Nebius Review for AI and Machine Learning Workloads

The research in brief

Nebius is a good fit for AI startups and data science teams that need self-service H100/H200/B200-class GPU capacity, multi-GPU training clusters, and optional preemptible pricing for GPU cloud compute for AI training. Four of six platforms named Nebius during the ranking stage (Gemini, Grok, Kimi, Perplexity), giving it a 66.7% share of included platform responses and an average listed rank of 5.5. The strongest reason to consider it is a technically complete training stack: current NVIDIA GPUs, 8-GPU HGX hosts, InfiniBand, Kubernetes, and Slurm, with published on-demand and spot rates. The main limitation is that spot capacity is dynamic and interruptible, and public contract, SLA, and framework-version details remain incomplete.

Research Snapshot

Why Nebius Qualified for This Study

Questions explored

• Is Nebius a good choice for GPU cloud compute for AI training?
• Why did multiple AI platforms rank Nebius for AI training workloads?

Nebius qualified because it appeared in the ranking stage of four of the six included platforms and was assessed for fit by all six. Its ranks ranged from 4 (Perplexity) to 7 (Grok), with Kimi at 5 and Gemini at 6, producing an average listed rank of 5.5 and a final rank of 5.

The qualification is not based on a single capability. Every platform that evaluated Nebius for this use case returned a positive fit rating: five rated it "good" and Gemini rated it "strong." No platform rated Nebius as a poor or unsuitable fit for GPU cloud compute for AI training.

The platforms converged on the same structural reasons: current-generation NVIDIA GPUs, multi-node InfiniBand clusters, managed Kubernetes and Slurm, self-service access, and published per-GPU-hour pricing with a preemptible tier. These map directly onto the study's category criteria of GPU model availability, GPU pricing per hour, multi-GPU cluster support, framework and driver support, and spot or preemptible pricing options.

One qualification caveat: platform mentions count only platforms that named Nebius during ranking discovery. All six included platforms evaluated fit, but only four surfaced Nebius in the ranking stage.

The hosting offering in focus

Questions explored

• Which Nebius product should a buyer choose for multi-node AI training on H100 or H200 GPUs?
• Does Nebius offer 8-GPU HGX systems with InfiniBand for distributed training?

The relevant offering is Nebius AI Cloud GPU instances and AI clusters, particularly NVIDIA HGX H100, H200, B200, and B300 systems with Kubernetes, Slurm, and InfiniBand support. Platforms described this offering under several names, including "Nebius AI Cloud," "GPU instances," "AI clusters," "HGX systems," and "preemptible VMs."

Nebius lists NVIDIA HGX H100, H200, B200, and B300 systems, plus RTX PRO 6000 and L40S options ^[16]. Kimi reported the range extends to GB200/GB300 NVL72 ^[29]. Google reported that the us-central1 (Kansas City) region hosts B200 and H200 configurations, while us-north1 (Woodbury) supports Blackwell B300 platforms ^[4].

For distributed training, Nebius describes 8-GPU HGX systems, NVIDIA NVLink within hosts, and Quantum-2 or Quantum-X800 InfiniBand networking, with advertised scaling to thousands of GPUs ^[18]. Grok reported multi-node clusters with InfiniBand at 400–800 Gbps and non-virtualized GPUs ^[14].

The reviewed materials do not identify one single universally defined plan named "Nebius GPU Cloud," and no consolidated product specification or service-level document was located. Buyers should confirm exact SKU naming with Nebius before contracting.

What the AI Platforms Agreed About

Questions explored

• What do AI platforms agree Nebius offers for AI training workloads?
• Is Nebius pricing for H100 and H200 GPUs competitive for training?

The platforms agreed on four points: current NVIDIA GPU availability, multi-node InfiniBand cluster support, managed Kubernetes and Slurm, and the existence of a preemptible or spot tier.

On GPU availability, ChatGPT, DeepSeek, Grok, Perplexity, Kimi, and Gemini all reported H100/H200-class capacity, with most also reporting B200 and B300 ^[16].

On cluster architecture, ChatGPT, DeepSeek, Grok, Perplexity, and Gemini all described multi-node InfiniBand-connected clusters suitable for distributed training ^[15].

On orchestration, ChatGPT, DeepSeek, Grok, Kimi, and Gemini reported Kubernetes and/or Slurm support, with Kimi and Gemini both stating managed Kubernetes and Slurm carry no separate charge ^[15].

