Contribution units
The reward formula: units follow tokens served weighted by layer share, then scale by performance tier.
How compute rewards are computed.
units += (tokens / 1k) × (node_layers / total_layers) effective = units × perf_tier × gateway_bonus infra : hub uptime/hr > gateway uptime/hr (summed on top)
One unit ≈ 1k tokens served, weighted by the node's layer share of each inference — a node running 12 of 49 layers earns 12/49 of each inference's units. The tier multiplier then rewards measured speed, and infra roles accrue hourly uptime on top.
Worked example
One inference, four nodes
| Node | Layers | Share | Tier | Effective units / 1k tokens |
|---|---|---|---|---|
| GPU | 15 / 49 | 0.306 | S ×1.5 | 0.459 |
| CPU | 12 / 49 | 0.245 | B ×1.0 | 0.245 |
| NPU | 12 / 49 | 0.245 | B ×1.0 | 0.245 |
| Phone | 10 / 49 | 0.204 | C ×0.7 | 0.143 |
- Rewards follow real work: a node that served nothing earns nothing, regardless of uptime (compute roles).
- Roles stack — one machine can be compute + gateway + hub, and its streams sum.
- Everything settles in KVR to the node's own owner wallet; the dashboard shows raw × tier = effective and a claimable balance.