Sovereign AI Box Canada is our hardware AI stack. The Box bundles an H100 GPU with an open-weight LLM and an agent runtime. We install it on your hardware in Canada. The stack ships sovereign by default. Data never leaves your jurisdiction. There is no per-token cost. There is no outside prompt logging. You get a full audit trail. You also get a regulator-ready evidence pack.
Most AI buyers route prompts through a SaaS LLM API. However, they send Canadian data to a US data centre. Additionally, they accept the outside logging that comes with it. In contrast, the Sovereign AI Box inverts that path. First, the hardware sits on your floor. Otherwise, it sits inside a Canadian-region cloud account you control. Next, the model weights sit on disk you control. Then, the agent runtime emits structured event logs. Specifically, it sends them to a Canadian-region log store. Therefore, every prompt stays inside your jurisdiction. Likewise, every completion stays too. Similarly, every tool call stays.
The Box ships in three hardware tiers. First, the single-H100 tier suits dev work and small inference. Second, the dual-H100 tier suits single-model inference at moderate scale. Third, the eight-H100 tier suits multi-model inference at scale. Notably, that includes 405B-parameter models running with NVLink. Furthermore, we size the tier against your token budget. Likewise, we size against your concurrency target. Ultimately, the choice is yours; we ship the spec.
Region runs four ways. AWS Canada Central suits buyers who want Canadian cloud. Toronto on-prem suits buyers with a Toronto datacentre. Montreal on-prem suits Quebec buyers under Treasury Board Directive on Service and Digital rules. Hybrid suits buyers who train on-prem. They then run inference in the Canadian cloud. The Box recipe lifts across all four regions.
LLM size runs four ways. First, the 8B-parameter tier suits prototypes and routing. Next, the 70B-parameter tier suits production chat and agent work. Then, the 405B-parameter tier suits research work where model size outweighs cost. Finally, the mixture tier routes prompts across many models in one Box. Specifically, it suits buyers who want specialised models for specialised tasks. In addition, open-weight Llama 3.1 and Qwen 2.5 are the default families. Ultimately, the operator picks the licence path.
Sovereign AI Box Canada speaks to regulated industries. For example, healthcare buyers run patient data under PHIPA. Similarly, fintech buyers run client data under OSFI rules. Likewise, defence buyers run procurement data under ITSG-33 with a Protected B target. In addition, public-sector buyers run citizen data under federal data-residency rules. Consequently, each industry hits the same wall with SaaS LLM APIs. Ultimately, the Box clears the wall.
The Box sits inside the Build trunk. It is the top tier on hardware. Agents that run on the Box come from our agentic systems work. Buyers who want a foundation pass first book our Sovereign Infrastructure Brief. That engagement produces the design record the Box runs against. Buyers with an AI footprint book our Intelligence Audit to map the migration. Buyers who want a managed cloud tier pick our Open Claw Enterprise instead. Patterns and runbooks sit in the Library. Threat-side context lives under the Defend trunk. New lab work is shared at Research.
Pricing runs as an AggregateOffer. The single-H100 tier starts at $45,000 CAD. The dual-H100 tier starts at $85,000 CAD. The eight-H100 tier starts at $280,000 CAD. Mixture and 11-plus-agent setups price on the kickoff call. Timeline runs 8 to 16 weeks. Hardware lead time on the chosen NVIDIA H100 SKU drives the wide end. The handover runbook covers install, observability, audit-trail tooling, and the 12-month monitoring contract. A quarterly review keeps the stack current.
Configurator (M1 stub; live configurator arrives in milestone 2)
The live configurator is plugin-deferred. Until milestone 2 ships, the table below shows the choice surface. Buyers walk it on the kickoff call.
Hardware tier
- Single H100: starts at $45,000 CAD. Dev workloads. Small inference. 8B or 70B models.
- Dual H100: starts at $85,000 CAD. Single-model inference at moderate scale. 70B models with headroom.
- 8x H100: starts at $280,000 CAD. Multi-model inference. 405B models with NVLink. Mixture routing.
Region
- AWS Canada Central. Canadian cloud. No colocation footprint needed.
- Toronto on-prem. Your own datacentre or a colocation contract.
- Montreal on-prem. Quebec data-residency rules supported.
- Hybrid. On-prem training. Canadian cloud inference.
LLM size
- 8B-parameter. Llama 3 or Qwen 2.5 small.
- 70B-parameter. Llama 3.1 70B production-grade.
- 405B-parameter. Llama 3.1 405B research-grade.
- Mixture. Multi-model routing. Specialised tasks.
Agent count
- 1 to 3 agents. Single workflow. Narrow scope.
- 4 to 10 agents. Multi-workflow. Departmental scope.
- 11 plus agents. Custom scope. Mixture routing. Priced on call.