Is Northern Inference a sovereign AI service for the Canadian government?

Yes. Northern Inference gives Canadian teams one approved API for the AI models their people use. Select a Tier 3 Managed Cloud route for Canadian data residency or a Tier 4 route for broader access. Tier 3 requests fail closed when the selected Canadian route is unavailable rather than silently rerouting outside Canada.

How does Northern Inference support data sovereignty and PIPEDA?

Tier 3 requests are pinned to Canadian-resident routes and blocked rather than silently rerouted outside Canada. Each routed model request has a chain-of-custody record containing the selected route, provider, jurisdiction, data-handling details, and credential source, plus an integrity digest and Northern Inference's Ed25519 signature when signing is available. Optional PII substitution can replace detected personal information before it reaches the provider. The record attests Northern Inference's selected and recorded path up to the third-party provider boundary, while provider documentation, attestations, and contracts govern execution inside the provider.

Is Northern Inference certified against a Government of Canada security standard?

Northern Inference is working toward a Government of Canada Authority to Operate (ATO), assessed against ITSG-33 at the Protected B (PBMM) profile. This is an in-progress program, not a completed certification.

Which AI models can I route to through Northern Inference?

Northern Inference publishes its currently enabled Azure, AWS Bedrock, and Vertex AI routes in the live model catalog. You select the full route key, which also selects the provider, route tier, region, and jurisdiction.

Now onboarding teams / Canadian residency / AI Chain of Custody

Frontier AI access,without losing control.

Route AI workloads across Canadian and global models through one sovereign control plane, with privacy controls, optional PII substitution, and a verifiable AI Chain of Custody for every request.

Join waitlist Read the docs

Planned service Need GLM-5.2 or another leading open model hosted in Canada?

We are planning Canadian-owned capacity for teams with data-residency, no-US-cloud, or controlled-goods requirements. Tell us which model you need and the minimum monthly usage you could commit after launch.

Requests start at $5,000 in monthly usage after launch. No payment, reservation, or contract is created today.

Next, we review your request and contact you about fit and a written launch proposal.

https://northerninference.ca/v1

You own your data NI processes prompts to complete your request. We do not train on them or keep a copy by default. Provider terms are published per route.

AI Chain of Custody A per-request record of the NI-selected route, provider boundary, jurisdiction, and credential source, signed by NI when signing is available.

Frontier and near-frontier models Claude, GPT, Gemini, Llama and more. The model ID selects provider, tier, and region.

quickstart.py

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://northerninference.ca/v1",
    api_key=os.environ["NI_API_KEY"],
)

response = client.chat.completions.create(
    model="azure/gpt-4.1-mini-ca",
    messages=[{"role": "user", "content": "Hello from NI"}],
    max_tokens=200,
)

print(response.choices[0].message.content)

import os
from anthropic import Anthropic

client = Anthropic(
    base_url="https://northerninference.ca/v1",
    api_key=os.environ["NI_API_KEY"],
)

message = client.messages.create(
    model="bedrock/global.anthropic.claude-sonnet-4-5-20250929-v1:0-global",
    max_tokens=200,
    messages=[{"role": "user", "content": "Hello from NI"}],
)

print(message.content[0].text)

How NI resolves this request Recorded

1 / request Exact route key

azure/gpt-4.1-mini-ca asks for the Canadian Azure deployment.

2 / enforce Policy passes

The API key permits Tier 3. Region, spend, and audit controls pass before provider access.

3 / resolve Azure OpenAI, Canada

The provider call uses an NI-managed credential and returns custody evidence.

Evidence returned: request ID / exact route / provider region / jurisdiction / NI-managed or BYOK credential / custody path

Live routes and tool compatibility

Azure OpenAI AWS Bedrock OpenAI SDK Anthropic SDK Claude Code Codex BYOK credentials

Government and public sector

Built on signed route evidence. Ready for government.

Each routed model request has a chain-of-custody record identifying the route Northern Inference selected, including provider, region, jurisdiction, and credential source. Tier 3 routing fails closed: a request that cannot stay on an eligible Canadian route is blocked, not silently rerouted. The portal and API record carries an integrity digest and, when signing is available, Northern Inference's Ed25519 signature. This attests the NI-selected path up to the provider boundary. Provider documentation, attestations, and contracts govern inference location and data handling inside the provider. See each model's data-use posture.

Canadian public-sector infrastructure

Canadian residency is an enforced route decision.

For a Tier 3 request, NI selects a supported Canadian provider deployment and blocks the call if that residency guarantee cannot be met. It does not silently fall back to another jurisdiction.

Tier 3

Canadian data residency

At Tier 3 we route to Canadian regions such as AWS ca-central-1 and Azure Canada East. The routing layer fails closed when a request cannot be served in the selected jurisdiction.

