One grid, ten thousand operators

Operators fall into four rough groups, and they behave very differently. Independent data centers bring the most stable capacity and the highest floor price. Render and VFX studios bring enormous burst capacity with strong seasonality. Research clusters are cheap and reliably available overnight in their local timezone. Individual operators are the smallest slice and the most volatile, but they reach metros no one else covers.

Together they make a supply curve that no single provider can reproduce, because most of this hardware would never have been sold as cloud capacity at all. It was bought for something else and is being resold in the gaps.

When a job arrives, the scheduler builds a candidate set of every node that could physically run it, then scores each one. Price matters, but so does measured interconnect bandwidth between candidate nodes, the operator's rolling completion record, distance to the data the job will read, and whether the customer has asked for execution proofs — which not all hardware can produce.

In practice the cheapest node wins about a third of the time. More often the scheduler pays a small premium for a node that will not drop the job at hour nine of a twelve-hour run.

If you do not own the hardware, you need some reason to believe the work happened as specified. ROBO offers three levels, and the right one depends entirely on what a wrong answer would cost you.

Attestation is nearly free and covers the overwhelming majority of jobs: the node signs a hardware-rooted statement of what image it ran and what it produced. Replication reruns a sample of the work on unrelated nodes and compares. Zero-knowledge proofs of execution are expensive and slow, and are used almost exclusively where a result settles money on-chain.

On a network of independently owned machines, a node disappearing is not an incident. It is Tuesday. An operator's power contract changes, a studio takes its farm back for a project, a residential connection drops. The median node is attached for 61 days and unattached for 9 before returning.

So the scheduler treats departure as expected behaviour rather than failure. Operators signal drain when they can, which earns a reliability credit; jobs checkpoint on a cadence tuned to the node's own history rather than a fixed interval; and the placement score prices an operator's unannounced departures directly into their future bids.

The network has capacity almost everywhere. What it does not have is your data everywhere. For training runs against a multi-terabyte corpus, the dominant cost of a bad placement is not a worse GPU price — it is moving the corpus.

ROBO's answer is to treat datasets as first-class residents of the network. A dataset declares the regions it may live in; the scheduler replicates hot shards toward capacity rather than dragging jobs toward storage, and refuses placements that would violate a declared residency boundary.