Nebius
An Amsterdam-headquartered, Nasdaq-listed AI cloud renting NVIDIA GPUs by the second, aimed at teams that want to train and serve models without a sales call.
Who owns it
Nebius Group N.V. is a Dutch company headquartered in Amsterdam and listed on Nasdaq as NBIS. It is the former Yandex N.V., the Dutch parent of the Russian technology group. In February 2024 Yandex N.V. agreed to sell its Russia-based businesses to a consortium of Russian investors, and on 15 July 2024 it announced the second and final closing of that sale, divesting its remaining 28% stake for around 5.4 billion US dollars in cash and its own shares. Chairman John Boynton's statement on completion was that "All connections with Russia have now been severed". The shares were delisted from the Moscow Exchange on 10 July 2024, the renaming to Nebius Group N.V. was announced on 16 August 2024, and Nasdaq trading resumed on 21 October 2024. The company that emerged kept the non-Russian assets, principally the Finnish data centre and cloud platform that is now Nebius AI Cloud. Its founder and chief executive is Arkady Volozh, co-founder of Yandex, who was placed on the EU sanctions list in June 2022 and resigned from all Yandex positions, and who was removed from that list in March 2024. At the relaunch the company had 1,300 staff, of whom roughly 1,000 were engineers who came across from Yandex, about 500 of them in Amsterdam and the rest in Israel, Belgrade and Austin, most having left Russia after February 2022. NVIDIA joined a 700 million US dollar placement in December 2024, invested a further 2 billion US dollars alongside a strategic partnership in March 2026, and disclosed a 9.3% beneficial stake in July 2026. Microsoft signed a contract in September 2025 worth about 17.4 billion US dollars through 2031, rising to about 19.4 billion with options. Nebius designs and builds its own servers, running them in a mix of its own data centres and colocation space: its own description is of a data centre it owns in Finland plus "greenfield projects primarily in Europe plus additional colocation capacity". It does not resell another cloud and runs no marketplace of third-party hosts. The cloud operating entity is Nebius B.V. of Amsterdam, but which entity signs a UK customer's contract is not published.
## What it costs
Prices are in US dollars only, with no GBP or EUR list for a UK visitor, and no VAT basis is stated: the documentation says tax treatment varies by account type and residency. Read on 22 September 2026, one NVIDIA H100 NVLink (SXM) with 16 vCPU and 200 GB of RAM costs $3.85 an hour on demand, which is $2,811 a month at 730 hours, and the same GPU as a preemptible instance costs $2.15 an hour. Nebius has published that this rate rises to $4.50 an hour on 1 October 2026, a 17% increase, with H200 going from $4.50 to $5.40, B200 from $7.15 to $8.50 and B300 from $7.85 to $9.50. Committed pricing exists but is sales-only: the price page promises "up to 35% less than on-demand rates by reserving large-scale clusters for multiple months" and publishes no rate, no minimum term and no prepayment terms. GPUs are sold individually rather than in fixed nodes, so an eight-GPU H100 configuration works out at $30.80 an hour on demand, about $22,484 a month, but nothing forces you to buy eight. The cheapest GPU is an NVIDIA L40S PCIe from $1.55 an hour, $1,131.50 a month at 730 hours, or from $0.74 an hour preemptible. Nebius does not sell the GB200 NVL72; its Blackwell line is B200 and B300 NVLink.
Billing is per second against an hourly rate, so half an hour of an H100 costs half of $3.85. A stopped machine costs nothing in GPU, vCPU or RAM, which is unusual and useful. Storage is the catch: volumes are charged by their size whether they are attached to anything or not and however empty they are, at $0.071 per GiB a month for network SSD, $0.053 for non-replicated SSD, $0.080 for a shared filesystem and $0.0147 per GiB for standard object storage. Ingress and egress are free, with the single exception of object storage egress at $0.0150 per GiB. You can pay by card, by bank transfer or on invoice, on pay-as-you-go or with commitment discounts, and there are budget alerts and a CSV usage export. For the persona here the smallest realistic monthly bill is one L40S left running for a month plus a terabyte of network SSD: $1.55 x 730 hours = $1,131.50, plus 1,024 GiB x $0.071 = $72.70, so about $1,204 a month with no transfer charges on top. A shorter piece of work is cheaper still, because 100 hours on an H100 is $385. New accounts get $25 of Nebius AI Cloud credit and $25 of Token Factory credit on email verification, repeated about 30 days later if the account stays active, all expiring after 90 days, and approved members of Nebius for Startups get $5,000.
