Skip to main content
Four nouns carry the entire model. Learn these and every screen in both consoles reads the same way.

​
Operator

An operator is one independent datacenter. It owns its own admin accounts, its own customer tenants, its own servers, and its own capacity pool. No other operator on the deployment can see any of it. Operator accounts are code-gated: NeoChamber issues the first admin invite code during onboarding, and that code names the operator, so redeeming it creates the datacenter and makes the redeemer its first admin.
Inside a datacenter, an operator account is effectively root. Everyone you add sees every one of your tenants’ machines and telemetry and your whole capacity pool. What they cannot see, ever, is another operator. There is no narrower role within an operator yet — so keep your own list short.

​
Tenant

A tenant is the isolation boundary inside a datacenter — one per customer company. Every tenant belongs to exactly one operator. Machines reach a tenant two ways:
  1. The customer rents whole nodes themselves from your live availability.
  2. You reassign a node to them from its detail page.
Either way the assignment is immediate, and telemetry starts from the moment of assignment.

​
Node

A node is one GPU server. It belongs to your operator permanently — there is no destination outside it — and sits either in your capacity pool (rentable, never visible to customers) or assigned to one tenant. Each node holds its own credential, issued at enrollment. That single fact explains a lot of the platform’s behavior: revoking an enrollment token never stops a machine that already installed with it, and deleting the teammate who enrolled it changes nothing about the node. A node’s GPUs are identified by UUID, which is the identity anchor for hardware-change detection. A vanished UUID raises a gpu_missing event; a new one is flagged for explicit swap acknowledgment.

​
Agent

chamberd is the agent on every GPU server. One static binary — no Docker, no runtime, no dependencies on the machine.
  • Reads GPU state every 15 seconds and streams it out.
  • Speaks outbound HTTPS on port 443 only. No inbound access is ever needed.
  • Buffers 24 hours of telemetry on disk and backfills with original timestamps after an outage, up to a 55-minute replay cap. Gaps older than that are shown as gaps, honestly.
  • Runs under a hardened systemd unit with Restart=always and a watchdog.

​
How they fit together

Every metric and log line is attributed to its tenant and its operator by the platform itself — never by anything the agent or the browser sends. Every query has its scope injected from the session. See Isolation for how that is enforced and proven.

​
The lifecycle, end to end

1

An operator account is created

From an invite code that names the datacenter. See Create your account.
2

An enrollment token is issued

Named, revocable, revealed exactly once. See Enrollment tokens.
3

Servers are enrolled

One command per box, the same command fleet-wide. Nodes appear in Fleet within two minutes. See Install the agent.
4

Tenants are created and invited

One tenant per customer company, with single-use member invites. See Tenants.
5

Customers rent from the pool

Whole nodes by GPU model, assigned on the spot. See Rent compute.