EN
How to buy Request a quote

Dedicated NVIDIA GPU servers

Whole servers and clusters on 36 to 60 month terms, built to our published delivery specification.

See every GPU server we serve

Hardware sized to the workload

Which servers fit training, inference, media pipelines and generative video, and why.

Compare the four workloads

We bring the customers

Our pipeline holds more demand for GPU servers than we can deploy on our own. Partners put it to work.

How partnerships work

Built to a published specification

Every cluster is burned in before handover and run under written operating commitments.

Read the delivery specification

Engineering notes

Sizing guides, hardware comparisons and what changes when you move to dedicated servers.

Read the blog

Your ownGPU cluster.

1Legion is a bare metal GPU cloud. We rent dedicated NVIDIA GPU servers and clusters on 36 to 60 month terms, built to a published specification, burned in before handover and run under written operating commitments.

Request a quote Read the delivery specification
Allocation open in Toronto
HGX B300
HBM3e per GPU
270 GB
East-west network per GPU
800 Gb/s
Contract terms, one fixed rate
36 to 60 months

Inside a 1Legion cluster

A whole cluster for one customer

Our largest reference design puts 288 HGX B300 nodes on one rail-optimized, non-blocking fabric, with its own storage, CPU nodes and control plane. Nothing is shared: no other tenants, no hypervisor.

GPU nodes
288
NVIDIA B300 GPUs
2,304
Estimated IT load, about
4.39 MW

Racks on two power paths

Air-cooled HGX B300 servers, placed only in facilities that commit to A and B power for every rack, a contractual power limit per rack and the cooling to hold it. We plan about 14 kW for each GPU node.

Per GPU node, planning figure
14 kW
Independent power paths per rack
2
Local NVMe per node, at least
9.2 TB

Eight GPUs, one NVLink domain

Each server holds eight Blackwell Ultra GPUs with 270 GB of HBM3e each, joined by fifth-generation NVLink through NVSwitch, and gives every GPU its own 800 Gb/s path into the cluster fabric.

HBM3e per server
2.1 TB
NVLink per GPU
1.8 TB/s
Fabric bandwidth per GPU
800 Gb/s

Start at 64 nodes. Grow to 288.

Three reference sizes share one specification, so a cluster grows without a redesign. Larger clusters are designed to order, and every deviation is numbered in your proposal.

64 nodes128 nodes288 nodes
NVIDIA B300 GPUs5121,0242,304
Usable storage512 TB1,024 TB2,304 TB
IT load, about0.98 MW1.95 MW4.39 MW
See the reference clusters

Customers and partners

AI companies, GPU clouds, data center operators, carriers and hardware makers we work with.

See all 36
  • Akash Network
  • Vast.ai
  • Viggle AI
  • Dell Technologies
  • AudioShake
  • DDN
  • WEKA
  • Hewlett Packard Enterprise
  • Massed Compute
  • Compute Exchange
  • Zayo
  • Supermicro
  • Cogent
  • eStruxture Data Centers
  • Enovum Data Centers
  • Lenovo
  • CloudRift
  • Cocoon
  • Skyportal
  • Cisco
  • GenAI Works
  • DVEO
  • Computer Modules
  • ASUS
  • WiseDV
  • ISS Intelligent Security Systems
  • Templus
  • GIGABYTE
  • UK Servers
  • Glide
  • Inflect
  • Aivres
  • PodWarden
  • Ignite
  • Dacentec
  • AMAX

One operator, four ways to work with us

Rent the servers, have us run the software on them, place hardware in our partner facilities, or let us operate and fill a cluster you own.

GPU servers and clusters

Whole NVIDIA servers and dedicated clusters on 36, 48 or 60 month terms, from a single HGX B300 server to a 288-node cluster with its own fabric and storage.

Browse GPU serversDedicated GPU clusters

Managed Kubernetes

We install Kubernetes on your dedicated servers with the NVIDIA GPU Operator and Network Operator, then run, upgrade and monitor it. Managed Slurm if your team prefers it.

How managed Kubernetes works

Colocation and operations

Space and power in partner data centers, resold per kW on multi-year terms, with 24/7 operations, remote hands and monitoring for hardware you own.

See the two colocation tracks

Operations for GPU investors

You own or finance the servers. We deploy and run them, find the customers from our pipeline, contract and bill them, and report what each server earns.

