The White House safety review that held GPT-5.6 past its original launch window did not delay OpenAI's compute trajectory. The Jalapeño inference chip, co-developed with Broadcom and built for OpenAI's own data centre footprint, is already in validation; the architecture offloads Nvidia H100 and B200 NVL72 dependency for inference workloads while Microsoft Azure remains the training substrate. The constraint the review imposed was political, not silicon: Sol, Terra, and Luna tier access is gated to named enterprise partners, which means the deployment surface that funds OpenAI's next training run is deliberately narrow at launch, not broad.

The APAC read is in the partner list, not the benchmark sheet. OpenAI's Singapore Point-of-Presence, which routes API traffic for most of Southeast Asia and handles enterprise contracts across financial institutions in the MAS regulated stack, will determine which regional counterparties get Sol-tier access before the general release window. PRC-domiciled labs running inference on Huawei Ascend 910B clusters in Suzhou and Shanghai cannot access the tier at all: BIS October 2023 controls plus the subsequent May 2024 expanded entity-list additions mean Alibaba Cloud, Baidu, and ByteDance run whatever OpenAI publishes in open weights, and GPT-5.6 is not open weights. The partner gate and the export-control schedule have already decided the distribution question; the White House review decided only the date.