On spot pricing, ChatGPT, Grok, Perplexity, and Gemini all confirmed a preemptible or spot tier exists ^[10]. Deepseek reported spot availability as unclear, which is a disagreement rather than a contradiction.

Platform agreement here reflects consistent reporting across AI systems. It does not independently prove product quality or performance.

Where the AI Platforms Disagreed or Were Uncertain

Questions explored

• Why do AI platforms disagree about Nebius GPU pricing and spot availability?
• What is unverified about Nebius capacity, SLAs, and framework support for training?

Pricing is the largest conflict. Openai reported H100 at $3.85/GPU-hour effective October 1, 2026 ^[16]. Grok, Perplexity, and Gemini reported H100 at $4.50/GPU-hour ^[3]. Grok cited a price increase on October 1, 2026, raising H100 from $3.85 to $4.50 ^[30]. Kimi reported $3.85 on-demand and $2.15 preemptible ^[29]. Deepseek cited a third-party aggregator listing H100 in the mid-$2/hour range ^[23].

H200 shows the same pattern: ChatGPT reported $4.50, while Grok, Perplexity, and Gemini reported $5.40 ^[16]. B200 was reported at $7.15 by ChatGPT and Kimi, and $8.50 by Perplexity and Gemini ^[16]. B300 was reported at $7.85 by ChatGPT and Kimi, and $9.50 by Gemini ^[16].

Spot availability is a second disagreement. Deepseek stated that spot/preemptible pricing "is not clearly documented in the sources reviewed" and advised buyers not to assume it is available ^[20]. Openai, Grok, Perplexity, and Gemini all confirmed it exists ^[10]. Deepseek's research date was 2026-01-15, roughly nine months before the other platforms, which may explain the gap.

Other unresolved items: exact U.S. regional GPU inventory and queue times, the contractual SLA and interruption notice period, refund and reservation-cancellation rules, and a complete CUDA, driver, PyTorch, TensorFlow, and NCCL support matrix. Kimi reported only two U.S. regions (Missouri and New Jersey), while Gemini reported Kansas City and Woodbury ^[29].

Features for this workload

Questions explored

• Does Nebius support Kubernetes, Slurm, and InfiniBand for distributed AI training?
• Which NVIDIA GPU models does Nebius offer for training, and in which U.S. regions?

Nebius covers all five category criteria, with the strongest evidence on GPU availability and cluster architecture.

On orchestration, the H100 product material states customers can use Serverless Jobs, managed Slurm, Kubernetes, Ray/Anyscale, SkyPilot, Terraform, APIs, and CLI tooling ^[15]. Kimi reported that Nebius does not pre-install specific frameworks and that teams bring their own container images ^[29].

On networking, Gemini reported non-blocking NVIDIA Quantum-2 InfiniBand fabric with bare-metal-level GPU and network interface performance ^[28]. Grok reported NVIDIA Exemplar Cloud validation for training on H200/GB300 ^[11].

On geography, Gemini reported us-central1 (Kansas City) and us-north1 (Woodbury) ^[4]. Kimi reported Missouri and New Jersey as the two U.S. regions ^[29]. These two reports do not match, and buyers should confirm current U.S. region names directly.

Pricing, renewal terms, and ongoing costs

Questions explored

• How much does Nebius cost per GPU-hour for H100, H200, B200, and B300, and are there setup fees?
• What are Nebius spot pricing rules, storage fees, and commitment discount terms?

Nebius uses usage-based, per-second billing with an hourly pricing unit, and the reviewed sources do not establish a setup fee ^[3]. A first payment minimum of $25 is stated on the pricing page ^[16].

On-demand rates conflict across platforms. The table below shows every reported figure without resolving the conflict.

Sources: ^[16]. Grok attributed the higher figures to an October 1, 2026 price increase ^[30]. Buyers should treat the current Nebius pricing page as authoritative.

Spot pricing is dynamic. Openai reported listed starting spot prices of $0.79/GPU-hour for H100/H200 and $0.99 for B200/B300 ^[16]. Grok reported spot H100/H200 at roughly $2.15/$2.45 ^[25]. Kimi reported preemptible H100 at $2.15 and B300 at $4.30 ^[29]. Google reported spot H100/H200 from $0.79 and spot B300 from $4.30 ^[17]. Spot prices depend on capacity and demand and can change as frequently as every 15 minutes ^[10]. Customers can follow the spot price or set a maximum price ^[9].

