AI and agents
Self-hosted, Cohere or Dlytica-managed: three ways to run LLMs on sovereign data
There are three practical ways to run large language models on sensitive data: self-host open-weight models on your own GPUs, use Cohere models through its API, Model Vault or private deployment, or let Dlytica run models for you inside your jurisdiction. Choose by data sensitivity, volume, in-house skills and how much capability you need.
Published . Last updated
On this page
What are the three options?
| Self-hosted open models | Cohere | Dlytica-managed | |
|---|---|---|---|
| Who runs the GPUs | You | Cohere, or you for private deployment | Dlytica, in your chosen region or your account |
| Models | Open-weight (for example Qwen, Mistral, Llama families) | Command, Embed, Rerank, North | Open-weight and Cohere, behind one gateway |
| Data boundary | Your network | Depends on option; private deployment stays in yours | Your jurisdiction, under a Canadian or local operator |
| Skills needed | MLOps, GPU operations, model evaluation | Integration | Integration |
| Cost shape | Hardware or GPU rental plus engineering time | Per token or per instance | Monthly service, sized to usage |
| Best for | High, steady volume; strict isolation; strong platform team | Enterprise RAG, multilingual answers, reranking | Teams that want sovereignty without running GPUs |
When does self-hosting make sense?
When data cannot leave your network, volume is high and steady, and you have people to run it. Practitioners consistently warn that raw GPU price is not the full cost: monitoring, upgrades, evaluation and idle GPUs add up. A GPU at 10% utilisation costs ten times as much per token as one kept busy.
When is Cohere the better choice?
When you want enterprise-grade generation, embeddings and reranking with a clear path to private deployment, and a Canadian vendor. Cohere's Embed and Rerank pair especially well with lakehouse search. See using Cohere with DataNature.
What does Dlytica-managed mean?
Dlytica operates the model serving for you: GPUs in a sovereign cloud such as R2i Edwin in Canada or EverestCloud in Nepal, or in your own cloud account. You get an OpenAI-compatible endpoint, the Atlas AI gateway for approvals, logs and budgets, and model upgrades handled for you. Prompts and outputs stay in the jurisdiction you choose.
Can we combine them?
Most teams do. Route sensitive workloads to private models, use Cohere for search quality, and keep a frontier API for non-sensitive tasks if policy allows. The Atlas AI gateway makes that a configuration choice, not a rewrite.
What should we decide first?
- Classify data: which prompts may leave the boundary, if any?
- Estimate volume per use case.
- Decide who will operate models day to day.
- Run an evaluation on your own tasks before choosing a model.
How does DataNature Atlas keep this secure and compliant?
Agents are governed like users: scoped, masked, logged and unable to act without approval. DataNature Atlas is the platform's governance module:
- Knows your data: a catalog with automated PII detection and classification across Bronze, Silver and Gold.
- Controls access: role-based grants plus attribute-based policies, row filters and column masks (null, redact, partial, hash, year only).
- Governs AI: Voir agents are scoped principals; masks, row filters and PII guardrails apply to every tool call and answer.
- Proves it: lineage from source to metric to decision, and an audit log of who or what accessed which data.
- Stays local: Atlas runs inside your deployment, so policies, lineage and logs never leave your jurisdiction.
These controls support GDPR, PIPEDA, Quebec Law 25, CCPA, HIPAA and sector rules such as OSFI E-23. Read DataNature Atlas in depth or see security and compliance.
Platform components in this article
Learn more about each DataNature product and module mentioned here.
See it on your own data.
Book a demoDemos and quotes are handled by Dlytica, the company behind DataNature.