Dedicated GPU servers for enterprise, with a written SLA

Dedicated GPU servers for enterprise teams need more than GPUs: isolation you can explain to a security reviewer, commitments in writing, and paperwork that fits procurement. We rent whole NVIDIA and AMD GPUs by the month, from single-GPU servers to 8-GPU bare-metal nodes and GB300 NVL72 racks, with a 99.9% availability SLA, a delivery date confirmed before payment and invoices in your company's name.

Isolation and data handling

• Whole GPUs only. Every GPU is a physical GPU assigned to one customer: no vGPU slicing, no MIG partitions, no time-sharing. 8-GPU servers are bare-metal machines; single-GPU plans are dedicated virtual machines with the full GPU passed through.
• Access you control. Root access over SSH with keys, a public IPv4 address and an IPv6 /64 with DDoS mitigation, and private networking between your servers on request. You manage the firewall, patching and hardening of the operating system.
• Your data stays yours. You keep all rights to the data on your servers. We access it only when necessary to operate or repair the service at your request, or when required by law. Disks are securely erased between customers.
• Backups are yours to run. Local NVMe storage is not backed up by us unless agreed in writing, so plan replication to your own storage.
• Location in writing. The server's location is stated in the written confirmation before you pay, so you can check it against your data-residency rules.
• A clean exit. When you cancel, the server runs until the end of the paid month, then it is decommissioned and its disks are securely erased.

Standard images are Ubuntu 22.04 and 24.04 with NVIDIA drivers, CUDA, Docker and the NVIDIA Container Toolkit, ROCm images for AMD, and Debian or Rocky Linux on request. See also our trust and security page.

SLA, delivery and support commitments

Every server carries a 99.9% monthly availability commitment for power, network and hardware. If we fall short, you receive service credits on that server's monthly fee:

Credits are requested within 30 days after the affected month and applied to the next invoice, or refunded if the service ends; scheduled maintenance announced at least 48 hours in advance is excluded. Failed GPUs and components are replaced as a priority, and when a repair cannot be completed quickly we offer an equivalent server where available. The full text is in our service level agreement, and our help center covers how to request SLA credits.

Delivery is a firm commitment: we confirm the date in writing before you pay, and if we miss it you can cancel for a full refund. Support is available 24/7 from engineers, through our contact form.

Sizing dedicated GPU servers for enterprise workloads

• Private assistants and retrieval on a 70B model. FP8 weights take about 70 GB, so one H200 holds a replica with about 70 GB left for KV cache. An HGX H200 node runs eight such replicas, or larger models with tensor parallelism over NVLink.
• Large open mixture-of-experts models. A model with about 671 billion parameters needs roughly 700 GB in FP8: one HGX H200 node with 1.13 TB of HBM3e, or one HGX B200 node with 1.44 TB.
• Fine-tuning on internal data. QLoRA on 70B models fits one H100; full fine-tunes run on 8-GPU nodes.
• Training and frontier scale. HGX B200 and B300 nodes, or GB300 NVL72 racks: 72 GPUs in one NVLink domain with about 20 TB of HBM per rack, from $7,990 per GPU per month on 12–36 month terms. Vera Rubin NVL72 reservations are open. See rack-scale systems.

Procurement: invoices, terms and verification

Prices are in USD, billed monthly in advance, and exclude VAT and other taxes, which are added where applicable. Invoices carry your company details and VAT number, and the same invoice is issued whether you pay by card (Visa, Mastercard, American Express) or crypto; see invoices and VAT. Your price is fixed for each paid month, price changes for future months come with at least 30 days' notice, and material changes to our terms are notified by email at least 30 days before they apply.

Volume and term pricing for 6–12 months or several servers, usually from three, is available on request. For large orders, such as several 8-GPU servers or rack-scale systems, we may ask for company details and registration, the identity of the person ordering, proof of address and a short description of the intended use, as set out in our verification policy. We tell you exactly what is needed and why, and confirm the delivery date once verification is done. Export control and sanctions rules apply to every order.

For your legal review: under our Terms of Service and to the extent permitted by law, our total liability for a claim is limited to the amounts paid for the affected server during the three months before the event, and SLA credits are the primary remedy for unavailability. Read the full terms alongside the SLA before you sign off a fleet.

Start with a pilot

A practical path is a one-month pilot on one server with your real workload, then term pricing for the fleet. Send the configuration, quantity, term and any requirements through the contact form. We reply with a written confirmation of configuration, price, location and delivery date; payment follows by card or crypto, and credentials arrive by email. For multi-server deployments, ask about private networking at the same time.