Nebius Startup Deals & Discounts

Nebius runs an application-based startup program for AI-native companies that need serious GPU capacity but want to manage early-stage cost. The program is organized into stage-based tiers spanning bootstrapped teams through pre-Series B companies, so the support matches where a startup actually is rather than forcing everyone into one bucket.

Approved startups receive an initial allocation of cloud credits upon account activation to begin training and experimentation, and can unlock substantially larger matched discounts by committing to a multi-month reserved contract. Beyond credits, accepted teams gain access to technical and solution-architecture support and priority access to current-generation NVIDIA infrastructure, which matters when high-demand GPUs are scarce.

This is a good fit for teams pretraining or fine-tuning models, building production inference, or running compute-heavy AI research who would otherwise burn runway on raw infrastructure. Eligibility is assessed by company stage and use case through a single streamlined application, with decisions typically returned within a short review window. As with any provider program, terms and credit amounts are set by Nebius and can change over time.

Training foundation models - Teams pretraining LLMs or large multimodal models use multi-node H100, H200, or Blackwell clusters over InfiniBand, where interconnect bandwidth determines whether distributed training scales or stalls.

Fine-tuning open-source models - Startups adapting Llama, Qwen, DeepSeek, or Mistral checkpoints to proprietary data run fine-tuning jobs on reserved GPU capacity, then deploy the result through Token Factory dedicated endpoints.

Production inference at scale - Applications serving hundreds of millions of tokens per minute rely on the OpenAI-compatible Token Factory API with a 99.9% SLA and autoscaling throughput, avoiding the work of building and operating their own serving stack.

Regulated and privacy-sensitive AI - Healthcare, finance, and EU-data-residency workloads use the SOC 2 Type II / HIPAA / ISO 27001 regions and zero-retention inference to keep sensitive requests inside a chosen jurisdiction and out of training pipelines.

Computer vision and agentic systems - Research and product teams building vision models or agentic AI tap Blackwell B200/B300 systems for the memory and throughput that reasoning and multimodal workloads demand.

Nebius prices GPU compute per GPU-hour with two self-serve tiers, preemptible and on-demand, plus reserved-capacity discounts:

GPU	Preemptible / hr	On-demand / hr
H100	$2.15	$3.85
H200	$2.45	$4.50
B200	$3.95	$7.15
B300	$4.30	$7.85
GB200 NVL72	Contact sales	Contact sales
GB300 NVL72	Contact sales	Contact sales

Storage (monthly): shared filesystem ~$0.08/GiB, WEKA ~$0.10/GiB, standard object storage ~$0.0147/GiB, block volumes ~$0.053-$0.118/GiB depending on replication.

Networking and orchestration: ingress/egress and Kubernetes/Slurm cluster management are provided at no additional charge; you pay for compute and storage consumption.

Commitments: reserving large-scale clusters for a multi-month term earns up to 35% off on-demand rates. Figures reflect publicly listed rates and can change; confirm current pricing with Nebius.