OpenAI tests Broadcom “Jalapeno” AI chip, a reported 50% cost-savings claim puts pressure on data-center suppliers

OpenAI is testing AI chips built by Broadcom, according to a market report published Tuesday, in a sign that the AI hardware arms race is shifting toward tighter cost control for large-scale model training and inference.

The report says OpenAI’s testing involves a Broadcom chip referred to as “Jalapeno.” It also attributes a headline number to the trial: the chip reportedly shows about 50% cost savings. In this context, “cost savings” typically refers to lowering the total operating cost per unit of compute, including power, datacenter infrastructure, and hardware efficiency, though the report does not detail how that percentage is calculated.

Broadcom, meanwhile, is expected to move these chips into wider data-center use later this year, the report adds. If confirmed, that timing would matter because AI capex plans are increasingly tied not just to raw performance, but to what it costs to run models at scale.

For Broadcom, AI accelerators and the supporting infrastructure are a strategic focus as cloud providers and model developers seek alternatives or complements to incumbent chip ecosystems. Even when performance differences are incremental, a large reduction in cost per workload can influence which suppliers win new orders.

The market report did not provide specifics on configuration, benchmark scope, power consumption targets, or whether the savings apply across multiple model sizes. It also does not clarify whether OpenAI would adopt “Jalapeno” as a primary compute platform, or whether it remains part of a broader evaluation alongside other chip options.

Sector-wide, the shift toward cost-efficient AI hardware is becoming a common theme. As companies scale inference to serve user demand in real time, margins are increasingly sensitive to datacenter energy costs, throughput, and hardware utilization, not just training speed.

Still, the public disclosure in the report is limited. Broadcom and OpenAI did not provide additional figures or engineering details in the cited post, and it is not clear what the “50%” claim is measured against, whether it is based on internal prototypes, or whether it will hold after production tuning and datacenter integration.

Investors and customers will likely watch for confirmation from official company updates and for any specifics that would allow comparison with competing accelerators, including deployment timelines, production readiness, and the practical cost basis for the reported savings.