Nebius AI Cloud
by Nebius Group N.V.
European full-stack AI cloud with published GPU pricing and enterprise compliance
Nebius AI Cloud is a full-stack GPU cloud for AI teams, offering NVIDIA H100 through GB300 capacity with published per-hour pricing, managed Kubernetes and Slurm, and inference services. It suits engineering teams that want hyperscaler-grade compliance and EU data residency without hyperscaler pricing, and are comfortable operating their own orchestration.
Nebius AI Cloud is the compute platform of Nebius Group N.V., a Dutch company founded in 2022 and listed on NASDAQ as NBIS, that sells GPU capacity plus the managed services around it rather than bare metal alone. The fleet spans NVIDIA L40S, RTX PRO 6000, H100, H200, B200, B300 and GB200/GB300 NVL72 systems across six regions in Europe and North America — Finland, France, Iceland, the UK, Missouri and New Jersey — and unusually for this market the per-GPU-hour rates are published: H100 at $3.85 on-demand or $2.15 preemptible, B200 at $7.15/$3.95, B300 at $7.85/$4.30, L40S from $1.55/$0.74, with commitment discounts up to 35% and NVL72 racks quoted by sales. Managed Kubernetes, egress and ingress traffic and public IP addresses carry no charge; storage runs $0.0147–$0.1100 per GiB/month. Above the infrastructure sit AI Studio and Token Factory for model serving and TractoAI for data processing. Nebius AI Cloud 3.0 'Aether', announced in October 2025, added a secrets manager, granular IAM, searchable audit logs and quota management, and the platform carries SOC 2 Type II including HIPAA (audited by Deloitte), ISO 27001, 27701, 27018, 27799, 27032 and 22301, with NIS2 and DORA verification. Growth has been extreme: Q2 2026 revenue of $582M was up 454% year over year, ARR reached $3.0B against FY2026 guidance of $7–9B, and contracted backlog stood at $40B with customers including Cohere and Reflection.
The AI platform or infrastructure lead at a company training or serving models at scale who has real Kubernetes and Slurm capability in-house and wants EU-resident, audited capacity at roughly a third off hyperscaler rates.
Published, materially cheaper per-GPU-hour access to current NVIDIA silicon on a platform that already holds SOC 2 Type II, HIPAA and the ISO 27001 family.
At a Glance
- Category
- Infrastructure & Cloud
- Pricing
- Usage-based, Subscription, Contact for pricing
- Target Market
- CTOs, VPs of Engineering, ML Platform Engineers, Data Scientists
- Deployment
- Cloud-only, Multi-cloud
- Founded
- 2022
- Headquarters
- Amsterdam, Netherlands
- Team Size
- 500+
- Customers
- Not disclosed as a count; $40B contracted backlog as of Q2 2026, with named customers including Cohere and Reflection
Key Features
- ✓Current NVIDIA GPU fleet
L40S through GB300 NVL72, including H200, B200 and B300, so teams are not stuck a generation behind.
- ✓Published per-GPU-hour pricing
On-demand and preemptible rates are listed publicly, which is rare in this market and makes budgeting possible without a sales call.
- ✓Free managed Kubernetes and networking
Managed Kubernetes, ingress/egress traffic and public IP addresses carry no charge, removing the usual data-transfer tax.
- ✓AI Studio and Token Factory
Managed model serving and inference sit above the raw compute, so teams can deploy without building a serving stack.
- ✓Enterprise compliance stack
SOC 2 Type II including HIPAA audited by Deloitte, plus ISO 27001, 27701, 27018, 27799, 27032 and 22301, with NIS2 and DORA verification.
- ✓Aether governance controls
Secrets manager, granular IAM permissions, searchable audit logs and quota management shipped with AI Cloud 3.0 in October 2025.
- ✓European data residency
Regions in Finland, France, Iceland and the UK let regulated EU customers keep training data inside the bloc.
Capabilities
Use Cases
- •Frontier model training
Multi-node H200/B200 clusters with strong all-to-all connectivity support large distributed training runs at predictable cost.
- •Cost-controlled batch fine-tuning
Preemptible H100 capacity at $2.15/GPU-hour roughly halves the bill for interruptible fine-tuning and evaluation jobs.
- •Regulated EU AI deployment
ISO 27701 and EU regions let healthcare and financial customers run models without exporting personal data.
- •Managed inference serving
Token Factory and AI Studio host models for teams that want endpoints rather than to operate their own serving infrastructure.
- •Hyperscaler cost arbitrage
Teams with steady, forecastable GPU demand move workloads off AWS or Azure for roughly 30% savings on compute.