Custody

Transparent data custody

Each model response identifies the resolved route, provider region, jurisdiction, custody path, and whether the provider call used NI-managed or customer BYOK credentials. These response headers are unsigned route evidence; the portal preserves the corresponding digest and, when signing is available, Northern Inference's signature.

Cited

Data-use posture, sourced

For every model route we publish where training, logging, and retention stand, each claim cited to the provider's own dated documentation with a link to the live source. Browse the posture.

Tiers

Privacy tier model

Select your privacy level by choosing the exact model route. Tier 1 customer-hosted and Tier 2 NI sovereign hardware are planned; Tier 3 Canadian cloud and Tier 4 provider-default access are live.

Opt-in

PII substitution

Opt-in detection and substitution replaces supported identifiers before the provider call and restores them in the response. Operational audit metadata records that substitution occurred without storing the original prompt.

Compliance

Built for federal procurement

Runs on AWS Canada infrastructure that holds SOC 2 Type II, ISO 27001, and PCI-DSS (AWS attestations, not NI's). ITSG-33 PBMM control mapping in progress. Full status on our Trust page.

Pricing

Procurement-friendly pricing

Transparent per-token rates and clear invoicing in CAD. Credit purchases carry the disclosed 9% service fee. Usage, fees, and balance deductions remain separately reviewable.

API

Standard API format

OpenAI-compatible chat completions and Anthropic-compatible Messages endpoints support common SDKs and tools. Feature compatibility depends on the endpoint and selected provider route.

Request a government briefing.

Priority access and dedicated onboarding for government and public sector organizations.

Full name *

Work email *

Department / agency *

Job title

Province / Territory

Level of government

Tell us about your AI use case

We will respond within 2 business days. Privacy Policy

PIPEDA-aligned AWS Canada attestations Tier 3 fails closed Transparent custody

The problem

Enterprise AI is no longer a model problem. It is a control problem.

Capability is abundant. Governance is not. Regulated teams need model access, residency, custody, spend control, and auditability to move together.

01 / fragmentation

Model access is splintering

Each provider, route, SKU, and SDK has its own limits, pricing, release cycle, and operational behavior.

02 / residency

Data location is easy to imply and hard to prove

Teams need the resolved jurisdiction for each request, not just a marketing label attached to a provider.

03 / custody

Provider credentials need attribution

BYOK should preserve customer custody while still producing request IDs, costs, hold behavior, and traceable routing.

04 / drift

Today's best model is not next quarter's

Hard-coding provider names turns routine model churn into application rewrites and security review churn.

05 / spend

Usage controls must live before the provider call

Rate limits, spend caps, model allowlists, and billing holds need to be enforced before tokens are generated.

06 / visibility

You cannot govern what you cannot inspect

Operators need logs, route decisions, pricing freshness, provider status, and alerts in one control surface.

The platform

One inference layer across providers, policies, and route tiers.

Northern Inference turns each model request into a governed path with an exact provider destination and reviewable custody evidence.

Exact route selection

The request model is the route key. It selects provider, jurisdiction, tier, and deployed model without hidden tier headers.

model = azure/gpt-4.1-mini-ca

Tier-aware access

API keys authorize the route tiers they may call. Canadian and global paths remain explicit in product and logs.

Tier 3 / Tier 4

BYOK custody

Attach provider credentials to NI API keys, route through NI, and keep provider billing in the customer's account.

provider credential = customer BYOK

Custody trail

Every routed model response identifies the route, provider region, jurisdiction, and custody path Northern Inference recorded. It also shows whether the provider call used NI-managed credentials or the customer's BYOK credential. Portal records add cost, an integrity digest, and Northern Inference's signature when signing is available.

request IDs / policy hashes

Data-use posture

Per route we publish where training, logging, and retention stand, each claim cited to the provider's own dated documentation with a link to the live source. We attest the route to the provider boundary; inside the provider we cite, not assert. Browse the posture.

cited per route

Alias negotiation

When tools send provider-native names, users can persist a safe mapping to the exact NI route they intend.

account-level mappings

Operator controls

Discovery, pricing freshness, workers, live logs, provider keys, users, billing, and security events remain visible.

admin control plane

One request / evidence at every layer Verifiable

IdentifyAssign the request ID and read the exact route selected by the caller.

EnforceCheck key permissions, tier, model, spend, billing, and optional PII policy before provider access.

RouteReach the exact provider deployment using either NI-managed credentials or the customer's BYOK credential.

VerifyReturn route and custody headers, then preserve cost, policy hash, and signed custody evidence in the portal.

Privacy tiers

Choose your level of trust.

