Training
The process used to create or substantially update a model.
Canadian AI infrastructure
AI has to run somewhere.
Every AI response ultimately depends on physical compute: GPUs, storage, networking and servers operating in a real location.
For organizations with residency, privacy or sovereignty requirements, the infrastructure beneath the model matters as much as the model itself. Underlabs designs the system around the level of Canadian location, operational control and infrastructure independence the workload actually requires.
The compute beneath the model
Every AI response is produced by physical hardware somewhere.
An application, database and documents can reside in Canada while inference runs in another region under another provider.
Sovereignty questions become more concrete as we move down the stack.
- 01User
- 02Application
- 03Company data
- 04AI model
- 05Inference
- 06GPU and compute
- 07Storage and network
- 08Physical facility
Where is it?Who operates it?Who controls it?
Compute, in plain language
The model has to execute on a machine.
Inference
Using an already-trained model to process an input and produce a result. For most Underlabs client systems, this is the primary infrastructure question.
- 01Web application · Montréal
- 02Database · Montréal
- 03AI API · United States
- 04Inference · United States
Application residency and inference residency are separate architecture decisions. See the complete data path.
Five different questions
Locate control, not only the servers.
These questions turn a general preference for Canada into requirements that architecture and procurement can verify. They are not legal advice.
Physical location
Where are the machines: in Canada, elsewhere, in a data centre or at the customer’s facility?
Provider
Who operates the infrastructure: a global provider, a Canadian company, a colocation operator or the customer’s IT team?
Administrative control
Who controls credentials, networking, operating systems, model deployment, monitoring and updates?
Dependency
Which external organization must remain available? Can the workload move to another environment?
Jurisdiction and governance
Which organizational and legal relationships apply? Location, ownership, operation and jurisdiction are related but not identical.
An important distinction
Canadian region ≠ Canadian-controlled infrastructure
A Canadian region operated by a global cloud provider can deliver genuine physical residency in Canada, mature tooling, high reliability, rapid scaling and enterprise support. For many organizations, that is entirely appropriate.
A Canadian region and Canadian-controlled infrastructure answer different requirements. Hardware may be in Canada while the organization operating or controlling it is based elsewhere. That does not make the service unsafe; it describes a different control boundary.
- Canadian cloud region
- Infrastructure physically in Canada, potentially operated by an international provider.
- Canadian data centre
- A physical facility in Canada that may host hardware belonging to providers, Canadian or foreign companies, and customers.
- Canadian provider
- A Canadian-controlled organization offering infrastructure or hosting.
- Customer-controlled Canadian infrastructure
- Infrastructure in Canada where the customer retains stronger direct control over deployment.
A spectrum, not a ranking
Direct control increases with operational responsibility.
No point on the spectrum is universally superior. Each option changes cost, control, maintenance, scalability, resilience and deployment speed.
- 01Managed AI APILowest operational responsibility; limited direct control.
- 02Cloud GPU, Canadian regionFast deployment, elastic capacity and managed tooling with physical residency in Canada.
- 03Dedicated Canadian environmentMore dedicated resources and a different dependency profile, evaluated provider by provider.
- 04Private Canadian infrastructureAn isolated or private environment with defined technical and operational boundaries.
- 05Customer-owned colocationThe customer owns the servers; the data centre supplies space, power, cooling and connectivity.
- 06On-premiseGreater direct physical control with greater operational responsibility.
- 07Edge or isolated environmentInference close to a device, sensor or operation, sometimes without continuous connectivity.
A complete infrastructure system
An AI system is more than a GPU in a server.
Model selection and infrastructure selection are connected. See how to choose a Canadian or international AI model.
GPU and memory
A GPU is a processor suited to the parallel calculations used by modern models. The model must fit in accelerator memory; larger models may need more memory, more GPUs, quantization or partitioning.
Storage
Model weights, documents, databases, embeddings, indexes, application files, outputs, logs and backups all need planned storage.
Networking
Application-to-model traffic, storage access, private network boundaries, latency, redundancy and interconnects affect performance and reliability.
Power and cooling
Private high-performance systems need electrical capacity, cooling, space, redundancy and maintenance. Owning the GPUs is not automatically simpler.
Current Canadian context
Sovereign compute has become a public-policy category.
