📌 OpenAI Jalapeño Rumor Drives Samsung HBM4 Demand

OpenAI is reportedly advancing its in-house inference chip, Jalapeño. What merits attention is not a single performance figure, but rather that AI infrastructure procurement may continue to extend from general-purpose accelerators to custom chips, advanced process nodes, and high-bandwidth memory.

According to related reports, Jalapeño is a collaboration between OpenAI and Broadcom and is expected to use TSMC’s N3P process. The reports also state that this 700W chip achieved per-user token throughput of more than 700 tokens/sec in DeepSeek R1 testing, with output per megawatt higher than NVIDIA Vera Rubin. If these parameters are ultimately verified, OpenAI’s direct benefit would be hardware better suited to its own inference workloads to meet computing demand; Broadcom and TSMC could in turn benefit from custom design and manufacturing orders, respectively.

The key supply-chain variable is memory. Another report says Jalapeño is configured almost entirely with Samsung HBM4. HBM is memory that provides high-speed data transmission for AI computing, and whether chip performance can be realized also depends on this configuration and supply capacity. If the proposed solution is adopted and actual shipments materialize, Samsung may gain a new source of HBM4 demand; however, this does not mean revenue has already been generated, as order volume, supply timing, and memory usage per chip all remain to be confirmed.

TrendForce expects memory to account for 68% of cloud service providers’ capital expenditures next year. This forecast can only indicate that memory is becoming a major expenditure item in AI expansion; it cannot directly prove that Jalapeño will generate corresponding procurement. What still most needs to be verified at present is the chip’s launch, performance testing, manufacturing progress, and HBM4 configuration. If the rumor has not received official confirmation, it likewise cannot be used to infer an actual impact on NVIDIA’s share or suppliers’ revenue.

▌ Sources

• Jalapeño performance and N3P manufacturing rumor by StockSavvyShay
• Jalapeño may use Samsung HBM4 by jukan05
• Memory may account for 68% of cloud providers’ capital expenditures by jukan05