The gap: Broadcom and OpenAI's new inference chip targets deployment cost, not the training constraint that actually gates PRC labs, and Korea's $550B pledge lands on exactly that mismatch.

OpenAI and Broadcom announced a jointly designed inference accelerator on July 1, aimed at cutting the per-query cost of serving GPT-class models rather than expanding training throughput. That distinction matters on the PRC side of the ledger: Alibaba and Baidu run inference at scale on Ascend 910B clusters domestically, but Ascend 910B lacks a verified HBM3-equivalent bandwidth figure matching Nvidia's B200 NVL72, and PRC labs remain bottlenecked at the training stage by BIS October 2023 controls on the H100/H800 class part, not at the serving stage. An inference chip solves a cost problem neither Anthropic, OpenAI, nor Google currently has priority-ranked above training capacity. Seoul's answer runs the other direction: the $550 billion chip package announced this week routes through Samsung and SK Hynix fabs in Icheon and Pyeongtaek, aimed squarely at HBM and foundry capacity, the exact layer PRC labs cannot buy from Seoul either, since HBM stacking equipment falls under the same multilateral control regime Washington has pushed Seoul to enforce since 2023.

The Broadcom-OpenAI chip and the Icheon-Pyeongtaek buildout sit on opposite sides of the same constraint: one lowers the cost of running models that are already trained, the other expands the physical capacity to train the next generation. PRC labs get neither. Ascend 910B's inference economics stay unverified against B200 NVL72 in any published benchmark, and Samsung's HBM roadmap, contracted through 2029 per this week's announcement, sits entirely outside the export channel Beijing can access. The next test is whether Seoul's package includes a carve-out for legacy-node capacity sales to PRC foundries, the one lever that has historically leaked around BIS controls; Korea's Ministry of Trade has not published implementing rules, and until it does, the $550 billion figure is a ceiling, not a delivery schedule.