## What you get and what to watch
A new account can launch the reference GPU the same day. Signup is self-serve with a card and the default quota in the Finnish region is 12 GPUs plus 8 preemptible machines, 12 non-GPU machines with 200 vCPU, 32 disks, 2 to 5 TiB of network SSD and up to 5 InfiniBand GPU clusters. Nebius has published that self-service GPU quotas run to 32 GPUs without contacting support, with instant console access to L40S and H200. Quota increases are made on the Administration, Limits, Quotas page of the console, and only allocations beyond that go through a request form. Watch one detail: a stopped machine still consumes quota for its whole life, from creation to deletion, so stopping is not the same as freeing capacity.
The regions, from Nebius's own documentation, are eu-north1 and eu-north2 in northern Europe, eu-west1 in Paris, eu-south1 in southern Europe, eu-west2, uk-south1 and uk-south2 in the UK, us-central1 in Kansas City, us-north1, and me-west1 in Israel. H100 is in eu-north1 in Finland and nowhere else, so a UK buyer who specifically wants an H100 is served from Finland. The live UK regions carry B300 and RTX PRO 6000. The first UK deployment went live in November 2025 on NVIDIA Blackwell Ultra, and in June 2026 Nebius announced about 1.7 billion pounds of UK expansion across three deployments reaching 65 MW when fully ramped in 2027, with a London hub for commercial and AI research work. On the platform, you get multi-GPU clusters over InfiniBand, Managed Kubernetes as a first-class service, Slurm through Soperator (the company's open-source operator for running Slurm on Kubernetes), object and block storage, shared filesystems, virtual private clouds, encryption at rest and in transit by default, and multi-factor authentication. There is a documented API, an official Terraform provider, an official CLI and an official MCP server. The MCP server is worth reading before you use it: Nebius's own repository describes it as beta and warns that it allows execution of any Nebius CLI command, "which may lead to destructive and insecure consequences". There is no bare metal or dedicated server line.
Nebius holds SOC 2 Type II including HIPAA and ISO 27001:2022 certified by an independent external auditor, along with ISO 27701, 27018, 27032, 27799, 42001 and 22301, and CSA STAR Level 1, and it states that it acts as an EU data processor under GDPR and supports data residency requirements. It does not address the CLOUD Act anywhere, which is worth noting given a Nasdaq listing and large US operations. The SLA, effective 28 September 2026, commits to 99.50% uptime for Compute, with a credit of 10% of fees between 99.50% and 99.00%, 15% between 99.00% and 95.00% and 30% below 95.00%. Claims must be made within 14 days of the end of the reporting period, which is a short window, and credits only discount future fees. Support is free with no paid tier, through the console support centre or support@nebius.com, with first-response targets of 30 minutes for critical, 60 minutes for high, 4 hours for medium and one business day for low. There is no phone line. Public review coverage is almost non-existent: there is no Trustpilot profile on nebius.com at all, the profile on nebius.ai carries a 2.9 score from just two reviews, both one star and both about surprise billing after a trial with no clear way to cancel, and the profile is unclaimed so none of it is invited. G2 has a seller profile but not enough reviews to publish a score. No outage with customer data loss, regulatory action or lawsuit against Nebius was found on the public record.
Not yet tested by Saaskly. Assessed on published pricing, vendor documentation, company filings, the vendor's own status and trust pages, and third-party reviews.
Pros
- +Per-second billing with no GPU, vCPU or RAM charge on a stopped machine, and free ingress and egress apart from object storage egress at $0.0150 per GiB
- +A new card-paying account can launch GPUs the same day: the default quota is 12 GPUs in the Finnish region and Nebius has raised self-service access to 32 GPUs with instant console access to L40S and H200
- +SOC 2 Type II including HIPAA and ISO 27001:2022 audited externally, a published 99.50% Compute SLA with named credit tiers, and free support with a 30 minute first-response target on critical tickets
Cons
- −Storage volumes are billed by their size whether attached or not and however empty, at $0.071 per GiB a month for network SSD, and both of the only two Trustpilot reviews are complaints about being charged after a trial with no clear way to cancel
- −Committed pricing is sales-only: "up to 35% less" for multi-month cluster reservations is the entire published disclosure, with no rate card, no minimum term and no prepayment terms
- −Prices are US dollars only with no VAT basis stated for a UK buyer, and the published rates rise on 1 October 2026, taking the H100 from $3.85 to $4.50 an hour, a 17% increase
- −The live UK regions carry B300 and RTX PRO 6000 only, so an H100 is served from Finland, and SLA credits must be claimed within 14 days of the end of the reporting period
Pricing
Minimum commitment: $2,811 ($2,811/mo × 1 seat × 1 month (no commitment)). The least you can spend before walking away; excludes numbers, minutes and add-ons.