How we run investor hardware

NVIDIA servers we deploy

Whole servers only: 8-GPU HGX systems, 72-GPU racks and PCIe servers for inference and media. Every configuration is confirmed in your proposal.

Allocation openBlackwell Ultra

NVIDIA HGX B300

Eight Blackwell Ultra GPUs with 270 GB of HBM3e each and 800 Gb/s of network per GPU, rented as whole servers or as dedicated clusters, built to our published delivery specification.

GPUs per server
8
Memory per GPU
270 GB HBM3e
FP4 per server
108 PFLOPS dense
Network per GPU
800 Gb/s
B300 servers Request a quote

See all GPU servers

Tested before you get the keys

Every cluster passes NVIDIA hardware validation for inference providers, five tests from HW-1 to HW-5, before handover, and you receive the acceptance report. It is why we do not run trials: the burn-in is the proof.

  1. HW-1Collective throughput. NCCL throughput at baseline with nccl-tests, with SHARP-accelerated collectives where the fabric supports them.
  2. HW-2Every network rail. Every InfiniBand or RoCE rail tested with ib_write_bw.
  3. HW-3GPUDirect and storage. GPUDirect RDMA and storage smoke tests, including NVMe read and write.
  4. HW-4NVLink and NVSwitch. NVLink and NVSwitch link count and bandwidth checked with nvbandwidth.
  5. HW-524-hour soak. A 24-hour soak under a representative inference load at a sustained power cap.
How burn-in and acceptance work
burn-in hgx-b300 node 0172, 8 GPUs illustrative
  1. HW-1 all_reduce_perf -b 8 -e 16G -f 2 -g 8 queuedrunningpass
  2. HW-2 ib_write_bw, every rail queuedrunningpass
  3. HW-3 ib_write_bw --use_cuda, NVMe read and write queuedrunningpass
  4. HW-4 nvbandwidth queuedrunningpass
  5. HW-5 24 h, inference load, power cap queuedrunningpass
  6. acceptance report issued to customer

Response times, in writing

Our operating commitments are written down, not implied. These are the clocks that start when something goes wrong, and the notice you get before we touch anything risky.

  1. 5 minInventory and state changes visible to you. DDoS scrubbing active.
  2. 15 minIncident acknowledged, a commander assigned, network operations engaged.
  3. 45 minFacility engineering or remote hands working on the problem.
  4. 60 minAn update at least every hour while an incident is open.
  5. 7 daysA written postmortem, and seven days' notice before risky maintenance.

Read the operating commitments

Where your servers run

Capacity in Canada today, Delaware next and Finland in development, inside a partner network of 49 data centers on four continents. We name data centers by city.

  • TorontoCanada In service NVIDIA HGX B300 servers and clusters.
  • MontrealCanada In service RTX PRO 6000 Blackwell and other inference servers, with onsite engineering support.
  • DelawareUnited States Coming online HGX B300 capacity for customers who need their compute in the United States.
  • HelsinkiFinland In development A brownfield data center site in development for future capacity in Europe.
  • Partner network26 cities, 12 countries 49 data centers North America, Europe, Africa and Australia, with at least 313 MW of power. See every city
Data centers

More demand than we can deploy on our own

Our pipeline holds requests for thousands of GPU servers. Investors, data center owners and brokers work with us to serve it.

Partner with 1Legion

Engineering notes

Sizing guides, hardware comparisons and what changes when you move to dedicated servers.

All articles
70GB of weights, four GPUs: what’s left for the KV cache at 80, 96, 141 and 270GB Load a 70B model in FP8 on 80, 96, 141 and 270GB GPUs and see how much memory each leaves for the KV cache, with the sizing arithmetic worked out. GPUs 4 min read 8 x 270GB: what fits on a B300 node What fits on an NVIDIA B300 node: 270GB memory math for 405B and 70B model inference, KV cache sizing, bandwidth ceilings, and where a single GPU stops. GPUs 6 min read 8 x 96GB: what fits on an RTX PRO 6000 Blackwell node What fits on an RTX PRO 6000 Blackwell node: the 96GB memory math for 70B model inference and QLoRA fine tuning, KV cache sizing, and where the limit sits. GPUs 4 min read

Tell us what you need to run.

Share the hardware, the number of servers, the term and the start date. You receive a written proposal built on our delivery specification.

Request a quote Talk to an engineer