# Subnets – first-principles view ## 1. Core idea A **subnet** is a self-contained slice of a distributed compute network. Every node inside that slice shares: - **Hardware similarity** – same qubit count, fidelity, connectivity, or other specs that matter to the workload. - **One service interface** – one algorithm family or API all nodes agree to run. - **Local governance** – its own rules for scheduling, result verification, and rewards. The global network is partitioned so each slice can specialise and optimise for a single task class. ## 2. How a subnet forms 1. **Partition hardware** Group nodes with matching performance profiles. Homogeneity lets compilers and runtimes push the hardware harder. 2. **Bind to one workload** The subnet contract states the job type it serves (e.g. search, optimisation). Sticking to one algorithm lets the stack pre-tune circuit depth, error mitigation, and scheduling. 3. **Route jobs by label** A dispatcher tags each job and sends it to the matching subnet. Other subnets never see the request, keeping queues short and latency low. 4. **Reach consensus locally** Nodes run the circuit. Validators in the subnet check fidelity, agree on the output, and release payments. Slashing and rewards stay inside the partition. 5. **Scale by adding partitions** A new algorithm or hardware type? Spin up another subnet. Existing ones keep running unchanged. ## 3. Why this design works | Goal | Subnet advantage | | ------------------ | ------------------------------------------------------------------------------------ | | Performance | Specialised code paths and error models push hardware limits. | | Isolation | Data and risk stay inside one partition. | | Economic clarity | Clear price curves per workload; operators know which hardware pays best. | | Regulatory control | Compliance rules can differ subnet to subnet without complicating the whole network. | | Modularity | New capabilities arrive as new subnets, not as sweeping upgrades. | ## 4. Key takeaway A subnet is a workload-focused micro-cloud with its own hardware profile, API, consensus, and incentive logic. Partitioning the global network this way gives performance, security, and flexibility without heavy central coordination.