Two blocks can both say H100 and refuse to run the same job. Driver, VBIOS and node shape decide whether a 32-node run even starts, so they are fixed at listing and frozen for the window.

Without it: Every quote is a different product wearing the same four characters.

Rail-optimised and non-blocking at 1:1 carries the same GPU count as a 4:1 oversubscribed leaf and clears materially higher, because one of them finishes the training run. No index holds topology constant, so no index can hold price constant either.

Without it: You are comparing a cluster to a pile of cards in the same building.

A quote with no filesystem, no capacity and no per-GPU throughput floor is quoting something else. That is why the buyer who brings their own storage trades a different code, not a discount on this one.

512 GB/s sustained across 256 GPUs. Capacity does not scale — 200 TB is a per-lot listing value.

Without it: The GPUs idle waiting on a checkpoint and both sides argue about whose fault it is.

Naming the operator of record before signature is what makes the escalation path priceable. The desk is a counterparty, not a NOC, and the contract says so in one line.

Without it: The first Xid at 3am becomes a discovery exercise.

Xid 13, 31, 48 and 79, ECC over threshold, NCCL bus bandwidth under the floor and thermal throttle all count as down here. That definition, plus a 45-minute replacement clock and one hot spare per sixteen, is most of the value in the contract.

32 delivered nodes carry 2 hot spares.

Without it: 99.9% availability and a job that never converges, both true at once.

The price is the price, the buyer accepts the cluster before funds release, and the tape records what actually changed hands. There is no cash settlement in this instrument and there never will be.

Without it: You have a derivative on a number two vendors cannot agree on.