Products · NovaTerra
Your AI, on a node you own.
NovaTerra is a personal AI operating system: a Twin that keeps your memory and files, a Muse you talk to, and a Studio that builds with you. This page is the technical approach, with each part labelled built, in progress or planned.
Status: deployed in production, invite-only. Sign-up is not open yet; novaterra.world has a waitlist.
1 · A personal cloud
One node, one owner.
Your data lives on a node that belongs to you, at home or hosted, and it talks to other nodes without a middleman that can read it.
- Built
One node per person
A NovaTerra install has exactly one owner, enforced in the database. It runs at home on the owner's machine (Windows and Linux installers), or hosted as one machine and one volume per member, with the same software either way.
- Built
Identity is a key you hold
Every member and AI has a portable Ed25519 keypair, addressed as did:key. A node can be exported under a passphrase and moved to new hardware.
- Built
A relay that cannot read
Nodes exchange signed messages, sealed with X25519, HKDF-SHA256 and AES-256-GCM, through a relay that can delay or drop a message but cannot read, forge or redirect it.
- Built
Hosted nodes back themselves up
Each hosted node backs itself up off-site, under a key only that node holds.
2 · Sharing GPU capacity
Your GPU, and later your neighbours'.
The node already knows its own GPU. Sharing capacity between members is designed, with the missing pieces written down, but not built.
- Built
The node knows its GPU
A node reads its VRAM with nvidia-smi and sizes what it can run locally.
- In progress
A slot for peer compute
Every model call already records where it ran, on a privacy ladder: on this node, on a peer's node (mesh), or on a rented API. The mesh rung is defined in the contracts; nothing fills it yet.
- Planned
Sharing GPU capacity between members
A member with a spare GPU runs verified jobs for others (render, train, background work) and is paid per verified job, never per raw hour. The missing pieces are written down: a peer provider, peer candidates in routing, a job-and-receipt message on the federation wire, and per-job accounting.
3 · Generators
The model is not asked to reinvent the wheel.
Every Studio step runs on one of four rungs. Templates and skills do the known work deterministically; the model is called only where judgement or new writing is needed.
L0
Template
A deterministic render.
No model call
L1
Skill
A deterministic tool.
No model call
L2
Model fills a shape
The structure is fixed; the model writes into it.
Model call
L3
Model invents
Only where nothing known applies.
Model call
- Built
No planning call
A generator run skips the planner model entirely; the plan is the generator.
- Built
Reuse
A matching earlier run is served again at zero cost, with no model call.
- Built
Tested, not assumed
Generators that should make no model call have tests that assert it.
Why it matters for hardware: when most steps are deterministic and the rest fill a known shape, the model's job gets smaller, so smaller models on a single consumer GPU can do more of it. We have not yet published a measurement of the tokens or calls saved across real runs, so this page gives the mechanism and no saving figure.
4 · Where NVIDIA fits
From hosted APIs to a GPU the member owns.
Today most tiers call hosted APIs. The path is to move each tier onto an NVIDIA GPU the member owns or rents, so private data stays on their node.
- Built
Per-tier routing with a local option
Each model tier (chat, embeddings, transcription, speech) is routed on its own. Setting a tier to local points it at any OpenAI-compatible server, such as llama.cpp, vLLM or Ollama, and a local tier never silently falls back to a remote one. The reference local box is an NVIDIA RTX 3090.
- In progress
Quanta for local decode
Quanta, from our lab, compiles a model to PTX for NVIDIA GPUs and checks every kernel bit for bit against a reference. On one RTX 3090 it reached 1.27x llama.cpp on Qwen2.5-0.5B batch-1 decode, and was slower at 7B and on prefill. It is a separate engine, not yet wired into NovaTerra.
- Planned
NVIDIA NIM and TensorRT-LLM as the local-model option
Both speak the OpenAI-compatible interface the local tier already uses, so they slot in as configuration, and TensorRT-LLM becomes a published baseline for Quanta.
- Planned
A home node on RTX or DGX Spark-class hardware
A reference build where a member's whole AI (memory, models and speech) runs on one NVIDIA device they own, so private data stays on their node.
Quanta on NVIDIA GPUs: the result and its limits · The same story in the NovaTerra docs
Questions to jon@vocabotics.com.