Home/Dedicated Servers
Bare metal & dedicated GPU serversSingle-tenant GPU servers, published prices, no hypervisor tax
Lease a whole machine — every GPU, every core, every NVMe lane, root and BMC access included. Monthly, annual or hourly, across 25 markets, with hardware from Dell, Supermicro, HPE, Lenovo, GIGABYTE and QCT.
Would you rather own the hardware? Buy from any OEM and let us run it — see Bring Your Own GPU. Renting and owning use the same halls, the same fabric and the same engineers.
NVIDIA HGX & NVL72 dedicated servers
Eight-GPU SXM nodes with NVLink and NVSwitch, dual-socket CPUs, 400G/800G fabric NICs and local NVMe scratch. Sold as whole nodes and as multi-node pods with InfiniBand between them.
| Configuration | CPU / RAM | Local NVMe | Fabric NIC | Monthly | Annual (per mo) | Hourly |
|---|---|---|---|---|---|---|
| GB300 NVL72 — 72 GPU | 36× Grace · 40 TB | 4× 7.68 TB | 18× XDR 800G | Quoted | Quoted | — |
| 8× B300 SXM 288GB | 2× Xeon 6 / EPYC 9005 · 3 TB | 8× 7.68 TB Gen5 | 8× XDR 800G | $44,800 | $39,900 | $66.72 |
| 8× B200 SXM 180GB | 2× Xeon 6 / EPYC 9005 · 3 TB | 8× 7.68 TB Gen5 | 8× NDR/XDR | $38,000 | $33,800 | $52.88 |
| 8× H200 SXM 141GB | 2× Xeon 8568Y+ / EPYC 9554 · 2 TB | 8× 3.84 TB Gen4/5 | 8× NDR 400G | $23,900 | $21,300 | $30.32 |
| 8× H100 SXM 80GB | 2× Xeon 8468 / EPYC 9474F · 2 TB | 8× 3.84 TB Gen4 | 8× NDR 400G | $18,400 | $16,400 | $24.80 |
| 8× A100 SXM 80GB | 2× EPYC 7763 / Xeon 8358 · 1 TB | 4× 3.84 TB Gen4 | 8× HDR 200G | $10,400 | $9,300 | $15.20 |
| 8× A100 SXM 40GB | 2× EPYC 7742 · 1 TB | 4× 1.92 TB Gen4 | 8× HDR 200G | $8,000 | $7,100 | $11.76 |
Indicative list pricing in USD per node, excluding tax, parallel storage and cross-region transit. Multi-node pods include InfiniBand leaf/spine, cabling and validation.
The configurations the market actually rents
Inference, fine-tuning, rendering and CI fleets rarely need an NVL72. These are the workhorse builds, with the same single-tenant guarantee.
| Configuration | GPU memory | Platform | Typical workload | Monthly | Hourly |
|---|---|---|---|---|---|
| 8× H100 NVL 94GB | 752 GB | 2× EPYC 9454 · 1.5 TB · PCIe Gen5 | High-density LLM inference | $14,700 | $21.60 |
| 4× H100 PCIe 80GB | 320 GB | 2× Xeon 8462Y+ · 768 GB | Fine-tuning, mid-size serving | $7,300 | $10.80 |
| 8× RTX PRO 6000 Blackwell 96GB | 768 GB | 2× EPYC 9455 · 1.5 TB · Gen5 | Inference, Omniverse, simulation | $9,500 | $17.44 |
| 8× L40S 48GB | 384 GB | 2× Xeon 8462Y+ · 1 TB | Inference, video, VDI, rendering | $6,000 | $10.80 |
| 4× L40S 48GB | 192 GB | 2× EPYC 9354 · 512 GB | Small-model serving, batch | $3,300 | $5.40 |
| 8× RTX 5090 32GB | 256 GB | 2× EPYC 9354 · 768 GB | Research, diffusion, dev clusters | $3,800 | $6.00 |
| 8× RTX 4090 24GB | 192 GB | 2× EPYC 7543 · 512 GB | Diffusion, CI, cost-sensitive inference | $2,700 | $5.04 |
| 4× RTX A6000 48GB | 192 GB | 2× Xeon 6338 · 512 GB | Workstation-class, CAD, legacy CUDA | $1,900 | $5.40 |
| 8× L4 24GB | 192 GB | 2× EPYC 9124 · 384 GB | Low-power inference, transcoding | $2,200 | $3.12 |
| 2× L40S 48GB | 96 GB | 1× EPYC 9254 · 256 GB | Entry inference node | $1,640 | $2.70 |
The rest of the estate
Data pipelines, tokenisation, head nodes, vector databases and checkpoint stores — all on the same network, in the same hall as your GPUs.
