roughly 40-60% above Nebius's, but the two companies structure the purchase itself very differently, and that structural difference matters as much as the sticker price.
CoreWeave's 8-GPU-Only Node Pricing
CoreWeave prices HGX H100 at $49.24/hr per 8-GPU node on-demand, which works out to $6.16/GPU-hr, and HGX H200 at $50.44/hr per node, or $6.31/GPU-hr (coreweave.com/pricing). There is no smaller unit to buy. A single-GPU inference deployment and an 8-GPU distributed training job are billed identically, at the full node rate, because CoreWeave's current pricing tier doesn't offer anything below eight.
That's a defensible design for a training job that genuinely needs the full NVLink domain across all eight GPUs. It's a worse fit for the much larger population of teams running single-model inference or fine-tuning under 70B parameters, who end up paying for seven idle GPUs to access the one they need.
Nebius's Per-GPU Pricing (1x and 8x Presets)
Nebius prices HGX H100 SXM at $3.85/GPU-hr on-demand and HGX H200 SXM at $4.50/GPU-hr, and both figures apply whether you provision a 1-GPU preset (1gpu-16vcpu-200gb) or the full 8-GPU preset (8gpu-128vcpu-1600gb) (nebius.com/prices, docs.nebius.com). A one-GPU checkout is a normal, self-serve option, not a sales exception. If your workload is inference or single-node fine-tuning, Nebius's published rate is the rate you actually pay, at the GPU count you actually need.
Why the Per-GPU Math Favors Nebius on Paper, and Where It Doesn't
Run the arithmetic on a continuous 8-GPU H100 cluster and the gap is stark: 720 hours a month on CoreWeave on-demand costs $35,452.80 ($49.24 x 720), versus $22,176 on Nebius on-demand (8 x $3.85 x 720), a savings of roughly $13,277 a month for identical hardware. That math holds cleanly for anyone actually running 8 GPUs continuously.
It breaks down the moment your real utilization is lower. A team that needs 2 GPUs pays Nebius for 2 GPUs and CoreWeave for 8, because CoreWeave has no smaller unit to sell. The "CoreWeave is 60% more expensive" framing is accurate at matched GPU counts, but the more common failure mode for smaller teams is paying CoreWeave's node rate for capacity they never use in the first place. For the hardware-level tradeoffs between the two generations before you pick one, see our NVIDIA H100 vs H200 comparison.
Spot and Preemptible Pricing Compared
The pricing gap narrows sharply once you move off on-demand. CoreWeave's H100 spot rate runs about $2.46/GPU-hr against Nebius's $2.15/GPU-hr preemptible rate, a difference of roughly 14%. On H200, CoreWeave spot at about $2.62/GPU-hr sits only about 7% above Nebius preemptible at $2.45/GPU-hr.
| GPU | CoreWeave spot | Nebius preemptible | Gap |
|---|---|---|---|
| H100 | ~$2.46/GPU-hr | $2.15/GPU-hr | ~14% |
| H200 | ~$2.62/GPU-hr | $2.45/GPU-hr | ~7% |
Both spot tiers carry the same caveat: capacity can be reclaimed without notice, and neither provider drops the 8-GPU-vs-1-GPU access difference just because you're on the discounted tier. CoreWeave's spot H100 and H200 nodes are still sold only in 8-GPU units.
Committed-Use Discounts and Contract Terms Compared
Both companies gate their best rates behind a sales conversation, but the ceilings and the underlying business models differ enough to matter for anyone evaluating a multi-month or multi-year deal.
CoreWeave: Up to 60% Off, 5-Year Average Contracts
CoreWeave advertises "up to 60% discounts over on-demand prices for committed usage" with no published rate tiers (coreweave.com/pricing). Applied to the $6.16/GPU-hr H100 on-demand rate, a full 60% discount would land around $2.46/GPU-hr, still above Nebius's uncommitted preemptible rate.
The company's business model runs on long-dated capacity commitments. CFO Nitin Agrawal told investors on the Q1 2026 earnings call that the "weighted average contract length for new capacity" sits at approximately 5 years, backing a $99.4 billion contracted revenue backlog, up nearly 50% sequentially, with 36% expected to be recognized within 24 months (Investing.com earnings call transcript). On the same call, Agrawal said: "We remain largely sold out of our 2026 capacity, with prices increasing across the board from Ampere to Hopper to Blackwell." That's a company financing its buildout on multi-year commitments, not a marketplace competing to fill idle GPUs, and it shows in how it prices short-term access.
Nebius: Up to 35% Off, Multi-Month Reservations
Nebius advertises "up to 35% less than on-demand rates by reserving large-scale clusters for multiple months" (nebius.com/prices), a lower ceiling than CoreWeave's but on a shorter, less binding commitment structure. Nebius doesn't publish specific rate tiers either, so getting an actual number still means a sales call. Nebius CEO Arkady Volozh explained the company's aggressive 2026 capex raise, from $16-20 billion to $20-25 billion, with a simple logic: "Everything we build, we sell, and we are still in the very early days" (Nebius Q1 2026 earnings call, May 13, 2026). Nebius reported $399 million in revenue for the quarter, up 684% year over year.
For a broader framework on when a reserved commitment beats staying on-demand, see serverless vs on-demand vs reserved GPU pricing.
GPU Catalog Breadth: CoreWeave's 11 Models vs Nebius's 8
CoreWeave's current pricing page lists 11 distinct GPU models: GB300 NVL72, GB200 NVL72, HGX B300, HGX B200, RTX PRO 6000 Blackwell Server Edition, HGX H100, HGX H200, GH200, L40, L40S, and A100 (coreweave.com/pricing). Nebius's self-serve catalog covers 8: HGX H100, HGX H200, HGX B200, HGX B300, GB200 NVL72 and GB300 NVL72 (quote-only), L40S, and RTX PRO 6000 (nebius.com/prices).
