How a GPU moves from one tenant to the next

A block is not a login. Moving it is a small project with a fixed order, and the order exists so that nobody pays for something that was not delivered.

The six steps

Consent, up to two days. The data centre has already agreed that its customer may sublet. The sub-tenant is confirmed against that agreement.

Screening, three to ten days. Who the buyer is, who owns it, and whether the hardware may go to them under export rules. Longer on Asian routes.

Partition and wipe, two to three days. The nodes come off the previous tenant. Fabric partitioned, disks wiped, image and firmware pinned to what was listed.

Attestation, hours. The serials, firmware and driver are read and compared to the listing. The network is measured: an all-reduce across the nodes has to clear a stated floor. Storage throughput is measured. The operator's health feed is connected to the record.

Acceptance, forty-eight hours. The buyer runs its own test inside the tenancy. If it fails, the buyer walks and no money moves.

Record. Delivery is hashed and timestamped, funds release, the trade prints.

What counts as down

A node is down when it throws a listed fault class, when memory errors pass a threshold, when the network drops below the floor, or when it throttles on heat. One spare node in sixteen is held, and a failed node is replaced within forty-five minutes. Those numbers are in the contract, so a buyer knows what they bought.