Provider tiers

The compute behind Kumulus comes from machines contributed by independent providers. Those machines are not equivalent: a laptop in a living room and a rack in a datacentre both usefully run workloads, but not the same workloads, and not with the same promises attached.

Tiers are how that difference is named.

Flare

Home PCs and small servers, usually behind a home router with no public address and no fixed IP. This is the tier most contributors start in, and the one the network is designed around: a Flare node needs no special hardware, no inbound ports and no datacentre.

Uptime is whatever the household provides. Workloads placed here are chosen to tolerate that — things that can be moved, restarted or rebuilt without anybody noticing.

Stars

Datacentre-grade capacity. Two quite different owners sit in this tier:

• Professional datacentres, contributing capacity on commercial terms.
• C-nodes — crowdfunded mini-datacentres, owned by a community or an institution rather than by a company. Same class of hardware, different ownership story.

Stars nodes have the reliability, the persistent storage and sometimes the GPUs or TEE-capable CPUs that Flare cannot promise. They are where demanding and long-running work goes.

Tutors

Rented capacity from an external cloud, reached through that provider's own API rather than by a Kumulus agent. Highest reliability, highest cost, and least interesting philosophically — a Tutor solves reachability and SLA by not being decentralised.

Tutors exist so that "nobody is available" does not have to mean "your workload does not run". They are a fallback, not the default.

A tier is a badge, not a product

This is the part that most often gets misread. Tiers do not divide Kumulus into separate products with separate interfaces. A single machine can improve — better connectivity, an added GPU, a move from a spare room into a rack — and move from Flare to Stars without re-registering, and without a different console.

That is why the provider console reacts to what a machine reports rather than to a label. A GPU panel appears because the node reports a GPU. Options relevant to demanding workloads appear when the operator opts into them. Nothing is gated on the tier name alone, because the tier name is a summary of facts that are tracked individually — and those facts are what MOSAIK describes and BIDEEW acts on.

What a tier does not tell you

A tier describes hardware, reliability and reachability. It does not describe how much the machine's owner can see.

On any tier, the person who owns the hardware has physical access to it. Stronger isolation between neighbouring workloads is available and is worth having, but no tier currently on offer makes a workload unreadable to the operator of the machine it runs on. Closing that gap needs confidential computing — encrypted memory backed by the CPU, with attestation the customer can verify — and that is not shipping today.

We would rather state that plainly here than let a tier name imply a guarantee it does not carry.