The gap matters for teams that need older or cheaper hardware alongside their frontier GPUs. CoreWeave's inclusion of GH200, L40, and A100 as self-serve line items gives it more range for mixed fleets, training on Hopper while running inference on L40; Nebius's leaner catalog is more current-generation-focused but leaves those older tiers to other providers.
Scale, Reliability Rating, and Business Model Differences
SemiAnalysis's ClusterMAX 2.0 GPU cloud rating puts CoreWeave alone at the top: "CoreWeave retains top spot as the only member of the Platinum tier." Nebius lands one tier down, grouped with the strongest of the rest: "Nebius, Oracle and Azure are the top providers within the Gold tier. Crusoe and new entrant Fluidstack also achieve Gold tier" (SemiAnalysis, ClusterMAX 2.0). Gold isn't a knock; ClusterMAX 2.0 scored 84 providers directly out of 209 tracked, and Gold sits one tier below the single Platinum slot CoreWeave occupies alone.
Scale tells a similar story. CoreWeave is targeting roughly 1.7 GW of active power capacity by the end of 2026, according to BofA Securities estimates (Finimize). Nebius is targeting 800 MW to 1 GW of active capacity over the same window, up from an estimated 220 MW at the end of 2025 (The Motley Fool). CoreWeave's target is roughly 1.7-2x Nebius's, not an order of magnitude apart, but CoreWeave is further along in absolute terms and backing that scale with a much larger, longer-dated contract book.
The business models diverge from there. CoreWeave finances data center buildout against multi-year enterprise contracts and sells whatever on-demand capacity is left over. Nebius runs closer to a growth-stage cloud provider, self-serve by default, funding expansion off equity and debt rather than a backlog of 5-year deals.
Billing Granularity and Hidden Costs
Both providers bill in hourly increments, no exception. A 40-minute job costs a full hour on either platform. Neither publishes storage or egress rates on its primary pricing page; CoreWeave's node bundle includes local storage but quotes persistent volumes and networking add-ons through sales, and Nebius charges standard Object Storage egress separately (not itemized on the core compute pricing page), with Enhanced-tier egress free. Budget for a follow-up sales conversation on either side before you finalize a number.
The biggest structural cost isn't a line item at all. On CoreWeave, it's the 8-GPU floor: if your real utilization is 2 GPUs, you're still billed for 8. On Nebius, it's the quota and sales gating that shows up once you scale past self-serve presets into larger committed allocations, adding onboarding time that has a real cost even though it isn't a dollar figure on the invoice.
CoreWeave vs Nebius: Which One Fits Your Workload
CoreWeave fits if: you're running distributed training that genuinely needs a full 8-GPU NVLink domain, you can commit to a multi-year contract averaging around 5 years, and reliability rating (Platinum ClusterMAX) matters more to your procurement process than the per-GPU sticker price.
Nebius fits if: you need H100 or H200 access at single-GPU granularity without an enterprise sales cycle, EU data residency is a requirement, and your workload is inference or fine-tuning rather than a full-node training run.
Neither fits if: you want per-minute billing, no quota process, and access to more than a handful of GPU generations without a sales call. That's the gap Spheron is built to fill.
Where Spheron Fits Between the Two
Spheron rents H100 and H200 by the single GPU, bills per minute instead of rounding up to the hour, and doesn't gate on-demand access behind a quota or sales conversation. Where CoreWeave forces an 8-GPU minimum and Nebius requires a quota process once you scale past self-serve presets, Spheron's on-demand and spot pricing is available immediately at whatever GPU count your workload actually needs. For the direct numbers against each provider individually, see Spheron vs CoreWeave and Spheron vs Nebius. If your evaluation is really about getting off one of these two specifically, the CoreWeave alternatives and Nebius alternatives roundups go deeper on the non-price tradeoffs. For the full 15+ provider table including AWS, GCP, Azure, and OCI, see the GPU cloud pricing comparison 2026, and for throughput and spec data beyond price, our GPU cloud benchmarks page covers A100 through B300. Deployment guides for connecting to a new instance are at docs.spheron.ai.
If an 8-GPU minimum on one side and a quota process on the other don't fit how you actually work, single-GPU H100 and H200 access with per-minute billing is the more direct route.
Spheron H100 instances → | H200 GPU pricing → | View all GPU pricing →
Frequently Asked Questions
Nebius is cheaper on paper. Nebius prices HGX H100 SXM at $3.85/GPU-hr on-demand versus CoreWeave's $6.16/GPU-hr, a roughly 60% gap. But CoreWeave only sells H100 in 8-GPU nodes, so the comparison only holds if you actually need all eight GPUs. Nebius also sells a 1-GPU H100 preset, which CoreWeave doesn't offer at any self-serve price.
Nebius again, at $4.50/GPU-hr on-demand versus CoreWeave's $6.31/GPU-hr, about a 40% gap. On spot/preemptible the gap narrows to roughly 7%: CoreWeave H200 spot runs about $2.62/GPU-hr against Nebius's $2.45/GPU-hr preemptible rate.
No. CoreWeave's current pricing page lists HGX H100 and HGX H200 exclusively as 8-GPU nodes, with no self-serve single-GPU option. Nebius sells both a 1-GPU preset and an 8-GPU preset for the same hardware, so a one-GPU inference job is a normal checkout on Nebius and effectively unavailable on CoreWeave's public pricing tier.
CoreWeave's current pricing page lists 11 distinct GPU models, from GB300 NVL72 down to A100. Nebius's self-serve catalog covers 8, topping out at HGX B300 and GB200/GB300 NVL72 (quote-only) and not listing GH200, L40, or A100 as standalone self-serve options.