Inside the Somix Sovereign AI Stack
How a neutral sovereign AI platform is actually wired — from gateway to settlement
Somix Engineering · September 2026
"Sovereign AI" has become one of the most used and least specified terms in the industry. In practice it usually means one of two things: train a national foundation model, or host someone else's model inside national borders. Both are partial answers, and neither addresses the harder question — who controls the system that sits between a nation's users, its data, and the models?
This post walks through how Somix answers that question in code and infrastructure: the layered stack, the MCP tool gateway, the metering design, and what end-to-end verification actually covered.
1. Sovereignty is an architecture problem, not a model problem
A model is a dependency. A platform is an asset. If a country's entire AI capability is an API call to a foreign provider, it has no leverage over routing, retention, cost, or the agent layer that increasingly is the product.
Somix's design decision was to treat sovereignty as a property of the control plane, not of the weights. Concretely, that means:
- The model layer is pluggable — Somix's own model family, global commercial models, open-weight models, and nationally hosted models coexist and are selected per workload.
- The data, compute and knowledge layers are local-first, and are the parts a national deployment actually owns.
- Security, governance and API form an explicit layer rather than being implicit in a vendor's terms of service.
The result is a stack where swapping a model provider is a configuration change, not a migration.
2. The stack, layer by layer
SOMIX Sovereign AI
│
┌─────────────────────────┼─────────────────────────┐
│ │ │
Models Agents Applications
│ │ │
Local Models Government Agents Government AI
Global Models Enterprise Agents Enterprise AI
Open Models Industry Agents Education AI
Custom Models Personal Agents Industrial AI
│ │ │
└─────────────────────────┼─────────────────────────┘
│
AI Platform Layer
│
┌────────────────────┼────────────────────┐
│ │ │
Data Compute Knowledge
│ │ │
└────────────────────┼────────────────────┘
│
Security / Governance / API
│
Local Sovereign Deployment
The top layer is intentionally heterogeneous. Local, global, open and custom models run side by side; agents are scoped by domain (government / enterprise / industry / personal) rather than being one generic assistant. The bottom three layers are what a sovereign deployment owns outright.
This is also the anti-lock-in argument, stated structurally: the layers a nation must control are separated from the layers it can commodity.
3. The MCP gateway: where industrial depth lives
The most differentiated part of the platform is not the chat surface — it is the tool gateway. Somix exposes 260+ engineering and manufacturing tools through an MCP (Model Context Protocol) gateway, including:
| Capability | What it does |
|---|---|
| Material query | Material property and selection lookup |
| CAD & DFM analysis | Design for Manufacturability feedback on geometry |
| Compliance screening | Regulatory and trade-compliance checks |
| Supplier search | Sourcing and supplier discovery |
| CNC / manufacturing workflows | Process planning and manufacturing orchestration |
The gateway is also where metering happens. Rather than instrumenting every consumer application, Somix runs a proxy gateway that performs injected, prefix-based tiered pricing — tool classes are mapped to pricing tiers, the gateway resolves the tier from the call prefix, and the platform settles accordingly. Because metering is centralized at the proxy, applications don't have to reimplement billing, and pricing policy can change without touching client code.
Two properties of this design are worth calling out, because they are the ones that break in production:
- Atomic settlement. Credits are deducted atomically at the data layer, so concurrent calls cannot double-spend a balance.
- Failure is free. A failed upstream call does not consume credits. This sounds obvious; it is not, and it is the difference between a metered platform users trust and one they don't.
4. Metering that survives contact with production
Streaming is where naive metering systems lose money. If you charge an estimate up front and never reconcile, long generations are effectively subsidized.
Somix settles streaming responses against actual token usage: the stream is parsed for real usage data, a provisional hold is placed at request time, and the difference is refunded once true usage is known. Non-streaming paths are charged directly. Usage remains visible to the user through usage endpoints and APIs.
The surrounding production engineering, in brief:
- Edge routing with real-client-IP rate limiting — the edge terminates and forwards verified client IPs; application rate limiting keys off the real address, so limits are neither globally shared nor trivially spoofable.
- Identity — OIDC-based authentication, with separate signing keys per subsystem to contain blast radius.
- Data layer — MongoDB for primary state, Redis for fast paths, Meilisearch for retrieval.
- Attribution — server-side conversion events are emitted on payment success and linked back to the originating marketing session.
- Operations — scheduled monitoring (container health, disk, certificate and gateway expiry), daily backups, and periodic cleanup jobs are automated rather than manual.
None of this is novel in isolation. The point is that it exists and is running, which is the actual bar for institutional buyers.
5. What "verified" means here
The National Sovereign AI (NSA) experience layer at somix.io/nsa was verified end to end:
- Agent run lifecycle — a multi-step run progresses through its execution state machine and reaches a completed state, with model routing and credit settlement operating together in the same run.
- Platform services — providers, agents, tasks, store, projects, billing and knowledge endpoints all live and responding.
- Routing and directory — model routing and tool directory endpoints operational.
- Interface — conversation, knowledge, agents, billing and research surfaces render correctly.
Two clarifications, stated deliberately so the claims stay falsifiable:
- The NSA experience layer ships its own built-in tool set. This is distinct from the 260+ tools reachable through the Somix MCP gateway. They are different scopes and should not be conflated.
- Verification confirms the system runs and responds correctly. It is not a claim about model quality benchmarks, nor about signed national deployments.
6. Why neutrality is a feature
For a national buyer, the largest risk is not capability — it is dependency. A platform that is economically or technically aligned with one model provider, one cloud, or one jurisdiction will, over a ten-year horizon, make decisions that are not the buyer's.
Somix's neutrality is structural: models are pluggable, deployment is local, governance is an explicit layer, and the tool gateway is protocol-based rather than proprietary. The commercial model is therefore compatible with the buyer's sovereignty rather than in tension with it.
7. What comes next
The near-term roadmap follows the partner model: deepen national deployments with governments, telecom operators, enterprises and universities, starting from Thailand, and work toward interconnecting sovereign deployments into a trusted network rather than a set of isolated national stacks.
About Somix — Somix is The Sovereign AI Company, building a neutral, layered sovereign AI platform. Current product foundation: a unified AI workspace with multiple models, 260+ engineering and manufacturing tools, a developer API, and the Somix Prime / Swift / Code model family.
Links: https://somix.io ｜ https://somix.io/nsa/
Top comments (0)