GPU compute without the cloud markup.
NVIDIA H200 NVL on dedicated hardware.
velia.net delivers GPU compute the way modern AI workloads actually want it: no preemption, no per-token egress, no 12-month commit. Just metal sized for the workload, on a monthly contract.
5,000+
Companies
40,000+
Servers
4.2 Tbit/s
Backbone
22 years
In hosting
It is a contract shortage. And a markup problem. Cloud GPU pricing rewards the wrong commitments.
The list price is a fiction.
The real price requires a multi-year reservation. The reservation locks you to a region, a SKU, and a generation. By the time the contract ends, the GPU is two generations behind the frontier.
Preemption breaks training runs.
Spot pricing looks cheap on a chart. It looks expensive at 3 a.m. when a run dies 89 percent through. Production AI does not run on best-effort scheduling.
Per-token egress eats inference economics.
Inference at scale is a network workload. Hyperscale cloud egress turns every served token into a margin compression event. The fix is to own the egress contract, not just the model.
The AI numbers behind the move to dedicated GPUs.
| Metric | Hyperscale cloud (typical) | velia bare metal |
|---|---|---|
| H200 NVL per-hour rate | Variable, often surge-priced | Flat monthly |
| Setup time | Hours to days, queue dependent | 1 to 4 hours |
| Preemption risk | Real on spot tiers | None |
| Egress per TB | Charged separately, often hidden | Included up to commit |
| Contract minimum for committed pricing | 12 to 36 months | 1 month |
| GPU memory per node | Capped by SKU | Up to 8x H200 NVL = 1.13 TB HBM3e |
A single 8x H200 NVL node on velia.net delivers 1.13 TB of HBM3e and 38.4 TB/s of aggregate memory bandwidth at a fraction of the cost of equivalent hyperscale reservations.
Built for the frontier.
NVIDIA H200 NVL
1x to 8x · Training, fine-tuning, large-context inference.
- 141 GB HBM3e per GPU
- 4.8 TB/s memory bandwidth per GPU
- 900 GB/s NVLink in 4x and 8x configs
- Up to 2 TB DDR5 system RAM
- NVMe Gen5 sized for dataset and checkpoints
141 GB HBM3e per GPUAMD EPYC 9555P
Dense GPU host CPU.
- 64 cores / 128 threads
- Pairs cleanly with multi-GPU configs
AMD EPYC 9575F
GPU host for latency-sensitive serving paths.
- 64 cores / 128 threads
- Zen 5 high-frequency
Six locations. One backbone.
- →4.2 Tbit/s multi-homed backbone
- →Always-on DDoS protection at the network edge
- →Frankfurt peering at DE-CIX for European inference
- →St. Louis and New York for US workloads
- →Singapore for Asia
Frankfurt
DE-CIX peering · European inference
St. Louis
US workloads
New York
US East Coast
Singapore
Asia-Pacific
4.2 Tbit/s
Multi-homed backbone. Always-on DDoS protection.
What teams run on velia GPUs.
Inference at scale.
Production serving without per-token egress. H200 NVL nodes, flat monthly, no surge pricing.
Training runs.
Reserved GPUs for the length of the run, monthly contract. No preemption. No queue wait.
Fine-tuning shops.
4x or 8x H200 NVL nodes on demand. Stand up for a quarter, stand down without a 12-month exit clause.
Research and evaluation.
Mixed workloads, multi-tenant teams. Predictable monthly cost.
The hardware behind the workloads.
Migrated production inference from a hyperscale GPU offering to 4x H200 NVL nodes in Frankfurt. Cost per million tokens dropped 71 percent.
Devi R.
CTO, AI Fine-Tuning Lab
Ran a six-week multi-model fine-tuning campaign on 8x H200 NVL. No preemption, no surge pricing, no 12-month commit. Stood the cluster down at the end.
Research Lead
ML Research Group
We were paying for vCPUs we never used and queuing for GPUs we needed yesterday. velia.net turned that around in one quarter.
Head of Infra
Production Inference Team
Egress on inference is the silent killer. With velia.net, it stopped being a line item.
VP Engineering
Generative AI Platform
| Hyperscale cloud GPU | Niche GPU clouds | velia bare metal | |
|---|---|---|---|
| Pricing model | Hourly with reservations | Hourly | Flat monthly |
| Preemption risk | Real on spot | Provider-dependent | None |
| Contract for committed price | 12 to 36 months | Often shorter | 1 month |
| Egress economics | Often hidden | Often hidden | Included up to commit |
| Hardware control | None | Limited | Full |
| Region count | High but expensive | Often 1 to 2 | 6 locations |
From order to running workload.
Configure on the shop.
Pick your GPU count (1, 2, 4, or 8) on H200 NVL, host CPU, RAM, and NVMe size.
Server live within 1 to 4 hours.
Welcome email contains dashboard credentials, DDoS profile, and remote management access.
Deploy your stack.
Bring your own image. CUDA, ROCm, vLLM, TensorRT, any framework. Full root.
Results from teams already on velia.net
“Cost per million tokens dropped 71%”
Migrated production inference from a hyperscale GPU offering to 4x H200 NVL nodes in Frankfurt. Throughput per node went up because the team could finally co-locate the model and the cache without paying egress for the round-trip.
“No preemption across six weeks”
Ran a six-week multi-model fine-tuning campaign on 8x H200 NVL. No preemption, no surge pricing, no 12-month commit. Stood the cluster down at the end of the campaign.
Hardware is live within 1 to 4 hours.
Monthly. No annual lock-in.
H200 NVL. Talk to the team for other GPU availability.
Yes. 1, 2, 4, or 8 GPU nodes with NVLink in 4x and 8x.
You pick the data center at order time.
Always-on at the network edge.
Yes. Bare metal means any OS, any driver stack.
Monthly contract. Cancel any time. No exit fees.
AI is a hardware problem.
We solved the hardware.
Real talk from 22 years in hosting: performance starts at the hardware layer. Everything else is overhead.
velia.net // H200 NVL // since 2003