OpenAI describes o3 as production-ready enterprise software. Its leaked P&L shows the lab losing money against B200 NVL72 training costs at a rate the APAC enterprise API contract book in Singapore and Tokyo cannot close. SpaceX's $60B acquisition of a developer coding platform (Ars Technica, June 2026) names the structural answer OpenAI's income statement does not contain: the interface layer above inference is where the enterprise margin lives. Barret Zoph's exit as VP of Research after five months is the staffing signal on the same arithmetic. A lab that cannot compound o3's 71.7% SWE-bench Verified performance into deployment revenue that services the next training run does not hold senior research talent against a vertical stack that owns the developer IDE.

Google's $2B Jurong data center expansion in Singapore (Q1 2026) runs per-token inference for the APAC enterprise book at margins the San Francisco labs do not capture. They do not own the interface. Baidu's ERNIE 4.0, running on Ascend 910B clusters at the Guizhou facility, faces a symmetric constraint from the PRC side: no equivalent commercial deployment surface to convert inference margin into training capital. The pressure on OpenAI's enterprise book now arrives from SpaceX's interface acquisition. OpenAI's IPO prospectus, expected Q3 2026, must price a training-to-deployment revenue path; SpaceX's $60B acquisition priced it first, and the path runs through the developer workflow, not through SWE-bench rank.