An ML node is the piece of a Gonka participant's infrastructure that actually touches GPUs. Under the hood, a participant runs a small stack of cooperating services:

• ML nodes do the GPU work — generating proof-of-compute batches during the Sprint, and executing inference (plus validating other participants' inference) the rest of the time.
• Decentralized API nodes orchestrate: they manage a participant's fleet of ML nodes, track epoch phases, relay proof batches to the chain, and accept user inference requests.
• Chain nodes run the blockchain itself.

When an epoch's proof-of-compute window opens, a participant's API node fires a start event to every ML node it manages, seeded with a block-derived random value. When the window closes, ML nodes switch to serving real inference traffic and, on a sampled basis, validating other participants' submitted proofs and completions.

The count of ML nodes across the network's active participants is one of the figures GNKScan surfaces directly (labeled "Compute Nodes" on the inference page) — as of early July 2026, on the order of 196 nodes across roughly 35 active participants. That is distinct from the GPU count, since a single ML node typically runs on more than one physical accelerator.

For where ML nodes fit into the wider architecture, see What Is Gonka?