OpenAI's Jalapeño inference chip, taped out at TSMC's Hsinchu facility on N3 process node and manufactured in volume through Broadcom's supply agreement, scored 3.1 TFLOPS per watt on INT8 inference workloads against Nvidia's H100 at 2.0 TFLOPS per watt on the same measure. The deployment claim is real. The independence claim is not: every Jalapeño die exits a fab OpenAI does not own, in a city the BIS October 2023 advanced-node controls were specifically written to keep out of PRC reach, on a process node that TSMC's Hsinchu campus runs and its Arizona fabs are not yet qualified to replicate at volume. Broadcom's role in the supply chain is packaging and co-design, not wafer production. OpenAI controls the architecture; TSMC controls the substrate.

The bilateral implication is not a forecast. BIS's October 2023 Entity List controls restrict Hsinchu-taped N3 derivatives from transfer to PRC-domiciled entities, which means Jalapeño's inference efficiency advantage -- the margin that would make it worth deploying at Alibaba Cloud or Tencent Cloud scale -- is structurally unavailable to either of those operators. Baidu's Kunlun 3 and Huawei's Ascend 910C, both running on SMIC 7nm and Huawei's own 6nm derivative respectively, are the inference substrate PRC hyperscalers actually run. On the eval that matters for enterprise API throughput (tokens per second per dollar at batch size 512), Kunlun 3 benchmarks published by Baidu in March 2026 show 18% lower throughput than H100 at equivalent power draw; Jalapeño, if the Broadcom co-design brief holds, closes most of that gap on the US side while the PRC side's access to the process node that would close it -- TSMC N3 -- remains blocked by BIS Supplement No. 2 to Part 742, with the next review window opening in Q4 2026.