Dragon 1 hosts your vaults and runs the compute on them, under on-chain commitments validators inspect. The cluster sees ciphertext. You hold the keys.
# deploy an agent to the cluster$ npm i @cef-ai/vault-sdk$ cef agent deploy --cluster dragon1 resolving manifest CID ............ ok binding vault (16/48 shards) ...... ok spawning V8 isolate ............... ok cubby ready ....................... <10ms✓ live on Dragon 1 testnet
Dragon 1 runs production AI workloads on its own GPU fleet, in-jurisdiction, next to your encrypted data. Not rented from a hyperscaler. Not shipped to someone else's region.
Running on Zettabyte infrastructure, wrapped as elastic GPU for inference and training.
Inference runs on the same nodes that hold your ciphertext. Your data never leaves the cluster to be computed on.
Dragon One for enterprise with custom components, Dragon Two for the community. One codebase, separate repos.
// Production workloads run on this fleet today. Pricing and capacity by request.
Content-addressed shards on the same nodes that run the compute. Ember computes, cyan stores and verifies. Everything client-keyed.
Every file splits into 48 erasure-coded shards, replicated to 20 peers. Lose 32 and it still reconstructs. Content-addressed by CID.
Agents execute next to the data, with Cubby memory that persists across sessions.
Events are the single ingress shape, over NATS JetStream and Redis Streams. Deployed Ansible to K3s.
Data is encrypted and client-keyed. The cluster never holds the keys.
Reed-Solomon erasure coding splits every object into 48 shards across the cluster, replicated to 20 peers. Any 16 reconstruct the whole. Lose 32 and your data is still intact.
Dragon 1 runs in UAE on Zettabyte hardware. Dragon 2 and 3 follow in other countries, so data stays where the law fits.
| vault | Wallet-keyed data root + event dispatch | base |
| agent-runtime | Agent code in V8 isolates | base |
| cubby | Persistent agent memory, per-agent SQLite | base |
| inference | Model serving on NVIDIA Triton | gpu |
| dac | Data Activity Capture, audited by validators | base |
Obligations are typed records on Cere Protocol, backed by CERE bonds and verified by validator swarms. Not marketing language.
| Commitment | What it guarantees | Enforced by |
|---|---|---|
| data availability | Stored shards stay retrievable within latency bands. | DA checks |
| compute integrity | Agent code runs as published, on the inputs given. | audits |
| SLA conformance | Uptime and latency targets per tier. | bonds + slashing |
| transparency | Cluster state and usage are observable and reconcilable. | DAC + settle |
// SLA targets are governance parameters and may evolve. User refunds are a marketplace-layer concern, not Dragon 1.
Deploy agents with @cef-ai/vault-sdk and pay per execution that touches storage or compute.
Stake CERE, join the cluster on-chain through the ddc-staking and ddc-nodes pallets, and earn a share of usage.
Dragon 1