Trust · Infrastructure
Where Digby’s models actually run
Storing data in Canada is half the answer. The other half is where the model that reads it is running — and for Digby, that is Canadian-located GPU hardware.
01
Storage residency is only half the story
Most “Canadian data residency” claims cover storage only: your data sits in a Canadian data centre, and then the prompt is shipped to a model running somewhere else to be answered. The bytes were at home; the thinking happened abroad.
For a privacy assessment that distinction matters. What a model reads, it processes; what it processes, another jurisdiction may be able to compel. A residency claim that stops at the disk does not answer the question a privacy officer is actually asking.
Digby answers both halves: your account, conversations, memory and uploaded files are stored in Canada, and the models that write the answers are served on Canadian-located GPU hardware operated in Canada.
02
The hardware, honestly described
Digby’s inference runs on NVIDIA GPU hardware in Canada — a data-centre-grade GPU server of the B300 class for production serving, plus NVIDIA DGX Spark nodes used for development and tuning work in a lab in Victoria, British Columbia.
Those are real machines, not a cloud-region checkbox: they are operated in Canada, and the models serving chat traffic are the same family of custom-tuned open-weight models described elsewhere on this site.
We deliberately do not publish facility addresses, network details, customer names or security architecture here. “Canadian-located GPU hardware operated in Canada” is the claim; a visitor does not need a floor plan to evaluate it, and publishing one would be a bad trade for everyone.
03
Why this changes what you can promise
When storage and inference are both in Canada, an organisation can say something concrete to its clients, its regulator and its board: the information stays in the country, and so does the processing. Digby’s models being open-weight is what makes that possible — a closed foreign API could not be moved onto Canadian hardware, however the contract was worded.
04
The three things that still leave Canada
We would rather list the exceptions than be caught by them. When Digby searches the live web, the search query goes to a search service and pages are fetched from the public web. When you connect an app — Gmail, Google Drive, Slack, Asana, Xero, GitHub, Discord, Telegram, iMessage — Digby talks to that app on your behalf, so that data goes to (and comes from) that service. And the marketing site’s analytics and advertising tags send usage data to Google and Meta.
None of those is the chat itself: conversations, files and memory stay in Canada, and none of it is used to train AI models.
05
Run it on your own GPUs
If inference should happen on hardware you own, self-hosted Digby runs the full stack — database, interface, integrations and coding workspace — on your servers, pointed at any OpenAI-compatible model endpoint: vLLM on your own GPUs, Ollama, LM Studio, or a managed GPU service you already operate.
That is the same product, not a cut-down edition: projects, knowledge base, scheduled work and the coding agent all run locally, and nothing reaches digby.ai unless you configure it to.
$ cp selfhost.env.example .env $ docker compose -f docker-compose.selfhost.yml up -d --build
06
Your own sovereignty checklist
- 01
Storage
in which country, and which region, is the database and file storage?
- 02
Inference
where does the prompt go to be answered, and can you name the operator?
- 03
Models
open-weight and replaceable, or a closed model you cannot move?
- 04
Training
is your content used to train or fine-tune anything?
- 05
Deletion
what happens to a file, a conversation and an account, and how fast?
- 06
Exit
could you run the system yourself next quarter if you had to?
Frequently asked questions
Do Digby’s models run in Canada?
Yes. Digby’s models are served on Canadian-located NVIDIA GPU hardware operated in Canada, not on foreign API endpoints.
Why does inference location matter?
Storage residency only covers where data sits. A model that reads your prompt processes it wherever that model runs, so the jurisdiction of processing matters as much as the jurisdiction of the disk.
What hardware does Digby run on?
Data-centre NVIDIA GPU servers of the B300 class for production serving, and NVIDIA DGX Spark nodes for development and tuning work in Canada.
Does anything leave Canada?
Three things, and we name them: a web search sends the query to a search service, a connected app (Gmail, Slack and so on) talks to that app, and the marketing site’s analytics and ad tags go to Google and Meta. Conversations, files and memory stay in Canada.
Can we run the models on our own GPUs?
Yes. Self-hosted Digby points at any OpenAI-compatible endpoint — vLLM on your own GPUs, Ollama or LM Studio — so inference happens on hardware you operate.
Is the infrastructure certified?
No. We hold no SOC 2, ISO 27001 or similar certification, and we do not claim one. This page describes where things run, not an audit result.
Try asking Digby
Keep reading
Try Digby now.
No account needed. Models and data stay in Canada.