GPU compute · Model engineering · Vancouver, Canada
Curemodel AI rents NVIDIA-certified GPU clusters from Vancouver and helps teams train and deploy models on them. Customers are AI drug discovery companies and research groups that need capacity for a run, and an engineer to call when it fails.
NVIDIA-certified infrastructure On-demand or reserved Canadian data residency
01 — What we do
Most customers start with capacity and later use the platform for tuning and deployment. Both run on the same hardware, so a project does not have to be rebuilt when it moves from experiments into production.
01 / Compute
Clusters on NVIDIA-certified hardware, with automatic recovery. Rent by the node, or reserve a fixed shape for the length of a project.
02 / Platform
Reinforcement learning, supervised fine-tuning and parallel LoRA sweeps on the same cluster you train on. Checkpointing and evaluation are part of the platform, so a run that dies at hour nine does not start again from zero.
03 / Operations
Build-out and operations for teams that buy their own hardware. We design the floor, commission it and run it, so you do not have to hire a facility team.
“A cluster that arrives three weeks late is worth very little, whatever hardware is inside it.”From our compute terms
When a lab reserves a cluster, the start date is in the contract. We hold the hardware for the term of the job and do not reallocate it, even if another customer asks for the same machines.
After that, most of the work is ordinary engineering: keeping the data pipeline up with the GPUs, choosing a checkpoint interval, and having someone on the phone who knows the cluster.
02 — Applications
Compute is the business today. These four are what we are building on top of it. All of them are in development and none is a released product.
Variant interpretation, pathogenic site screening and therapy matching from whole-genome data. The heaviest of the four in compute terms, and the one we spend most of our time on.
Structure and conformational change prediction for target discovery, epitope analysis and binder design.
Generative design with multi-objective screening on activity, selectivity, developability and toxicity. The output is a ranked candidate list for synthesis.
Detection, segmentation and quantification across pathology, CT and MRI, linkable to genomic data for combined studies.
03 — How it runs
Three stages. A project can start at any of them, and most start at the first.
01
We look at the workload: model size, sequence length, batch shape and how long a run takes. Single node for a fine-tune, multi-node with fast interconnect for pre-training. You get a start date and the cluster for the term of the job.
02
Our engineers work alongside your ML team on the parts that break at scale: data pipelines, distributed training, checkpointing, evaluation. A distributed job that is going to fail usually fails early, and it is cheaper to find that on day one than on day five.
03
A deployment with monitoring, a rollback path and a number that reaches an engineer during the engagement. Once the model is in front of users, support does not move to a ticket queue.
04 — Sizing
A rough guide to how we shape a cluster. Most of the time the limit is interconnect and checkpoint throughput rather than the speed of a single card.
| Workload | Typical shape | What decides the outcome |
|---|---|---|
| Pre-training from scratch | Multi-node, 8 GPUs per node, high-bandwidth fabric | Interconnect and checkpoint throughput. One slow node holds up the rest. |
| Fine-tuning (LoRA / QLoRA) | Single node, 8 GPUs, fast local NVMe | Memory headroom and dataset feeding. Storage is usually the bottleneck. |
| RL post-training | Training pool plus a separate rollout/inference pool | Scheduling between generation and learning. Idle GPUs on either side are waste. |
| Batch inference | Autoscaled pool, mixed precision, queued jobs | Cost per sample and time in the queue. Peak throughput matters less. |
| Genomic pipelines | GPU-accelerated alignment and variant calling | I/O on TB-scale inputs. Speeding up compute without fixing storage only moves the queue. |
Shapes are guidance, not a price list. We quote after scoping the workload.
05 — Data and paperwork
We do not train shared or public models on customer datasets. What you bring in is used for your workload only, and removed when the engagement ends.
Clusters and stored datasets can be kept in Canada for teams with jurisdictional or institutional requirements.
NDA, data processing terms, security questionnaire, architecture review. These are normally completed before a GPU is allocated.
Tell us the model, the dataset size, how long a run takes and when you need to start. With those four we can say whether we can hold the capacity and what it costs.