OpenAI filed the Jalapeño inference chip into production this week alongside Broadcom, and the announcement landed as a compute-independence claim: a custom ASIC designed to run GPT-class inference without Nvidia H100 or B200 NVL72 hardware. The gap between that claim and the deployment record is the read. Broadcom's ASIC tape-out schedule runs twelve to eighteen months from commit to volume silicon, and OpenAI has disclosed no hyperscale colocation agreement in APAC -- not in Singapore's Tuas data-centre corridor, not in Tokyo's Inzai cluster, not in the Hsinchu-adjacent capacity Nvidia and TSMC supply-chain partners have pre-positioned. The Jalapeño is a roadmap chip. OpenAI's inference load this quarter still runs on Azure's B200 NVL72 racks.

The White House request to delay GPT-5.6's release sits in the same structural frame. The delay is not a safety intervention in any named risk class -- no named evaluator, no named benchmark threshold, no export-control item number attaches to it. It is a sequencing signal: the administration prefers to control the announcement cadence of frontier-class US model releases while Broadcom's ASIC production schedule, Amazon's USD 13 billion India infrastructure commitment, and BIS controls on H100 and A100 exports to non-ally jurisdictions all remain unsettled. GPT-5.6 scores above 90 on GPQA-Diamond in OpenAI's own evals and was staged for June release; the delay pushes the public deployment window into Q3 2026, by which time Broadcom and OpenAI expect first Jalapeño silicon for internal testing. Whether that silicon reaches volume before PRC labs running on Ascend 910C clusters close the GPQA gap is the constraint the delay has not resolved.