Sovereign compute generally refers to infrastructure designed to keep strategically important workloads within an intended national, organizational or legal control boundary. It is not a universal technical certification.
Canada’s current national strategy presents compute, cloud, connectivity, data and infrastructure as foundations of AI sovereignty. It calls for expanding sovereign capacity while also treating hyperscaler investment as part of Canada’s broader compute ecosystem.
National infrastructure
The AI Sovereign Compute Infrastructure Program supports Canadian public advanced-compute capacity.
View the programGovernment infrastructure
Shared Services Canada is deploying a platform combining Canadian compute, storage and AI services for federal departments and agencies.
View the GC AI Platform2026–27 direction
SSC’s departmental plan describes sovereign AI infrastructure, hybrid options, edge compute and expanded Canadian capacity.
Read the departmental planCanada’s National Artificial Intelligence Strategy. This public-policy context is neither an architecture recommendation for every company nor an endorsement of Underlabs.
Capacity planning
Benchmark before buying.
Capacity planning estimates the infrastructure actually needed based on users, requests per second, model size, context, response length, concurrency, latency and availability targets.
Too little capacity makes the system slow or unreliable. Too much leaves expensive GPUs idle.
Cloud can fit when…
- the workload is uncertain or variable
- experimentation is ongoing
- rapid scaling matters
- operations staff are limited
Owned or dedicated infrastructure can fit when…
- the workload is stable and utilization is high
- control requirements justify it
- long-term economics support it
- local or offline operation matters
The full calculation includes hardware, power, cooling, maintenance, redundancy, staff and refresh cycles.
Hybrid infrastructure
One organization can combine several environments.
Private AI does not always require customer-owned hardware. See private and on-premise AI topologies.
Sensitive workload
- 01Private Canadian compute
- 02Private model
General AI tasks
- 01Approved external model
Immediate local decision
- 01Camera or sensor
- 02Local AI device
- 03Decision
- 04Central system
Edge AI can reduce latency and data transfer or enable offline operation, but adds device capacity, deployment and management constraints.
Proportionate resilience
Resilience is the ability to keep operating through a failure. Failover switches to another resource. Depending on criticality, plans may address GPU, network, storage, provider, model API or region outages.
From requirement to working system
Underlabs connects architecture, infrastructure, model and product.
Underlabs is not a data centre, hardware manufacturer or GPU provider. We determine the appropriate environment, deploy the AI stack and build the useful software above it.
- 01
Define the workload
Specify what AI must do, for whom and with which data.
- 02
Determine sovereignty requirements
Identify what must remain in Canada, private, customer-controlled or locally available.
- 03
Benchmark models
Measure real capability on the system’s tasks, languages, formats and failure modes.
- 04
Estimate compute
Connect model, volume, concurrency, latency and availability to required resources.
- 05
Select the deployment architecture
Compare Canadian region, dedicated, private, colocation, on-premise, edge and hybrid environments.
- 06
Build the application layer
Develop software, APIs, retrieval, agents, automation, authentication and integrations.
- 07
Deploy and monitor
Connect software and infrastructure into an operable, observable and recoverable system.
- 08
Preserve portability
Reduce unnecessary dependence on one model or infrastructure provider where practical.
Procurement guide
Questions to ask an AI infrastructure provider
Precise answers to these questions are more useful than a general cloud, Canadian-hosting or sovereignty label.
- 01Where will inference physically run?
- 02Where will model weights and customer data be stored?
- 03Who has administrative access?
- 04Is the infrastructure shared or dedicated?
- 05Can the workload move to another provider?
- 06Where are backups and logs stored?
- 07What happens during an outage?
- 08Who manages security updates?
- 09Can the environment operate without external dependencies?
- 10What data leaves the environment?
The complete loop
A Canadian GPU server does not automatically make the whole system sovereign.
- 01Canadian GPU
- 02Private model
- 03External logging provider
- 04Sensitive data leaves the environment
Sovereignty emerges from the architecture of the complete system.
Start with requirements
You do not need to choose the infrastructure before talking to us.
Share the workload, sensitive data, Canadian requirements, current systems and operating goals. Underlabs can translate them into an architecture, benchmark models and capacity, then build and deploy the system.