Storage and other fees are billed separately from GPU usage ^[16]. Reported storage rates include shared filesystem at $0.0800/GiB-month, WEKA filesystem at $0.1000/GiB-month, object storage from $0.0147/GiB-month, and block storage from $0.0530/GiB-month ^[16]. Kimi reported storage at $0.0147–$0.1100 per GiB/month ^[29]. Google reported network SSD at roughly $7.10/month per 100 GB and object storage standard tier at roughly $19.42/month per 1 TB ^[24]. Prices exclude applicable taxes ^[16].

On egress, Grok and Kimi reported free VM egress ^[25]. Google reported that egress from VMs is free but egress from Object Storage is billed separately ^[24].

On contracts, on-demand and self-service access are advertised without a required long-term commitment ^[16]. Commitment discounts reach up to 35% for large-scale clusters reserved for multiple months ^[16]. Cancellation, refund, termination-notice, and reserved-capacity release terms are unclear from the reviewed public sources ^[16]. Kimi reported that GB200/GB300 NVL72 capacity requires direct sales contact with no published terms ^[29].

Best Suited For

Questions explored

• Who gets the most value from Nebius for AI training workloads?
• Is Nebius best for teams that already use Kubernetes or Slurm?

Nebius is best suited for AI startups and data science teams training or fine-tuning models on H100, H200, B200, or B300 GPUs, and for teams needing 8-GPU hosts or larger InfiniBand-connected clusters ^[18].

It also fits workloads that can tolerate preemptions in exchange for lower GPU costs, and buyers comfortable using Kubernetes, Slurm, Terraform, APIs, or cloud-console workflows ^[15].

Kimi added that Nebius suits teams with in-house Kubernetes or Slurm expertise seeking lower-cost alternatives to AWS, Azure, or GCP, and workloads needing EU data residency or compliance certifications such as SOC 2 Type II and ISO 27001 family ^[29]. Google noted teams that prefer managed Kubernetes or Slurm orchestration without extra platform fees ^[24].

Cost-sensitive teams with checkpoint-tolerant training pipelines are a strong match, because preemptible capacity roughly halves the effective GPU-hour cost ^[29].

Probably Not Best Suited For

Questions explored

• Who should not choose Nebius for GPU cloud compute for AI training?
• Is Nebius a bad fit for teams that need guaranteed uninterrupted training runs?

Nebius is probably not best suited for mission-critical training jobs that cannot tolerate interruption or capacity variability ^[10]. Preemptible VMs are intended for non-critical workloads, experimentation, data processing, and short fine-tuning jobs ^[10].

It is also a weaker fit for buyers requiring extensive independent customer-performance evidence, since the reviewed sources are primarily company-controlled and independent comparative training benchmarks were not established ^[16].

Teams needing a single standardized hyperscaler environment across many non-GPU services should look elsewhere, because Nebius does not offer a broad general-purpose cloud ecosystem ^[29]. Buyers who require fully published long-term reserved-capacity terms before contacting sales are also poorly matched, since large-scale reservation terms are handled partly through sales ^[16].

Kimi additionally reported that teams without container orchestration experience may struggle, that idle resources are easy to leave running and bill unexpectedly, and that independent field reports describe enterprise support as slower than AWS or CoreWeave ^[29]. These are platform-reported claims, not independently verified findings.

When Another Option May Be Better

Questions explored

• What is a better alternative to Nebius for a buyer who needs the lowest interruptible GPU price?
• When should a buyer choose a hyperscaler instead of Nebius for AI training?

Choose a major hyperscaler when the buyer needs broad regional redundancy, mature identity and networking integrations, or standardized enterprise procurement (ChatGPT). Hyperscalers also fit buyers whose data infrastructure, pipeline, and data gravity are locked into AWS, GCP, or Azure, where cross-cloud egress fees and migration engineering hours outweigh GPU compute savings (Gemini).

Choose a specialized GPU marketplace or cloud when the primary objective is the lowest interruptible price and the buyer can accept more variable capacity and platform maturity (ChatGPT). Kimi named CoreWeave spot at roughly 60% discount and Vast.ai for experimentation as alternatives for interruption-tolerant workloads ^[29].

Choose a provider with independently published benchmark results when model-training throughput and cost per completed training run matter more than headline GPU-hour pricing (ChatGPT). Buyers needing specialized non-NVIDIA silicon, such as Google TPUs or AMD Instinct GPUs, should also look elsewhere (Gemini).

