# Sovereign LLM gateway: Cohere, open-weight and managed models behind one control point

> A sovereign LLM gateway on DataNature routes every AI request through the Atlas AI gateway to an approved model: Cohere through API or private deployment, open-weight models served by DataNature AI, or Dlytica-managed models. It enforces per-team budgets, logs prompts and answers in your jurisdiction, and applies the same permissions as your data.

Source: https://datanature.ai/solutions/sovereign-llm-gateway
Publisher: Dlytica Inc. (DataNature is a Dlytica company)
Updated: 2026-11-05

Approved models, budgets, logs and routing for every AI use case, inside your jurisdiction

## Components

- **Atlas AI gateway**: Approved models, routing rules, budgets, prompt and response logs
- **DataNature AI**: Serves open-weight models and embeddings inside the deployment
- **Cohere**: Command, Embed and Rerank via API, Model Vault or private deployment
- **Voir**: Agents and assistants that call models through the gateway
- **Atlas**: Identity, permissions, lineage and audit

## Steps

1. **Classify use cases**: Decide which data each use case touches and which models are allowed for it.
2. **Register models**: Add Cohere endpoints, self-hosted open models and managed models to the gateway catalogue.
3. **Set routing and budgets**: Route each team and agent to approved models, with token and cost limits and alerts.
4. **Enforce data permissions**: Retrieval and tool calls run with the user’s or agent’s Atlas permissions.
5. **Log and review**: Prompts, sources, outputs and costs are logged in your jurisdiction for audit and evaluation.

## Documentation

- [DataNature documentation](https://docs.datanature.ai)

## Why put a gateway in front of models?

Without one, every team integrates models its own way, keys spread, costs are invisible and nobody can answer "which data did this AI see?". A gateway makes AI governable.

## Which models can sit behind it?

Cohere models through API, Model Vault or private deployment; open-weight models served inside DataNature; and Dlytica-managed models in a sovereign region. Frontier APIs can be added where policy allows.

## How does it help with regulators?

It gives an inventory of models and use cases, logs for every call and lineage to source data, which supports model risk expectations such as OSFI E-23. See [OSFI E-23 explained](/blog/osfi-e-23-guideline-canadian-financial-institutions).

## Common questions

### How much does a self-hosted LLM cost?

It depends on volume and utilisation. See [self-hosted LLM vs API](/blog/self-hosted-llm-cost-break-even).

### Can we use Cohere privately?

Yes, through private deployment in your VPC or data centre. See [using Cohere with DataNature](/blog/cohere-integration-with-datanature).

## Implementation details

1. **Register model endpoints**: a self-hosted model on your GPUs, a Dlytica-managed model in-country, or an approved provider such as Cohere.
2. **Point the AI gateway** at the default endpoint; until configured it reads "Not configured".
3. **Route by sensitivity**: regulated twins to in-country or self-hosted models, low-risk work to cheaper endpoints.
4. **Keep embeddings inside** with DocuContext and Milvus, so documents never go to an external embedding API.
5. **Enable PII guardrails** to redact sensitive values in answers.
6. **Log usage** per twin and endpoint for cost and audit.