| Configuration | Cores / threads | Memory | Storage | Network | Monthly |
|---|---|---|---|---|---|
| Compute — Dual EPYC 9654 | 192C / 384T | 768 GB DDR5 | 2× 3.84 TB NVMe | 2× 25G | $1,490 |
| Compute — Dual Xeon 6740E | 192C / 192T | 512 GB DDR5 | 2× 1.92 TB NVMe | 2× 25G | $1,210 |
| General — Single EPYC 9354 | 32C / 64T | 256 GB DDR5 | 2× 1.92 TB NVMe | 2× 10G | $680 |
| Head / login node | 16C / 32T | 128 GB | 2× 960 GB NVMe | 2× 10G | $370 |
| Storage — NVMe dense | 64C / 128T | 512 GB | 24× 7.68 TB NVMe (184 TB) | 2× 100G | $3,800 |
| Storage — capacity HDD | 32C / 64T | 256 GB | 60× 22 TB SAS (1.3 PB raw) | 2× 25G | $4,500 |
| Memory-optimised | 96C / 192T | 3 TB DDR5 | 4× 3.84 TB NVMe | 2× 100G | $2,800 |
Priced above the median, deliberately
Velyrix list pricing is set at 15% above the market median — deliberately, and we publish the arithmetic. We are not trying to be the cheapest GPU-hour on the internet, because the cheapest GPU-hour is usually oversubscribed, blocking, unsupported, or all three. The premium buys engineering: non-blocking fabric, a signed acceptance report, four-hour hardware replacement and engineers who answer at 3 a.m.
| Rental unit | Market range | Market median | Velyrix list | vs median | What the premium buys |
|---|---|---|---|---|---|
| H100 SXM — per GPU-hour | $1.90 – $3.50 | $2.70 | $3.10 | +15% | 1:1 non-blocking NDR fabric, single tenancy |
| H200 SXM — per GPU-hour | $2.40 – $4.20 | $3.30 | $3.79 | +15% | Guaranteed capacity, no oversubscription |
| B200 — per GPU-hour | $4.00 – $7.50 | $5.75 | $6.61 | +15% | Liquid-cooled halls, sustained clocks under load |
| B300 / GB300 — per GPU-hour | $5.50 – $9.00 | $7.25 | $8.34 | +15% | Rack-scale commissioning and CDU redundancy |
| A100 80GB — per GPU-hour | $1.10 – $2.20 | $1.65 | $1.90 | +15% | Maintained fleet, burn-in before every handover |
| L40S — per GPU-hour | $0.75 – $1.60 | $1.18 | $1.35 | +15% | Enterprise SLA rather than best effort |
| RTX 4090 — per GPU-hour | $0.30 – $0.80 | $0.55 | $0.63 | +15% | Tier III-class facility, not a colocated shelf |
| 8× H100 SXM 80GB node — per month | $10,000 – $22,000 | $16,000 | $18,400 | +15% | 24×7 NOC, 4-hour replacement, acceptance report |
| 8× H200 SXM 141GB node — per month | $13,500 – $28,000 | $20,750 | $23,900 | +15% | Named architect, monthly SLA reporting |
| 8× B200 SXM 180GB node — per month | $21,000 – $45,000 | $33,000 | $38,000 | +15% | Direct-to-chip cooling, validated fabric |
Market ranges and medians are Velyrix's own observation of publicly advertised list pricing across GPU cloud and bare-metal providers, provided for orientation only; they are not a representation about any third party's current pricing, and no comparison is endorsed by any other provider. Velyrix rates are indicative list prices confirmed in a written quotation. Committed multi-year terms and large volumes are quoted individually.
