How it actually works

Ask for a capability. We own the vendor.

Under the hood it's the Sovereign Routing Protocol — and it speaks plain OpenAI. Keep your SDK, your code, your tooling. Set model to a capability and sovrgn resolves the right provider behind it, under your jurisdiction's residency rules. The capability is the contract; the vendor is ours to manage — so switching providers, or losing one, never touches your code.

smart

Frontier reasoning — the strongest model for the job.

fast

Low latency — fast, cheap, good enough at scale.

sovereign

In-jurisdiction residency — routes in-country, or fails closed.

Works withOpenAI SDK · LangChain · LlamaIndex— anything that already speaks OpenAI
# Point at your country's instance — nothing else changes.
from openai import OpenAI

client = OpenAI(
    api_key="sk_...",
    base_url="https://au.sovrgn.ai/v1",
)

resp = client.chat.completions.create(
    model="sovereign",            # capability, not a vendor
    messages=[{"role": "user", "content": "Hello"}],
    extra_headers={"X-Sovereign-Residency": "au-only"},
)

print(resp.choices[0].message.content)
It works like an electricity market

Firm your inference, or float it.

Power grids have priced this for a century: firm capacity is reserved, guaranteed and can't be curtailed; interruptible capacity is cheaper because it yields when firm load needs it. Inference is now priced the same way — and sovrgn lets you choose per call, under your jurisdiction.

⬢ Firmed inference

Reserved. Guaranteed. Yours.

Committed, in-jurisdiction capacity with guaranteed throughput and an SLA — the baseload your critical paths run on. It can't be bumped when the market tightens. The price you pay for certainty.

Grid analog: firm supply — agreed capacity, highest dispatch priority, can't be curtailed except under unforeseeable conditions (71% of US power-plant gas, 2016).
Real today: AWS Bedrock Provisioned Throughputhourly commitment · guaranteed units
◇ Floating inference

Spot-priced. Deferrable. Cheap.

Batch and off-peak work that rides spare capacity at a discount — and steps aside when firm demand arrives. Run it overnight, on the cheap side of the planet, while the sun is somewhere else.

Grid analog: interruptible supply — lower priority, short-term, curtailable, priced below firm to reflect the risk of being bumped.
Real today: Bedrock batch −50%Bedrock Flex −50%Anthropic Batch −50%

The market mechanics are the real ones: merit-order dispatch (cheapest capacity first), locational pricing (price tracks where the work lands), and a day-ahead/real-time split — firm reserved ahead, spot cleared on a ~5-minute cadence. Provider tiers and discounts cited are vendor-published and current; they move.

The continuity machinery

We don't just route. We guarantee the route.

"Plan B" only means something if it engages without you. The 2024 Red Sea cable cuts made the lesson physical: routes fail, not just vendors. These are the three mechanisms underneath every sovrgn instance — the parts that turn a bad day for a provider, a cable, or a jurisdiction into a non-event for you.

Automatic failover

Every capability resolves an ordered chain of providers. A retryable upstream failure reroutes to the next candidate transparently — one inbound call, one billed unit.

Circuit breaker

A per-provider breaker trips open after consecutive failures, skips the target that's down, then half-open probes before closing. A bad provider day never becomes your bad day.

Fail-closed residency

Send X-Sovereign-Residency: au-only and sovrgn routes only in-country. No silent offshore fall-through — if nothing in-jurisdiction can serve it, you get a clean, structured rejection.

⬢ Sovereign by definition, not by configuration. The terminal node of every in-jurisdiction chain is a national sovereign-silicon backstop — accredited against the SRP standard. AloomU is Australia's. Each nation plugs in its own.