Pillar 05 · Algorithmic

On September 3rd Nvidia released PAIR, the Personal AI Router. Free, open source, announced at IFA. It discovers the computers on your home network over mDNS, pairs them with a 6-digit code, and routes AI inference jobs to whichever machine is sitting idle. Your gaming PC, your kid's laptop, a Mac in the kitchen. One private cluster. Jobs go to free nodes. Nodes join and leave gracefully. If someone fires up a game, PAIR routes around them.

Sound familiar? It should if you've seen the SkyeGXU architecture. Here's ours in three sentences. A Coordinator orchestrates the work through CSR, client-side reconstruction, meaning shards only ever come back together inside the client's own boundary. Federated, stateless agents pick up chunks over the LAN, do their piece, and hand back results. Agents join, agents leave, the Coordinator adapts, and no single box is a bottleneck or a point of failure.

That's the compute plane, and PAIR just rebuilt it almost line for line. The difference is what's inside the chunks. Our agents receive custody objects. Encrypted on the client before anything moves, erasure coded so no agent ever holds a reconstructable copy, keys the agents never see. An agent in our fabric is a worker with amnesia. It computes, it returns, it retains nothing usable. We documented this pattern a while ago, and it sits inside our patent pending filings covering data custody and the ML lifecycle. Nvidia just validated it with a global launch and a keynote slot in Berlin.

I'm genuinely pleased. When a four trillion dollar company ships your architecture diagram, you stop having to explain the diagram. That's marketing budget I didn't have to spend.

Now the part they skipped.

PAIR secures the pipe. mTLS between nodes, mutual authentication, encrypted transport. Good. Table stakes. And then the job lands on a node and the data sits there in plaintext. The prompt, the context, the documents you fed the model, the output. All of it, readable, on whatever machine happened to be idle.

On a home network that's fine. Your threat model is your teenager. Nobody is serving a production order on the PC in your basement.

But look at where Nvidia is pointing this. The same IFA week featured their Span partnership: outdoor nodes with sixteen RTX PRO 6000s bolted next to your electrical panel, a hundred-home pilot, expansion planned for 2027. The trajectory is obvious. This pattern doesn't stay in the den. It scales to businesses, then agencies, then nations. Nvidia has been selling "sovereign AI" to more than 20 countries this year using exactly this distributed logic.

And at that scale, the missing layer stops being a footnote.

Transport encryption protects data in motion. It does nothing for data at rest on the node. It does nothing when the node is seized, subpoenaed, compromised, or simply owned by an operator subject to a foreign disclosure law. Every node in a PAIR-style cluster holds complete, reconstructable, readable data. The orchestration is distributed. The custody is not. There is no custody.

The fix is boring and it's math, and you already read it three paragraphs up. Encrypt on the client, code the shards, keep the keys away from the operator, run the ML lifecycle against custody objects. Then a seized node yields ciphertext fragments and a compelled operator has nothing to produce. That's the difference between distributing your compute and actually controlling your data.

Nvidia built the router. The vault is still missing.

To be clear, PAIR is good software solving a real problem, and I'd tell anyone with two idle RTX cards to go install it. Nvidia is very good at what they do. Keep the silicon, keep the orchestration, keep the ecosystem. Just don't confuse a secure channel with data custody, because the vendors selling this pattern upmarket absolutely will.

The compute plane is now a commodity. The custody plane is where sovereignty lives.

We know. We built it, and we filed on it.

Ross Norrie is Founder and Architect of SkyeConnex, the Data Sovereignty as a Service company.