Nvidia's own quarterly filing puts the AI Compute Partnership program — the one it stepped back from on 08-28 — at $36 billion of total commitments, structured as take-or-pay/revenue-share agreements with a typical six-year term. This is the first dollar figure on the specific program, distinct from the $105bn OpenAI guarantee and $500bn+ third-party platforms already disclosed in the same 10-Q: if utilization ran at 80% of capacity Nvidia would take a cut of the excess revenue, and if it ran at 20% Nvidia's take-or-pay obligation would require covering the shortfall itself. ⚠️ Reached via pickups of a WSJ story (reporters Anissa Gardizy and Berber Jin, 08-27) rather than the filing text directly re-verified here. (SEC EDGAR, NVDA 10-Q) What actually drew the antitrust concern, reported for the first time: Nvidia told some cloud providers they could rent the guaranteed capacity only to Nvidia-approved customers, and said it preferred capacity spread across several smaller AI companies rather than concentrated with one large operator. That is customer allocation by a dominant supplier, not a pricing term — courts treat the former with materially more suspicion, which is the specific legal exposure behind the vague "control" language in the 08-28 entry. Firmus Technologies (170,000-GPU Batam, Indonesia campus, up to $30bn of offtake — the underlying deal itself dates to 06-30 and isn't new) is named alongside SharonAI as an initial program signer now exposed to the pause. ⚠️ Same WSJ-derived sourcing chain as above; no primary Nvidia statement beyond the 08-28 spokesperson quote. (businessmodelanalyst.com, summarizing WSJ)