01
On-demand
Billed by the hour. Suited to experiments, ablations and short fine-tunes where the shape of the workload is still unclear.
- Single node or small multinode
- Start within the same working day
- Stop and restart at will
Compute
Reserved GPU capacity, with a start date.
Hourly GPU capacity depends on what happens to be free that day, which is unreliable when the market is tight. We rent clusters with a fixed start and end date, sized to the run you are planning.
01 — Terms
Three ways to buy.
All three run on the same hardware and the same platform. The difference is how long you commit and how much of the cluster is yours alone.
02
Reserved
A named cluster held for a fixed term. The usual choice for pre-training, where a week of lost wall-clock time costs more than the reservation.
- Guaranteed start date
- Fixed term, fixed shape
- Capacity that cannot be allocated to anyone else
03
Dedicated infrastructure
A single-tenant deployment, in our facility or yours. We handle build-out, procurement and operations, and the hardware is not shared with anyone else.
- Single-tenant hardware
- Build-out, power and cooling
- Optional managed operations
We quote after scoping: model size, sequence length, batch shape and expected wall-clock per run. A rate card without those inputs would be a guess.
02 — The stack
What sits under the job.
Bare metal, with the scheduler your team already uses. You get root, your own container images, and an interconnect that holds up when every node is talking at once.
For genomic pipelines, the same cluster runs GPU-accelerated alignment and variant calling, so you are not moving TB-scale output between two vendors.
None of this is unusual. It is a standard stack, installed properly, with one team accountable when a node drops.
Cluster profile
Reference
- Accelerators
- NVIDIA data-center GPUs, current generations, 8 per node typical
- Interconnect
- High-bandwidth node-to-node fabric for distributed training
- Storage
- Local NVMe for checkpoints and dataset cache; shared object storage for long-term data
- Scheduling
- Kubernetes or Slurm, provisioned to your team's preference
- Images
- NVIDIA container runtime, PyTorch and CUDA stacks, your own images welcome
- Location
- Canada, with residency commitments available in the agreement
- Support
- Engineer on the account, escalation out of hours
03 — Reliability
What happens when hardware fails.
At cluster scale, a card or a network interface will fail during a long run. What matters is whether that ends the job or costs twenty minutes.
Automatic recovery
Failed nodes are drained and replaced, and the scheduler is allowed to resume from the last checkpoint instead of restarting from zero.
Checkpoint strategy, agreed up front
We agree the checkpoint interval and storage path before training starts. It is the cheapest insurance in a project and the step most often skipped.
Maintenance you can plan around
Windows are scheduled with you rather than announced during a run. Reserved clusters are not repurposed between your jobs.
04 — Questions
What teams ask first.
How fast can we start?
On-demand capacity is usually live the same working day. A reserved cluster needs a scoping call and a start date — the limit is hardware lead time rather than paperwork.
Can we keep the data in Canada?
Yes. Canadian residency can be written into the agreement, covering both compute and stored datasets. If your institution requires it, say so at the scoping stage rather than after.
Do you support Slurm, or only Kubernetes?
Both. Most research groups want Slurm and most platform teams want Kubernetes; we provision either, and we can run Slurm inside Kubernetes if your team has opinions about both.
What happens to our data when the engagement ends?
It is returned to you and removed from our systems on an agreed schedule, with written confirmation. We do not retain customer datasets for our own model training.
Can you run our pipeline instead of just giving us hardware?
That is the second half of what we do. Our engineers can take on data pipelines, distributed training and deployment, either alongside your team or as the delivery team. See Platform.
What are you planning to train?
Send the model family, dataset size and target start date. We will come back with a shape, a date and a price.