Sovereign AI Box Canada: H100 LLM agents on your hardware

Sovereignty here covers four guarantees. First, data never leaves your Canadian jurisdiction at any point in the prompt and completion path. Second, there is no third-party prompt logging because no third party sits in the path. Third, every prompt, every completion, and every tool call lands in a tamper-evident audit trail you control. Fourth, the evidence pack we ship is regulator-ready against ITSG-33 with a Protected B classification target and against the Treasury Board Directive on Service and Digital. SaaS LLM APIs hit a wall on at least one of these guarantees. The Box clears all four.

The Box ships four region options. AWS Canada Central suits buyers who want a Canadian-region cloud footprint without colocation. Toronto on-prem suits buyers who already have a Toronto datacentre or a colocation contract. Montreal on-prem suits Quebec buyers under provincial data-residency mandates, including Law 25. Hybrid suits buyers who run model training on-prem and run production inference in the Canadian cloud. The Box recipe lifts cleanly across all four. The choice happens on the kickoff call and is recorded in the architecture decision record.

The default model families are Llama 3.1 and Qwen 2.5. Llama 3.1 ships under the Meta community licence with the usage and naming clauses that licence requires. Qwen 2.5 ships under Apache 2.0 with the more permissive redistribution terms. The operator picks the licence path on the kickoff call. We also support other open-weight families if the operator has already licensed a model from the Mistral, DeepSeek, or research-foundation pools. Custom model training is excluded from the Box engagement and runs through a separate scope.

The configurator matches your workload to a specific hardware tier, region, LLM size, and agent count. A buyer running a 70B production chat workload at moderate concurrency lands on a dual-H100 tier with Toronto on-prem region and 4-to-10 agents. A buyer running 405B research workloads at heavy concurrency lands on an eight-H100 tier with NVLink and a single high-quality model. The configurator codifies that decision before procurement starts. The live M2 configurator ships in a future plugin milestone; until then, the dimension table on this page renders the choice surface buyers walk through on the kickoff call.

Timeline runs eight to sixteen weeks from order to operational. Hardware lead time on the chosen NVIDIA H100 SKU drives the wide end of the range. The single-H100 tier typically lands at eight to ten weeks. The dual-H100 tier typically lands at ten to twelve weeks. The eight-H100 tier typically lands at twelve to sixteen weeks because the NVLink fabric ordering and the power-and-cooling validation steps both add lead time. We start agent runtime work and observability staging in parallel with hardware delivery so the operational date does not slip on installation alone.

Three upgrade paths exist. First, you add a second Box in parallel and route workloads across both via the agent runtime. This is the cleanest path when the bottleneck is concurrency, not model capability. Second, you scale the tier in place by upgrading the GPU configuration, typically single H100 to dual H100 or dual H100 to eight-H100 with NVLink. This is the right path when the bottleneck is model size and you need 405B in production. Third, you move from a single-model deployment to a mixture deployment where the agent runtime routes prompts across multiple specialised models on one Box. Each upgrade path is documented in the handover runbook with cost and timeline estimates.