Ideal For
Best For
- ✓Large-scale model training and fine-tuning on current NVIDIA silicon at published, predictable hourly rates
- ✓Regulated European organisations that need EU data residency plus SOC 2 Type II and ISO 27001 evidence
- ✓Cost-sensitive AI teams with in-house Kubernetes or Slurm expertise who can operate their own orchestration
- ✓Batch and preemptible workloads where interruptible capacity roughly halves the effective GPU-hour cost
- ✓Teams that want managed Kubernetes and inference services rather than pure bare-metal GPU rental
Not Ideal For
- ✗Organisations that need a broad general-purpose cloud — managed databases, ETL and non-AI services must be sourced elsewhere, fragmenting the stack
- ✗Teams that prioritise premium, fast-response enterprise support, which independent field reports say is slower than AWS or CoreWeave
- ✗Beginners without Kubernetes or Slurm experience, who report difficulty tearing down resources and avoiding unexpected charges
- ✗Latency-critical inference shops needing dynamic GPU allocation, model compression and intelligent orchestration, where specialist inference platforms are further ahead
Integrations
Deployment
Market & Ratings
Not disclosed as a count; $40B contracted backlog as of Q2 2026, with named customers including Cohere and Reflection
Market Analysis
Pros
- ✓Transparent published pricing at roughly 30% below major hyperscalers, with preemptible tiers that roughly halve it again
- ✓Access to current NVIDIA silicon — H200, B200, B300 and NVL72 racks — with users reporting consistently strong, predictable GPU performance and good all-to-all connectivity
- ✓Unusually deep compliance for a challenger cloud: SOC 2 Type II with HIPAA audited by Deloitte, plus the ISO 27001/27701/27018/27799/22301 family and NIS2 and DORA verification
- ✓Free managed Kubernetes and zero egress charges remove two of the costs that quietly dominate hyperscaler GPU bills
- ✓Financial transparency as a listed company: Q2 2026 revenue $582M (+454% YoY), $3.0B ARR and $40B contracted backlog
Cons
- ✗Support response times are reported as slower than AWS or CoreWeave, which is a real problem for time-sensitive production workloads
- ✗Documentation thins out for advanced multi-node and edge-case deployments, limiting self-sufficiency on complex builds
- ✗Inference optimisation lags specialist platforms — no dynamic GPU allocation, model compression or intelligent orchestration — so latency-critical serving is a weaker fit
- ✗Narrow service catalogue: databases, ETL and other non-AI workloads must be run elsewhere, fragmenting the overall architecture
- ✗Newcomers report difficulty shutting resources down cleanly and incurring unexpected charges, and the platform assumes Kubernetes or Slurm competence
- ✗Very large capacity commitments are concentrated in a handful of contracts, so smaller customers compete for allocation against multi-billion-dollar deals
Pricing
On-demand GPU
From $1.55/hour
- ✓L40S from $1.55/GPU-hour, H100 $3.85, H200 $4.50
- ✓B200 $7.15/GPU-hour, B300 $7.85/GPU-hour
- ✓Free managed Kubernetes, ingress/egress and public IPs
Preemptible GPU
From $0.74/hour
- ✓L40S (AMD CPU) from $0.74/GPU-hour, H100 $2.15
- ✓H200 $2.45, B200 $3.95, B300 $4.30
- ✓Suited to interruptible batch and fine-tuning workloads
Committed capacity
Contact for pricing
- ✓Up to 35% below on-demand rates
- ✓GB200 and GB300 NVL72 racks quoted by sales
- ✓Reserved multi-node clusters
Metering is per GPU-hour and rates are fully published, which is unusual in this market: H100 is $3.85 on-demand and $2.15 preemptible, B300 $7.85/$4.30, CPU-only instances from $0.05/hour and storage $0.0147-$0.1100 per GiB/month. Commitment discounts reach 35%, and only GB200/GB300 NVL72 capacity is gated behind sales. There is no free tier, though managed Kubernetes and network traffic are not charged; field reports warn that idle resources are easy to leave running and bill unexpectedly.
Security & Compliance
Sources
This page was written from 6 sources, 3 on domains other than nebius.com.
- 1.nebius.com — pricesvendor
- 2.nebius.com — soc 2 type ii hipaa iso 27001 enterprise security standardsvendor
- 3.nebius.com — nebius introduces nebius ai cloud 3 0 aether delivering entevendor
- 4.getdeploying.com — nebius
- 5.truetheta.substack.com — nebius user experience a field report
- 6.finance.biggo.com — US NBIS 2026 08 12
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Emerald Conductor
Makes AI data centres power-flexible so utilities will connect them faster
Etched Sohu
Transformer-only inference ASIC shipped as rack-scale frontier inference clusters
fal
Serverless generative-media inference: 1,000+ image, video, audio and 3D models behind one API
Daytona
Sub-90ms stateful sandboxes that give every AI agent its own disposable computer