Hardware sized to the workload
The right server depends on what it has to hold in memory and how far it has to talk. These pages match NVIDIA servers to four kinds of work.
1Legion rents dedicated NVIDIA GPU servers for four kinds of work. Each page below explains which servers fit, what decides the size, and where they run.
Four workloads
Each workload puts a different limit first: fabric for training, memory for inference, encoders and drivers for media, and memory for resolution and length in generative video.
AI training
Long training runs need burned-in hardware and a non-blocking fabric. HGX B300 clusters of 64 to 288 nodes and HGX B200 servers, with GB300 NVL72 and Vera Rubin by reservation.
AI inference
Dedicated servers keep latency predictable. Size by weights plus KV cache: 270 GB per GPU on B300, 96 GB on RTX PRO 6000 Blackwell, 48 GB on L40S.
Media and entertainment
Rendering, VFX, virtual production and transcoding on whole servers, with no egress fees for moving large media.
Generative video
Diffusion pipelines such as ComfyUI, where GPU memory decides resolution, clip length and batch size.
How we size a server
Four questions decide the hardware, and a proposal answers each one in writing.
- What has to fit in GPU memory: the weights, the KV cache or activations, and headroom for the runtime.
- How far the GPUs have to talk: within one server over NVLink, or across a cluster over an 800 Gb/s per GPU fabric.
- How many servers, and for how long: whole servers on 36, 48 or 60 month terms, with lower rates for longer terms.
- Where they run: Toronto and Montreal are in service, Delaware is coming online, Finland is in development, and our partner network covers more cities on four continents.
Our GPU server catalog lists every model we deploy, and the delivery specification sets out how HGX B300 clusters are built.
Questions buyers ask
How do you decide which GPU server fits our workload?
We start from memory: the model weights, the KV cache or activations, and headroom. Then interconnect, cluster size and location. A quote request with your model sizes and traffic is enough to start.
Can one contract mix training and inference servers?
Yes. A proposal can include HGX B300 servers for training and RTX PRO 6000 Blackwell servers for inference, each on 36, 48 or 60 month terms.
Do you rent GPUs by the hour for a test?
No. We rent whole servers on 36 to 60 month terms, and every cluster passes burn-in and acceptance tests before handover instead of a trial.
Where do the servers run?
HGX B300 capacity is in Toronto, RTX PRO 6000 Blackwell and inference servers in Montreal, more B300 capacity is coming online in the United States, and a site in Finland is in development. See data centers.
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.