What ships with every dedicated server
- Root and BMC access: IPMI/Redfish, virtual media, serial-over-LAN and remote reboot
- OS of your choice: Ubuntu, Rocky, RHEL, Debian, Windows Server or a custom image via iPXE
- Driver stack pre-validated: NVIDIA driver, CUDA, Fabric Manager, DCGM, NVIDIA Container Toolkit
- Burn-in before handover: 24–72 hour GPU stress, ECC and thermal validation report
- Bandwidth: 10G unmetered or 25/100G metered, dual-homed, with DDoS mitigation
- Remote hands: 24×7 facility engineers on site, four-hour hardware replacement target
| Add-on | Unit | Price |
|---|---|---|
| Parallel FS (WEKA/VAST) | per TB / month | $42 |
| NVMe block storage | per TB / month | $100 |
| Object storage (S3 API) | per TB / month | $20 |
| Backup & snapshot vault | per TB / month | $14 |
| 100G unmetered uplink | per port / month | $1,600 |
| Additional IPv4 /29 | per month | $46 |
| Private cloud interconnect | per 10G / month | $545 |
| InfiniBand pod networking | per node / month | $805 |
| Managed OS & patching | per server / month | $230 |
| Managed Slurm / Kubernetes | per cluster / month | $2,400 |
Own the GPUs. Let us run them.
Buy your NVIDIA servers from any OEM or distributor you like, ship them to a Velyrix hall, and we handle the rest — deployment, fabric, cooling, monitoring and support. Or rent ours. Either way, you get a plan in one business day.
Frequently asked questions
What is the difference between a dedicated GPU server and a GPU cloud instance?
A dedicated server is a whole physical machine leased to one customer: all GPUs, all CPU cores, all NVMe and the BMC belong to you for the term, with no hypervisor overhead and no shared tenancy. A cloud instance is provisioned per second and may be virtualised. Dedicated servers are usually 25-45 percent cheaper per GPU-hour on terms of three months or more.
How much does it cost to rent an 8x H100 server per month?
Velyrix lists an 8x NVIDIA H100 SXM node from $18,400 per month, falling to $16,400 on a twelve-month term and $15,600 at thirty-six months. The list price is deliberately 15 percent above the market median of roughly $16,000: it reflects non-blocking fabric, a signed acceptance report, 24x7 NOC coverage and a four-hour hardware replacement target rather than best-effort hosting.
Can I rent a single GPU rather than a whole node?
Yes, through Velyrix GPU Cloud, which offers 1, 2, 4 and 8 GPU instances billed per second. Dedicated servers are sold as whole nodes because the NVLink and NVSwitch topology inside an HGX baseboard cannot be meaningfully divided between tenants.
How quickly is a dedicated server delivered?
Configurations held in stock in an active region are typically live within two to twenty-four hours. Custom builds, large pods and Blackwell-class nodes depend on OEM lead times and are normally delivered within five to twenty business days.
Do you charge for bandwidth or egress?
Standard plans include a 10G unmetered uplink with no egress charges. Higher-capacity 25G and 100G ports, private cloud interconnects and dedicated transit are priced as add-ons.
Can I upgrade or swap hardware mid-term?
Yes. Within a term you can move to a higher tier at the difference in rate, keeping your data on the same storage estate. Downgrades are applied at the next renewal date.