Models on the owner's GPU live
A coding model, a vision model, an embedding model and a safety model, served on the node's own NVIDIA DGX Spark. Nothing leaves the machine unless the owner lends it.
What it does · November 2026
Each node is a complete AI stack its owner runs for their own people. Linked by agreements between owners, nodes lend each other capacity, run each other's tasks, carry each other's messages and watch over each other, under a charter every owner signs.
Live means running between the pilot's two nodes. Tested means built and passing on simulated federations of two to four nodes, but not yet used between real ones: the pilot has only two. Where it stands has the details.
A coding model, a vision model, an embedding model and a safety model, served on the node's own NVIDIA DGX Spark. Nothing leaves the machine unless the owner lends it.
Agents send ordinary chat requests to one address. A router picks the model, a scheduler shares the slots by priority and per-project limits, and a conversation stays where it started.
Facts, conversation turns, documents, an entity graph and code, each in its own memory. Only what is related is added to a request. Work done for a partner leaves no summary behind.
One login in front of every page. The dashboard shows status, projects and limits, live traces, memory, the federation, alerts and tasks.
A suite of 42 checks measures the whole stack daily. Watchdogs follow the network, the gate and the partners.
Made in the owner's own browser and kept on the node only encrypted with the owner's passphrase. It signs every act that is the owner's to make: the charter, consents, endorsements, overrides.
Two owners exchange join cards and confirm their nodes' certificates by voice. Each node pins the other's; there is no certificate authority and no list of members anywhere.
One page, step by step: the owner key, the release, partners, the charter, the node's key, relays, messages and charter checks.
A node can change its key without pairing again: each partner accepts the change. A warning comes 90 days before any certificate expires.
Every node serves its release: the full source, the installer and the images, endorsed by owners with their keys. A newcomer downloads it from a node and checks the endorsements.
Sets up a new DGX Spark from a release, with the models pinned to their exact versions, and upgrades it in place later.
No node takes part unless its owner signed the charter with their own key. Without it, a node gets no work and gives none.
Every new instruction in a model request is judged before it runs, by the node that runs it. A refusal names the principle and the point. If the judge does not answer, the work is refused.
A task is judged before it is accepted, by its purpose or likely direct effect: what it declares, and what its commands and inline scripts would do.
The judge lists every point a request matches, and an absolute one decides. An owner may override a wrong refusal on their own node, signed and visible to partners, but never for an absolute point.
A new version must keep every absolute point of the previous one. It takes effect between two nodes only when both owners sign it.
Messages sealed for the node they go to and signed by the one that sends them. Kept and retried while the other node is away, with receipts, replies and an inbox on the owner's page.
Applications can ask about their own tasks, and the answer stays with the task on both nodes.
A message can travel through relays. If no route is known, the node asks its partners, who ask theirs, up to six hops, to find an owner who receives messages. Only the owner's identifier travels, never the message.
A node gone silent, a model down, a certificate near expiry: its partners notice and tell their owners, encrypted node to node, with no outside messaging service.