Every deployed model carries a privacy tier. You select it by choosing an exact route key, and your API key authorizes which tiers it may call. No tier headers, no per-request flags.

Tier 1 · Coming soon

Self-Hosted Hardware

Your hardware, your premises, open-source models. NI tunnels API requests into your machine; only you ever see prompts and completions. The strongest privacy guarantee we offer. Phase 5 of our roadmap.

Only you process data

Tier 2 · Coming soon

Sovereign Hardware

NI's own bare-metal, single-tenant, in a Canadian-owned data centre. We hold the keys and no US-owned cloud sits in the path. Sovereignty by control, not just residency. Phase 5 of our roadmap.

NI-owned Canadian hardware, no US-owned cloud

Tier 3 · Live

Managed Cloud

AWS Bedrock and Azure OpenAI routes in Canadian regions. Provider training, logging, and retention posture is published per route from the provider's current documentation.

Cloud provider processes in Canada

Tier 4 · Live

Global Access

Broader model access through NI-routed upstream providers outside the Canadian-residency boundary. This can include direct APIs, Bedrock in US regions, Azure GlobalStandard, Vertex US regions, and other provider-default routes.

Exact route shown per request

Tier 2 / coming

Sovereignty, not just residency.

Residency answers where your data sits. Sovereignty answers whose law reaches it. We are standing up Tier 2 on our own Canadian hardware: bare-metal we own, in a Canadian-owned data centre, with no US-owned cloud anywhere in the path.

Coming

We own the machine

Single-tenant bare-metal in a locked cage in a Canadian-owned facility. Not a VM on someone else's cloud.

Coming

We hold the keys

The planned architecture keeps encryption keys under NI control and removes a third-party cloud provider from the inference path.

Coming

No US cloud in the path

The planned Tier 2 path uses NI-owned hardware in a Canadian-owned facility, reducing reliance on foreign-owned cloud infrastructure.

Coming

True per-tenant isolation

The planned inference engine will isolate prompt and KV-cache state per tenant. The implementation must pass security review before this tier is marked live.

Tier 4 is access. Tier 3 is Canadian residency by contract. Tier 2 is sovereignty by control, the next rung on the ladder. The need is not unique to Canada or to controlled-goods programs: any government, regulated industry, or enterprise that wants control instead of a residency checkbox fits here.

For whom

Built for teams that cannot adopt AI by ignoring reality.

Public sector

Canadian routing for protected workloads

Give teams a sanctioned path for model access with explicit route IDs and operator-visible custody. Tier 3 residency is enforced at the routing layer: a request that cannot be served in a Canadian region is blocked, not silently rerouted.

ca-central-1Tier 3audit

Healthcare

PII-aware application experiments

Use per-key controls, optional PII substitution, and route permissions while teams validate sensitive internal workflows.

PII policykey scopeslogs

Financial services

Model access with reviewable spend

Control model allowlists, token rates, caps, and custody so AI usage can be reviewed instead of guessed.

spend capsbilling holdsrate limits

Engineering and legal

Client-sensitive knowledge work

Keep provider credentials in customer custody with BYOK while NI records routing, request IDs, and fee state.

BYOKcustodymodel mappings

For developers

Compatible endpoints. Policy in the route key.

If your code already talks to OpenAI-compatible chat completions, start by changing the base URL and choosing an exact NI model route.

✓Use https://northerninference.ca/v1 as the API base.
✓Select Canadian or global routing with the exact model route key.
✓Use documented Claude Code, Codex, OpenAI-compatible, Anthropic-compatible, and IDE integrations.
✓See request IDs and custody headers in responses for operator follow-up.

View docs Browse models

your_app.py

from openai import OpenAI

# Before: locked to one provider
client = OpenAI()

# After: NI routes by exact model key
client = OpenAI(
    api_key=os.environ["NI_API_KEY"],
    base_url="https://northerninference.ca/v1",
)

response = client.chat.completions.create(
    model="azure/gpt-4.1-mini-ca",
    messages=[
        {"role": "user", "content": "Summarize this memo."}
    ],
    max_tokens=200,
)

print(response.choices[0].message.content)

from anthropic import Anthropic

# Before: locked to one provider
client = Anthropic()

# After: NI routes by exact model key
client = Anthropic(
    api_key=os.environ["NI_API_KEY"],
    base_url="https://northerninference.ca",
)

response = client.messages.create(
    model="azure/gpt-4.1-mini-ca",
    max_tokens=200,
    messages=[
        {"role": "user", "content": "Summarize this memo."}
    ],
)

print(response.content[0].text)

Trust

Built for security review, not against it.

Northern Inference exposes the evidence operators need: exact route resolution, security events, billing state, provider status, and custody attribution.

Enforced

Session-gated portal and admin

Authenticated shells are served only after server-side session validation.

