Nebius, the Amsterdam-headquartered AI infrastructure company spun out of Yandex, builds and operates its own data centres in Finland, France, the US and elsewhere, and sells GPU capacity from single VMs to InfiniBand clusters of thousands of GPUs. It has a full platform around the compute: managed Kubernetes, Slurm-based clusters, object storage, and a per-token inference service.
packet.ai is smaller and lower. It does not sell H100, H200 or B300 yet, and its cluster product is quoted rather than self-served. What it has is a B200 at roughly half Nebius's on-demand price and a set of workstation and Ampere cards Nebius does not list at all.
Platform overview: packet.ai vs Nebius
What Nebius offers
Nebius's published on-demand rates in August 2026: HGX H100 $3.85/hr (preemptible $2.15), HGX H200 $4.50/hr (preemptible $2.45), HGX B200 $7.15/hr (preemptible $3.95), HGX B300 $7.85/hr (preemptible $4.30). Committed-use discounts run to 35%. Egress is free; standard object-storage egress is about $0.015/GiB. Earlier in 2026 the B200 was listed at $5.50/hr, so watch the trend.
- Own data centres in Finland, France, Iceland, the US and more; strong European data-residency story
- Managed Kubernetes, Slurm clusters, InfiniBand fabric, shared filesystem and object storage
- Preemptible tier at roughly 55% of on-demand across all cards
- No consumer or workstation GPUs; the rate card starts at the H100
What packet.ai delivers
Nebius's rate card starts at the H100. There is no A100, no L40S, no RTX 6000 Pro, no RTX 4090, nothing below datacentre-class Hopper and Blackwell silicon. If your workload fits an entry or mid-tier card, Nebius simply does not sell one. packet.ai's range runs from RTX 4090 at $0.39/hr up through B200, so most workloads that do not need an H100-class card land on a cheaper tier by default.
On the one card both platforms sell, B200, packet.ai is cheaper even against Nebius's preemptible rate. B200 starts at $3.75/hr Dynamic, non-interruptible, on packet.ai. Nebius's preemptible B200 is $3.95/hr, which can still be reclaimed at any time, and its on-demand B200 is $7.15/hr, a 48% gap against packet.ai's Dynamic rate.
- Dynamic: scheduler-enforced multi-tenant, never pre-empted. B200 starting at $3.75/hr, cheaper than Nebius's interruptible tier.
- Dedicated: a whole card, single-tenant, 99% SLA. RTX 4090 at $0.39/hr, L40S at $0.92/hr, A100 80GB at $1.43/hr, B200 at $6.99/hr, none of which have a Nebius equivalent below H100.
What Nebius has that packet.ai does not: H100, H200 and B300 at scale with InfiniBand, managed Kubernetes and Slurm clusters, and committed-use discounts to 35% for teams doing sustained large-scale training. If your workload genuinely needs Hopper-class or newer silicon at cluster scale with European data residency, Nebius is built for it. If your workload fits anything from RTX 4090 up, packet.ai gets you there for less, without a preemption risk on the cheap tier.
Billing is hourly with monthly commits up to 20% off. There are no platform fees and no ingress charges; egress is $0.04/GB. Up to 2 TB of local NVMe per node is included. Dynamic instances are SSH-ready in under five minutes, Dedicated in five to ten. Capacity is live in the US (California, Virginia, Texas, Oregon) and Europe (Frankfurt, Amsterdam, Paris, London, Dublin). H100 SXM, H200 and RTX 5090 are on the notify list.
Against Nebius, packet.ai's B200 is the comparison that matters: $3.75/hr Dynamic against $7.15/hr on-demand, and even Nebius's preemptible $3.95/hr is above packet.ai's non-interruptible rate. Below the B200, Nebius has nothing to compare against packet.ai's RTX 6000 Pro, L40S, A100 and RTX 4090. Above it, Nebius has Hopper and B300 that packet.ai does not list yet.
packet.ai vs Nebius at a glance
Published starting rates in USD per GPU-hour, on-demand unless noted. Percentages compare the first price in each cell.
