For AI teams · Rent GPUs

Dedicated GPU servers. Bare metal. Live in days.

Reserve a whole multi-GPU server — no virtualization, no noisy neighbors, no capex. RTX PRO 6000 Blackwell from $1.50/GPU-hour, with root access and hardware support included.

Single tenant: the whole box is yours for the term
Flat monthly price — no egress games, no per-second metering
Month-to-month, 6-month, or 12-month terms

See Pricing Reserve Capacity

$0.00/hrPer GPU-hour, 12-month

0 GBVRAM per N8 server

0Tenant per server

0Term options

Pricing

One server. One price. Pick your term.

Commit longer, pay less per hour. Every plan is the whole machine.

Nexus N4

Inference and fine-tuning

$4,380/mo

$1.50 per GPU-hour · 4 GPUs

32% under the $2.20 managed-cloud median

GPUs4× RTX PRO 6000
VRAM384 GB
CPUDual EPYC 9005
RAM512 GB DDR5
NVMe8 TB
Network2× 25GbE

Single tenant, bare metal
Root + IPMI access
Hardware swaps included

Reserve the N4

Most requested

Nexus N8

Multi-tenant serving, 100B+ models

$8,760/mo

$1.50 per GPU-hour · 8 GPUs

32% under the $2.20 managed-cloud median

GPUs8× RTX PRO 6000
VRAM768 GB
CPUDual EPYC 9555
RAM768 GB DDR5
NVMe16 TB
Network2× 100GbE

Single tenant, bare metal
Root + IPMI access
Hardware swaps included

Reserve the N8

Nexus H8

Training on HGX with NVLink

$16,060/mo

$2.75 per GPU-hour · 8 GPUs

21% under the $3.49 managed-cloud median

GPUs8× H100 SXM5
VRAM640 GB HBM3
CPUDual Xeon / EPYC
RAM2 TB DDR5
NVMe30 TB
Network8× 400Gb IB

Single tenant, bare metal
Root + IPMI access
Hardware swaps included

Reserve the H8

Monthly price assumes 730 hours. Managed-cloud median observed November 3, 2026. Taxes, cross-connects, and extra IPs quoted separately. Availability confirmed at reservation.

Why dedicated

Cloud flexibility without cloud markup.

Shared clouds charge a premium for the same silicon and slice it across tenants. A reserved Nexus server is the whole machine at a flat rate, in a certified facility, with the hardware on us.

	Nexus dedicated	Shared cloud
Tenancy	Whole server, single tenant	Virtualized, shared host
Pricing	Flat monthly, per server	Metered, plus egress and storage
Access	Root + IPMI, your OS	Managed image, limited control
Hardware	Fixed box, same GPUs every day	Instance placement varies
Support	Named ops team, hardware swaps	Ticket queue

Included on every server

Everything you need to ship. Nothing you have to babysit.

Bare metal, single tenant

The whole machine for your term. Never virtualized, never shared.

Root and IPMI

Ubuntu + CUDA + Docker on delivery, or bring your own OS and image.

Fast networking

25–100GbE per server, InfiniBand on H8, private VLAN between your boxes.

Hardware on us

Failed GPU, fan, or drive? We swap it. Remote hands included.

Monitoring

GPU health, thermals, and ECC alerts — we catch failures before your jobs do.

Certified facility

Tier III-class sites with A+B power, badge access, and isolated management.

How it works

Request to root in four steps.

Request

Pick a server, GPU count, and term. We confirm availability, usually same day.

Sign the order form

One page: server, term, monthly price, start date. First month due at activation.

Provision

We image, burn-in check, and hand you SSH keys and IPMI. In-stock units go live in days.

Run

Your stack, your jobs. Flat invoice monthly; scale up by adding servers to the same VLAN.

Built for

Every stage of the AI pipeline.

Production inference

vLLM and TensorRT-LLM serving with predictable latency and no cold starts.

Best on: N8

Fine-tuning

LoRA, QLoRA, and full fine-tunes with 96 GB per card.

Best on: N4 / N8

Multi-node training

Distributed training over NVLink and InfiniBand.

Best on: H8

Image & video generation

Flux, SD, and video models that need dense VRAM per card.

Best on: N8

Embeddings & RAG

High-throughput embedding jobs and vector pipelines.

Best on: N4

Private & regulated

Isolated hardware for data that can't sit on shared clouds.

Best on: Any

Common questions

How fast can I get a server?

In-stock units typically go live within a few business days of a signed order form. If the configuration you want needs to be built, we give you a date when we confirm availability.

What's the minimum term?

Month-to-month is the minimum. 6- and 12-month reservations lock a lower per-hour rate for the full term.

Can I get hourly or spot access?

Not on reserved plans — these are whole servers by the month. Some Nexus-operated machines are listed on public GPU marketplaces between contracts if you need short hourly bursts.

What happens if hardware fails?

We replace failed GPUs, fans, drives, and PSUs at no cost. Monitoring usually flags a failing part before it takes the box down.

Can I scale up later?

Yes. Additional servers join the same private VLAN. Ask about multi-server pricing when you reserve.

Who owns the hardware?

Nexus operates every server and is your only counterparty. Some servers are owned by Nexus program participants; that never changes your contract, access, or support.

Reserve capacity

Tell us what you need.

We confirm availability, send the order form, and give you a go-live date — usually same business day.

Running 24/7 for a year or more?

Owning usually beats renting past the one-to-two-year mark. Buy the same server and we host it on a flat fee.

Buy a Server

Need racks or megawatts?

For multi-rack and multi-megawatt deployments, Nexus sources dedicated colocation across its network.

AI & GPU Colocation