Kimi noted that teams needing maximum simplicity, premium support responsiveness, or coast-to-coast U.S. latency optimization should evaluate CoreWeave or hyperscalers alongside Nebius ^[29].

Questions to ask before choosing a provider

Questions explored

• What should a buyer confirm with Nebius before signing a GPU training contract?
• Which Nebius capacity, SLA, and framework details must be verified before committing to large training runs?

The reviewed sources leave several items unresolved. Buyers should confirm each in writing before committing to large training runs.

Capacity and placement: which U.S. region has immediately usable H100, H200, B200, or B300 capacity, and what are the expected queue times (ChatGPT). Can the buyer reserve a specific number of GPUs with guaranteed placement and a defined start date (ChatGPT)? Which U.S. data centers offer which GPU SKUs today (DeepSeek)?

Spot and interruption rules: what are the exact preemptible interruption rules, notice period, restart behavior, and data-persistence guarantees (ChatGPT)? Are the published spot prices available to this account and region, and how frequently can they change (ChatGPT)? How often can preemptible instances be interrupted for the target GPU type and region (Perplexity)?

Framework and driver support: what CUDA, NVIDIA driver, NCCL, PyTorch, TensorFlow, and container versions are supported (ChatGPT)? Are 8-GPU hosts allocated as a single NVLink domain, and what topology and bandwidth are guaranteed across hosts (ChatGPT)? What pretrained ML frameworks or container images are available, or is bring-your-own-image mandatory (Kimi)?

Cost and contract: what are the storage, snapshot, filesystem, egress, public-IP, and support charges for the proposed architecture (ChatGPT)? What SLA, service credits, incident response times, and refund terms apply to on-demand and committed GPU capacity (ChatGPT)? What are the minimum commitment size, cancellation terms, and effective discount for a multi-month reservation (ChatGPT)? What are the specific commitment discount tiers, minimum terms, and any overage or underutilization penalties (Kimi)?

Performance evidence: can Nebius provide buyer-specific benchmark results for the intended model, batch size, precision, and distributed-training topology (ChatGPT)? Can Nebius provide benchmark data or customer references for multi-node distributed training at the required scale (Kimi)?

The consensus assessment

Nebius is a good fit for GPU cloud compute for AI training, with one platform rating it strong and five rating it good. It was named by four of six platforms during ranking discovery, with an average listed rank of 5.5 and a best rank of 4.

The consensus case rests on a technically complete training stack: current NVIDIA GPUs through the Blackwell generation, 8-GPU HGX hosts, InfiniBand networking, managed Kubernetes and Slurm, self-service access, and a preemptible tier that can roughly halve effective GPU-hour cost.

The consensus case is limited by three unresolved issues. First, published on-demand prices conflict across platforms, with H100 reported at both $3.85 and $4.50 per GPU-hour. Second, spot capacity is dynamic and interruptible, and the reviewed sources do not specify a universal interruption notice period or availability guarantee. Third, contract, SLA, refund, and framework-version details are incomplete in public sources, and independent comparative training benchmarks were not established.

The practical recommendation is to shortlist Nebius, run a paid pilot on the target GPU type and region, and obtain written confirmation of capacity, SLA, interruption rules, and framework versions before committing to large training runs.

How This Review Was Produced

This review evaluates Nebius only for GPU cloud compute for AI training. It is not a broad company review.

The study used a research prompt asking which providers would be recommended for a machine learning team frustrated by limited GPU availability and high prices on general-purpose cloud platforms, needing a range of GPU models, multi-GPU cluster support, framework and driver compatibility, and spot or preemptible pricing.

Six platforms were included in the fit-research stage. Four named Nebius during ranking discovery: Gemini, Grok, Kimi, and Perplexity. All six platforms evaluated Nebius for fit. Five returned a "good" rating and Gemini returned a "strong" rating.

The research date is 2026-10-03. Platform-reported research dates differ: DeepSeek reported 2026-01-15, while Gemini, Grok, Kimi, ChatGPT, and Perplexity reported 2026-10-03.

Research limitations

Platform-reported research dates differ from the authoritative run date and do not independently prove freshness. Deepseek's responses were dated roughly nine months earlier than the other platforms, which may explain its uncertainty about spot pricing.

Company-owned citations materially outnumber independent citations in the reviewed evidence. Company claims are not described as independently verified in this review.

