Depending on the deal, the hyperscaler takes an IaaS fee and/or a revenue share while the lab acts as a seller of record  and counts the gross ARR. The same token spend can produce different revenue numbers and margins at the lab and at the hyperscalers. While the revenue share is positive for hyperscalers who monetize thes Token as a Service businesses higher than their IaaS businesses, the increased incremental margins that go to hyperscalers ultimately hurt lab EBIT margins (who book this cost in Sales & Marketing). Compute follows the same logic: indirect workloads run on hyperscaler MWs but serve lab models, so the capacity allocation needs deal-level detail to be right. (2/4)