Google's $920 million monthly contract with SpaceX prices the compute directly. At that rate, Google is paying SpaceX roughly $11 billion a year to run Gemini training on B200 NVL72 hardware outside the capacity ceiling of its own GCP data centers, a ceiling that exists because TSMC's Hsinchu N3P fabs cannot simultaneously clear the combined wafer demand from Nvidia, AMD, and Apple without one queue waiting on another. TSMC's board said so this week, plainly. The US lab procurement stack is now legible: Anthropic on Amazon and Google cloud agreements, OpenAI deploying its $40 billion SoftBank round into Azure capacity, Google contracting SpaceX rather than building its own cluster. All three are bounded by how many N3P wafers TSMC's Fab 18 pulls per quarter, a figure TSMC does not publish but that each contract has already priced into its per-token cost structure.

Baidu, Alibaba, and Huawei Cloud run training on a different constraint map entirely. Wafer allocation is not the constraint. Ascend 910B and the forthcoming 910C operate on SMIC's N+2 process node, approximately 7nm equivalent, paired with domestic HBM alternatives that deliver lower memory bandwidth than SK Hynix HBM3e; the performance gap on transformer attention layers at batch sizes above 2,048 is documented in Ascend MLPerf submission logs. The binding ceiling is SMIC defect density rate and the HBM bandwidth that BIS October 2023 Entity List additions removed from legal reach. AirTrunk's $30 billion India commitment, with its first Bengaluru tranche scheduled for 2027, lands in that gap: as of June 2026, BIS controls govern chip-level exports but not data center operations, and no published BIS instrument covers a facility that is neither US-operated nor PRC-affiliated.