H100 NVLink 80 GB, one GPU
$2,811/mo · no commitment
Same price on any term
- ✓One NVIDIA H100 NVLink SXM 80 GB (16 vCPU / 200 GB RAM) at $3.85 an hour on demand, $2,811 a month at 730 hours; preemptible $2.15 an hour. Committed pricing is sales-only, quoted as up to 35% off for multi-month cluster reservations, with no rate, term or prepayment published. Nebius has published a rise to $4.50 an hour from 1 October 2026. USD, VAT not stated, read 22 Sep 2026.
L40S PCIe, one GPU
$1,132/mo · no commitment
Same price on any term
- ✓One NVIDIA L40S PCIe (8 to 40 vCPU / 32 to 160 GB RAM) from $1.55 an hour, $1,132 a month at 730 hours; preemptible from $0.74. Network SSD $0.071 per GiB a month billed attached or not; object storage $0.0147 per GiB a month; ingress and egress free except object storage egress at $0.0150 per GiB.
Specifications
| Rating | 2.6 |
| Pricing and plans | |
|---|---|
| Committed H100 | Sales only |
| Cheapest dedicated server | None |
| Billing | Per second |
| Egress | Included |
| Committed pricing | Sales only |
| Free credit | $25 Nebius AI Cloud credit plus $25 Token Factory credit on email verification, repeated after about 30 days, expiring after 90 days; $5,000 for approved Nebius for Startups members |
| Features | |
| H100 per GPU-hour* | $3.85/h (NVIDIA H100 NVLink (SXM) 80 GB) |
| Cheapest GPU | $1.55/h (NVIDIA L40S PCIe) |
| 8xH100 node | $30.80/h |
| Spot / interruptible | ✓ |
| Storage (per GB/mo) | $0.071 |
| Self-serve signup | ✓ |
| GPU quota | None |
| UK location | ✓ |
| H100 in the UK | No |
| EU locations | 5 |
| Locations | 10 |
| Origin | Spun out of Yandex N.V. in July 2024; founded and led by Yandex co-founder Arkady Volozh |
| Cluster networking | ✓ |
| Kubernetes | ✓ |
| Slurm | ✓ |
| Bare metal | No |
| Stopped instance billed | No |
| Sales-only above | Committed and reserved pricing, and GPU allocations beyond the self-service quota |
| UK cities | Not published; the live UK regions are named only as uk-south1 and uk-south2 |
| NVIDIA GPUs | ✓ |
| AMD GPUs | No |
| Multi-GPU nodes | ✓ |
| Dedicated servers | No |
| Object storage | ✓ |
| Block storage | ✓ |
| Private networking | ✓ |
| Integrations and developers | |
| API | ✓ |
| Terraform | ✓ |
| MCP server | ✓ |
| CLI | ✓ |
| Ownership and trust | |
| Hardware** | Own data centres |
| Owned in | EU-owned |
| Owner | Nebius Group N.V. |
| Uptime SLA (%) | 99.50 |
| Trustpilot (reviews) | 2.9 (2) |
| HQ | Netherlands |
| Publicly listed | ✓ |
| Sovereign offering | None |
| ISO 27001 | ✓ |
| SOC 2 | ✓ |
| 2FA | ✓ |
| Support | |
| Phone support | No |
| Support channels | Ticket, Email |
* One NVIDIA H100 80 GB per GPU-hour, on demand, in the currency the vendor shows a UK visitor; where the vendor sells no H100 the nearest GPU is named. Committed prices are the cheapest published rate and its minimum term.
** Who owns and runs the GPUs. "Third-party hosts" means the vendor is a marketplace for other people's machines and the SLA is theirs, not the vendor's.
Researched and drafted with AI assistance; reviewed and approved by Tim Meredith. How we use AI
Work at Nebius? Something not look right? Register in the vendor portal to see what we publish and send corrections.