The question every donor should ask: what does it cost to run this intelligence for a million people? Our answer is structural. Active inference agents are sparse and small — they run where the community lives, on modest local hardware, speaking to each other through the Spatial Web. There is a real compute bill in this plan, and it belongs to research, not to scale.
Large language models are compute-hungry because every answer passes through billions of parameters in a central data center — cost grows with every user, forever. An active inference agent works differently. Grounded in the Free Energy Principle, it maintains a compact generative model of just the variables that matter inside its Markov blanket — for a community agent, that is hundreds of variables, not billions of parameters. It updates beliefs as events arrive rather than recomputing the world each time. VERSES’ enterprise deployments already run this class of agent on ordinary server CPUs.
The claim held to the maturity hierarchy: FEP is established theory; sparse active-inference agents at community scale are an engineering bet with strong precedent, not a shipped commodity. That is exactly why the research track below exists — and why it is firewalled from the live network’s cost structure.
Coherence Labs runs a protected simulation environment: a network of synthetic agents carrying the candidate mathematics of coherence — generative models, factor graphs, the equations under test. Pilot data enters only under consent, is held inside the firewall, and never leaves raw. Here Unity is simulated before it circulates (the cadCAD discipline), coherence measures are stress-tested, and the Framework earns its v1. This is where the real compute bill lives: ~$140k over 24 months — GPU time for simulation sweeps, secure storage, reproducibility infrastructure — inside the Labs budget.
Each community runs its own lightweight agent on local hardware — a small server, eventually a steward’s laptop. Agents hold only their community’s model, communicate peer-to-peer over Spatial Web protocols (HSTP), and keep working when the internet doesn’t — which matters on citys and in remote territories. There is no central inference bottleneck to pay for, defend, or trust. Refined models flow outward from Labs as published updates; raw community data never flows back without consent. ~$3k hardware + ~$6k/yr per CD.
Curves are illustrative shapes, not forecasts — the point is structural: in a distributed mesh, adding the millionth member adds a node to a community, not a rack to our data center.
The Spatial Web standard (IEEE 2874, ratified 2025) was designed so that every entity — a person, a community, a watershed — is a first-class holder of its own data with programmable permissions. We inherit sovereignty from the architecture instead of bolting it on:
Each member’s contributions, exchanges, and care records live in their own wallet; each community’s ledger lives on its own node. There is no master database of human lives at CoH.
HSML expresses who may see what, for what purpose, for how long — machine-readable consent. A community grants Labs access to anonymized patterns for research; it can revoke that grant, and the revocation is enforced by protocol, not by policy.
A community that leaves takes its node, its ledger, and its history with it, intact. For the indigenous partnership this is foundational: sovereignty that survives disagreement is the only kind worth the word (CARE principles as reference standard — see Legal).
Data sovereignty is not a product line — it is a property of the system every community gets by default. Anything less would make protection a privilege.
| Line | What it is | Y1 | Y2 |
|---|---|---|---|
| Labs simulation compute (firewalled) | GPU sweeps, secure storage, reproducibility — the research bill | $60k | $80k |
| CD local nodes | Hardware + connectivity: CD 1 in Y1; four more nodes in Y2 (incl. indigenous CD, provisioned for offline-first) | $8k | $34k |
| Interface hosting & services | Wallet + dashboard delivery, monitoring, backups — boring cloud, kept small | $36k | $70k |
| VERSES / Genius engagement | Scoped, severable; accelerates the agent build while native capacity grows Review | $180k | $80k |
| Infrastructure total | ~4% of the round — the structural proof the model scales | $284k | $264k |
Compare: a centralized real-time inference architecture for 65,000 members would carry a $300–500k annual compute line that grows with every community — and a single point of failure, control, and subpoena. The mesh costs less and cannot be captured. That is not a coincidence; it is the same design decision seen twice.
The question that decides whether 2031 is architecture or fantasy. Yes — for one reason: nothing in this system has to know about everything. A community’s agent models its own members and bioregion; it never needs the state of the other 59,999 communities. Adding the sixty-thousandth community adds no load to the first.
| Scale | Communities | Members / node | Commons infrastructure | What is centralized |
|---|---|---|---|---|
| 2027 · pilot city | 1 | 5,000 | ~$280k/yr | Labs research firewall only |
| 2028 · four cities | 5 | ~13,000 | ~$320k/yr | Labs + Starter Pack distribution |
| 2029 · self-serve | 2,000 | ~1,000 | ~$650k/yr | Labs + published model updates |
| 2031 · 100M | ~60,000 | ~1,700 | ~$3.5M/yr | Nothing that holds member data. Ever. |
Node hardware and hosting are carried locally — often on donated, municipal or cooperative infrastructure — which is why the commons line grows sub-linearly. The costs that would explode in a centralized design never appear, because that data never comes to us.
At every scale there will be a good argument for one big database, one inference cluster, one canonical registry — faster matching, cleaner analytics, easier compliance. Each would work. Each would convert a commons into a platform and coherence measurement into surveillance. The distributed mesh is the architecture of a promise.