The supplied URLs were collected from platform responses and were not independently validated by the writer stage. Citations are platform-reported evidence, not independently verified facts.

Platform mentions count only platforms that named Nebius during ranking discovery. All included platforms evaluated fit, but not all named Nebius in the ranking stage.

Conflicting product names, pricing, and capabilities were not resolved by guessing. Where platforms disagreed, this review describes the conflict and identifies what buyers should verify.

The reviewed sources do not independently verify actual U.S. GPU inventory, queue times, sustained training throughput, or failure rates. Claims about support quality, optimization, and benchmark-level performance are primarily company claims.

No personal testing, customer experience, or independent verification was performed for this review.

See the broader AI and Machine Learning Workloads consensus index for comparisons across qualified options.

Explore more vps cloud and dedicated server hosting guidance in the category directory.

Sources

Company-Owned Sources

1. Types of virtual machines and GPUs in Nebius AI Cloud: https://docs.nebius.com/compute/platforms
2. Compute pricing in Nebius AI Cloud: https://docs.nebius.com/compute/pricing
3. Compute pricing in Nebius AI Cloud: https://docs.nebius.com/compute/resources/pricing
4. Nebius AI Cloud regions: https://docs.nebius.com/overview/regions
5. Managing endpoints in Serverless AI - Nebius AI Cloud: https://docs.nebius.com/serverless/endpoints/manage
6. Unified billing for GPUs, vCPUs and RAM for GPU virtual: https://docs.nebius.com/signup-billing/billing-changes-oct-2025
7. Spot pricing and pricing policies in Nebius AI Cloud: https://docs.nebius.com/signup-billing/pricing-policy
8. Nebius Group: https://nebius.com/
9. Nebius AI Cloud: Q3 2026 updates: https://nebius.com/blog/posts/compute-cloud-platform-q3-2026
10. Introducing spot pricing for preemptible VMs: https://nebius.com/blog/posts/introducing-spot-pricing-for-preemptible-vms
11. Nebius achieves NVIDIA Exemplar Cloud on NVIDIA GB300 for training: https://nebius.com/blog/posts/nebius-achieves-nvidia-exemplar-cloud-on-nvidia-gb300-for-training
12. Introducing spot pricing for preemptible VMs: https://nebius.com/blog/spot-pricing
13. Nebius opens the first availability zone in the US: https://nebius.com/blog/us-availability-zone
14. Compute - GPU and CPU infrastructure for AI on Nebius: https://nebius.com/compute
15. NVIDIA HGX H100 on Nebius AI Cloud: https://nebius.com/h100
16. NVIDIA GPU Pricing | Nebius AI Cloud: https://nebius.com/prices
17. NVIDIA GPU Pricing | Nebius AI Cloud: https://nebius.com/pricing
18. Instant access to NVIDIA GPUs | Self-service AI clusters: https://nebius.com/self-service
19. Nebius AI Cloud services: https://nebius.com/services
20. Nebius AI Cloud: https://nebius.com/services/ai-cloud
21. Official pricing and terms source: https://docs.nebius.com/legal/terms-of-use

Independent Sources

22. Cloud GPU Pricing Comparison 2026 - Cantech: https://cantech.com/gpu-pricing-comparison-2026
23. GPU cloud pricing comparison: https://cloud-gpus.com/
24. Nebius: Review, Pricing & Alternatives: https://getdeploying.com/providers/nebius
25. Nebius GPU Pricing from $1.55/hr (2026) - 6 GPU models | GPU Finder: https://gpufinder.dev/providers/nebius
26. How to Choose a GPU Cloud Provider for AI Model Training | GPUSmith: https://gpusmith.com/articles/en/choose-gpu-cloud-provider-ai-training
27. ClusterMAX 3.0: The Industry Standard GPU Cloud Rating System Returns: https://newsletter.semianalysis.com/p/clustermax-30-the-industry-standard
28. Nebius vs AWS GPU: The Cost Allocation Problem: https://opslyft.com/blog/nebius-vs-aws-gpu
29. Nebius AI Cloud - Overview, Features & Use Cases | THE DAILY BRIEF: https://www.beri.net/tools/nebius-ai-cloud
30. Nebius price change - Nebius raises H100, H200, B200 and B300 GPU: https://www.usagepricing.com/blueprint/activity/nebius-ai-2026-09-24-gpu-price-increase