OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first custom inference accelerator, on Wednesday, with Broadcom chief executive Hock Tan telling Bloomberg the ASIC delivers roughly 50 percent cost savings against standard AI GPUs. Tan and Charlie Kawwas, Broadcom’s president of semiconductor solutions, hand-delivered the first engineering samples to Sam Altman and Greg Brockman in a staged photo op that did most of the narrative work the press release didn’t.

The co-development timeline is the headline buried beneath the headline: nine months from initial design to tape-out. That’s aggressive even by hyperscaler-ASIC standards, and it tells you how compressed OpenAI’s runway feels relative to its Nvidia dependency, which remains intact for training workloads.

Engineering samples are already running ML workloads at production target frequency and power, including GPT-5.3-Codex-Spark. OpenAI described early performance-per-watt results as “substantially better than current state-of-the-art” and promised a detailed technical report in the coming months. Translation: trust the framing now, audit the silicon later.

Initial deployment is scheduled for the end of 2026, scaling at gigawatt scale alongside Microsoft and other partners, with volume building in 2027 and full tilt in the first half of 2028. Celestica is handling board and rack system integration.

Broadcom shares are up 10 percent year-to-date in 2026, and Tan, speaking to CNBC, called compute demand from his six custom-silicon customers “simply insatiable.” The word choice is the point. Hyperscaler capex has graduated from cyclical to structural in the executive lexicon, which is the kind of vibe shift that historically precedes either a generational platform or a 2000-style overshoot. The bet baked into Jalapeño is that this time it’s the former.

Sources