Visible

Provider and pricing freshness

Operators can inspect discovery, pricing jobs, worker state, and stale sources.

Recorded

Security and audit events

Cross-tenant denials, BYOK lifecycle events, and admin actions are durable records.

Inspectable

Request custody headers

Responses include the NI request ID, resolved route, custody path, and whether the provider call used NI-managed or customer BYOK credentials.

Live on the platform

Available models.

All models available through the NI API with transparent per-token pricing. Choose by provider, tier, or data residency.

Tier 3: Canadian Cloud

Managed Canadian cloud routes. Data residency stays in Canada when the exact route is Tier 3.

Tier 4: Global Access

Provider-default or global routes. Use when model access matters more than Canadian residency.

Built for developers

One API. Every enabled route.

Point OpenAI-compatible or Anthropic-compatible clients at NI and switch provider, route tier, and region with the exact model key.

Standard API format

Supports OpenAI-compatible chat completions and Anthropic-compatible messages. Change the base URL and choose a route from the live catalog.

Frontier and near-frontier models

The live catalog includes enabled Claude, GPT, Gemini, Llama, Mistral, and other routes. The exact model key determines provider, tier, and region.

Transparent billing

See the provider cost and our fee separately on every request. Thinking tokens billed at the rate shown per model. No expiring credits. No surprise overages.

Per-request privacy

Choose the exact route key on every API call. Route sensitive prompts through your own hardware (coming soon) and casual queries through cloud providers.

PII substitution

Names, emails, and identifiers are replaced with realistic fakes before reaching the model, then restored in responses. Opt-in per API key or per request.

Real-time analytics

Live dashboard with per-model, per-team cost breakdowns. Budget alerts and hard limits built in.

Canadian control plane

NI's routing control plane operates in Canada. The exact model route determines the upstream provider region and jurisdiction, which are recorded per request.

Automatic prompt caching

On supported routes, NI can mark stable prompt prefixes for provider caching automatically. Cache behavior, retention, jurisdiction, and price remain specific to the selected provider route and are disclosed in its data-use and pricing records.

How we compare

Built different.

Northern Inference is developer infrastructure, not a consumer chatbot.

	Northern Inference	Venice.ai	OpenRouter	Direct APIs
Per-request privacy control	✓	✗	✗	✗
Canadian data residency (Tier 3 routes)	✓	✗	✗	~
Transparent pricing	✓	✗	~	✓
Standard API format	✓	✓	✓	~
Per-request custody trail	✓	✗	✗	✗
PII substitution	✓	✗	✗	✗
No crypto required	✓	~	✓	✓

Simple pricing

No surprises. No games.

Pay for what you use.

See provider-rated usage cost and any NI per-request fee separately. Credit purchases carry a 9% service fee. There are no monthly minimums, credits do not expire, and thinking-token rates are shown per model when applicable.

Transparent route rates No monthly minimums Credits never expire Thinking tokens priced openly

See live rates for every enabled route →

Private rollout

Get access for your team.

Join the waitlist for portal access, model routing setup, and integration support.

Work email * Full name (optional) Organization (optional) Your role Primary use case Job title (optional)

Frontier AI access,without losing control.

Built on signed route evidence. Ready for government.

Canadian data residency

Transparent data custody

Data-use posture, sourced

Privacy tier model

PII substitution

Built for federal procurement

Procurement-friendly pricing

Standard API format

Request a government briefing.

Enterprise AI is no longer a model problem. It is a control problem.

Model access is splintering

Data location is easy to imply and hard to prove

Provider credentials need attribution

Today's best model is not next quarter's

Usage controls must live before the provider call

You cannot govern what you cannot inspect

One inference layer across providers, policies, and route tiers.

Exact route selection

Tier-aware access

BYOK custody

Custody trail

Data-use posture

Alias negotiation

Operator controls

Choose your level of trust.

Self-Hosted Hardware

Sovereign Hardware

Managed Cloud

Global Access

Sovereignty, not just residency.

We own the machine

We hold the keys

No US cloud in the path

True per-tenant isolation

Built for teams that cannot adopt AI by ignoring reality.

Canadian routing for protected workloads

PII-aware application experiments

Model access with reviewable spend

Client-sensitive knowledge work

Compatible endpoints. Policy in the route key.

Built for security review, not against it.

Session-gated portal and admin

Provider and pricing freshness

Security and audit events

Request custody headers

Available models.

One API. Every enabled route.

Standard API format

Frontier and near-frontier models

Transparent billing

Per-request privacy

PII substitution

Real-time analytics

Canadian control plane

Automatic prompt caching

Built different.

No surprises. No games.

Pay for what you use.

Get access for your team.

Share to move up the waitlist