$17.5 billion is what Amazon borrowed this week. The debt offering, priced June 10, directed proceeds at what the filing describes as "cloud and artificial intelligence capacity," meaning Trainium2 cluster expansion at AWS northern Virginia and Oregon data centers against capacity pre-committed to Anthropic under the November 2023 partnership agreement obligating up to $4B in staged investment. OpenAI filed a confidential S-1 with the SEC the same week, converting five years of Microsoft Azure deployment data into a public equity narrative at a reported valuation above $300B. Google's pricing memo to enterprise Gemini 2.5 Pro customers, circulated around June 9, set floor rates that position Gemini directly against OpenAI's enterprise API pricing.

The structure resolves in Singapore, at Marina Bay data facilities where the compute geography becomes a balance-sheet line. OpenAI, Anthropic, and Google each run the same conversion: enterprise API contract revenue into capital-market access into NVIDIA B200 NVL72 hardware that BIS Export Administration Regulations Item 3A090, issued October 2023, prevents from shipping to PRC addresses. Baidu, Zhipu, and 01.AI train on Ascend 910B clusters at Suzhou facilities, without a public equity runway or enterprise deployment base at comparable scale. Beijing's targeting of US AI policy debates, noted in an OpenAI threat-intelligence memo circulated this week, is information-environment pressure. The capital-structure pressure runs the other direction. The OpenAI S-1, once public under SEC Regulation S-K review, puts the bilateral compute asymmetry on a prospectus that Goldman Sachs, Morgan Stanley, and every institutional investor holding PRC AI exposure will need to price against their existing book.