OpenAI Says Its New Chip Outperforms Nvidia's Blackwell As Nvidia Prepares Earnings Release

OpenAI has unveiled its first in-house artificial intelligence accelerator, codenamed Jalapeño, co-developed with Broadcom and Celestica. According to self-reported benchmarks, the custom silicon delivers 1.5x to 1.9x higher inference throughput and up to 3.6x lower latency compared to Nvidia's Blackwell-generation systems. However, volume production is not expected until 2027, and the accelerator is dedicated strictly to inference workloads, leaving Nvidia's dominance in AI training unaffected. The announcement arrives just as Nvidia prepares to release its fiscal second-quarter financial results. Wall Street consensus projects Nvidia's quarterly revenue between $91 billion and $92 billion—roughly doubling year-over-year—with earnings per share expected around $2.09. Despite consistently beating expectations over the past three quarters, Nvidia's stock has dropped following four of its last five earnings reports. While tonight's earnings figures will not be impacted by OpenAI's announcement, investors remain focused on long-term headwinds as hyperscalers increasingly develop in-house chips, posing a multi-year risk to Nvidia's GPU market share.