| Category | packet.aiYOU | Nebius |
|---|---|---|
| B200 (180–192GB) | $3.75/hr Dynamic · $6.99/hr Dedicated−48% | $7.15/hr on-demand · $3.95/hr preemptible |
| A100 80GB | $1.43/hr Dedicated | Not on current rate card |
| RTX 6000 Pro (96GB) | $0.66/hr Dynamic | Not offered |
| L40S (48GB) | $0.92/hr Dedicated | Not offered |
| RTX 4090 (24GB) | $0.39/hr Dedicated | Not offered |
| H100 / H200 / B300 | H100 and H200 launching soon (notify list) | H100 $3.85/hr · H200 $4.50/hr · B300 $7.85/hr |
| Interruptible tier | None; Dynamic is shared but not pre-empted | Preemptible at ~55% of on-demand |
| Commitment discounts | Monthly up to 20% off; clusters ~30% below retail on term | Committed-use up to 35% |
| Multi-node | InfiniBand clusters to 1,024+ GPUs, quoted, 2–6 weeks | Self-serve and managed Slurm/K8s clusters, InfiniBand, thousands of GPUs |
| Data transfer | No ingress fee; egress $0.04/GB | Free egress; object storage egress ~$0.015/GiB |
| Regions | US (4 sites) + EU (5 sites) | Finland, France, Iceland, US and others |
Detailed comparison
The one card they share
Both platforms sell the B200 on demand. Nebius asks $7.15/hr; packet.ai asks $3.75/hr on Dynamic and $6.99/hr on Dedicated. Dynamic is 48% cheaper, and Dedicated, a single-tenant card with a 99% SLA, is still two percent under Nebius's list. Nebius's preemptible B200 at $3.95/hr gets close to packet.ai's Dynamic rate, but a preemptible instance can be reclaimed; a Dynamic instance cannot.
If you are running a B200 for a training job that checkpoints well, Nebius preemptible is competitive. If you are serving inference, packet.ai Dynamic is the cheaper hour and it stays up.
Where the rate cards do not overlap
Nebius does not sell anything below the H100. For a team whose model fits in 96GB, packet.ai's RTX 6000 Pro at $0.66/hr is a fifth of Nebius's cheapest hour. The same goes for the L40S, A100 80GB and RTX 4090; they simply are not options on Nebius.
The reverse is also true. If you need H100 or H200 today, or want early access to B300, Nebius has published rates and packet.ai has a notify list. That is the honest shape of this comparison: little overlap, and a clear winner on each side of it.
Platform depth and scale
Nebius operates as a full AI cloud: managed Kubernetes with GPU operators, Slurm-based training clusters, a shared filesystem, object storage, and a per-token inference API. It signs capacity deals in the thousands of GPUs and has the balance sheet to build data centres.
packet.ai's cluster product reaches 1,024+ GPUs on InfiniBand at wholesale pricing around 30% below retail, but it is quoted and delivered over two to six weeks rather than self-served. For a 64-GPU fine-tuning cluster on a 12-month term, packet.ai's $2.80–3.20/GPU-hr B200 guidance is competitive; for a 4,000-GPU pre-training run, Nebius is the more proven counterparty.
Residency and regions
Both are strong in Europe. Nebius has Finland, France and Iceland with a well-documented sovereignty story; packet.ai has Frankfurt, Amsterdam, Paris, London and Dublin, plus four US sites. If your requirement is specifically Nordic, Nebius has more there; if you need a UK or Irish region, packet.ai has them and Nebius does not list them.
Who should choose which
Choose packet.ai if
- You want a B200 at the lowest non-preemptible on-demand rate with no cluster minimum
- Your model fits an RTX 6000 Pro, L40S or A100 - cards Nebius does not sell
- You need 1 to 64 GPUs hourly without a contract or enterprise procurement process
- You need UK or Irish region capacity
Choose Nebius if
- You need H100, H200 or B300 on-demand today
- You are running a checkpointed training job and can exploit preemptible pricing
- You want managed Slurm and Kubernetes with a deep platform layer behind it
- You are procuring thousands of GPUs and need a builder-operator with its own data centres
Conclusion
Nebius and packet.ai share one card, the B200, and packet.ai is 48% cheaper on it without interruption risk. Everything else on the two rate cards is exclusive to one side: Nebius owns Hopper and B300 today, packet.ai owns everything below the H100. Pick by the card you need; where they overlap, pick packet.ai.
Same silicon. Smarter economics.
Every rate on this page is a published starting price you can deploy against today. Dynamic launches in under five minutes; no credit card to start.
Additional resources
- packet.ai GPU pricing: the full Dynamic, Dedicated and Clusters rate card.
- packet.ai documentation: API, CLI and platform guides.
- How intelligent scheduling works: why Dynamic keeps peak performance at a lower price.
- Clusters and wholesale quotes: multi-node InfiniBand capacity at ~30% below retail.
- Rent GPUs on packet.ai: specs, pricing and workload fit for every card.