Jun 26, 2026

OpenAI Unveils Jalapeño: Its First Custom AI Chip for LLM Inference

OpenAI and Broadcom have unveiled Jalapeño, OpenAI's first custom-designed AI chip, an inference-only accelerator. This application-specific integrated circuit (ASIC) is built for large language model (LLM) inference, with initial deployment targeted for late 2026.

Compute & infrastructure

OpenAI unveils Jalapeño, its first custom AI chip

Last updated: 26 June 2026

On 24 June 2026, OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom-designed AI chip — an inference-only accelerator the companies say went from initial design to manufacturing tape-out in nine months. OpenAI disclosed no specifications, benchmarks, or cost figures; a technical performance report is promised in the coming months.

Disclosure: this story concerns OpenAI, a competitor to Anthropic; this newsroom uses Anthropic's Claude in production.

TL;DR

Jalapeño is OpenAI's first chip: an ASIC built only for large language model inference, co-developed with Broadcom and revealed 24 June 2026.
OpenAI says it reached tape-out in nine months — what it calls the fastest such cycle it believes has been achieved.
Engineering samples are running in the lab, including the GPT-5.3-Codex-Spark model. No specs or numbers were released.
OpenAI claims performance-per-watt "substantially better" than state-of-the-art, but disclosed no figures; an audited report is months away.
Initial deployment is targeted for end of 2026 at gigawatt scale with Microsoft and other partners.

What was announced

OpenAI and Broadcom jointly unveiled Jalapeño on 24 June 2026, calling it OpenAI's first "Intelligence Processor." Broadcom President and CEO Hock Tan and semiconductor chief Charlie Kawwas physically delivered the chip to OpenAI's Sam Altman and Greg Brockman. It is an application-specific integrated circuit (ASIC) — silicon tuned for one job rather than a general-purpose GPU — and that job is inference: running already-trained models to answer user requests in ChatGPT, the API, and agentic products.

OpenAI says it designed the chip from scratch around LLM inference, with Broadcom supplying silicon implementation and networking (including its Tomahawk networking silicon) and Celestica handling board, rack, and system integration. The company also says its own models helped accelerate parts of the design.

What's confirmed vs. what isn't

The announcement is heavy on milestones and light on data. The table separates what OpenAI stated from what it withheld.

Disclosed	Not disclosed
Name (Jalapeño), purpose (LLM inference), partners (Broadcom, Celestica)	Any benchmark or throughput number
Nine-month design-to-tape-out timeline	Cost per chip or cost per token
Lab samples running, incl. GPT-5.3-Codex-Spark	Specs: process node, memory, die size
End-2026 deployment target, gigawatt scale	How it compares to specific Nvidia or AMD parts

On efficiency, OpenAI says only that early testing points to performance-per-watt "substantially better than current state-of-the-art" claim, with a detailed report to follow. No "cheaper-by-X-percent" figure was released by either company; any such number circulating in coverage is not from OpenAI or Broadcom.

Why a chip, and why now

Inference is where the running costs sit: every ChatGPT reply and every API completion is an inference event, and OpenAI pays for that compute at enormous volume. A chip purpose-built for inference can lower cost per token, which translates either to higher margins at the same price or lower prices at the same margin. Greg Brockman framed the logic plainly: "By designing more of the stack ourselves, we can serve more intelligence with greater efficiency."

The move also loosens OpenAI's reliance on Nvidia, whose GPUs have powered nearly all of its products to date. OpenAI has separately signed compute deals with AMD, AWS, and Cerebras; Jalapeño pushes that diversification down to the silicon itself. The companies target initial deployment by the end of 2026, at gigawatt scale with Microsoft and other partners, over multiple chip generations projection.

The Broadcom angle

For Broadcom, Jalapeño is continuation, not departure. It is the custom-silicon partner of choice for hyperscalers: it co-designed Google's TPUs and works with Meta and ByteDance on bespoke accelerators. Adding OpenAI gives Broadcom a foothold with the most-watched name in frontier models. Hock Tan told CNBC that demand across Broadcom's custom-silicon customers is "simply insatiable," and signaled the ramp is gradual — a small prototype phase late in 2026 before scaling through 2027 and 2028. None of that displaces Nvidia at the frontier of training; it chips away at the inference layer, which is where the volume — and the bill — lives.

Performance and cost claims here are OpenAI's own and unaudited; the company has released no benchmarks and promises a technical report later. Figures labeled claim are vendor-stated; projection marks a stated future plan, not a shipped product. Verify against the vendor before relying on this.

Sources

OpenAI, "OpenAI and Broadcom unveil LLM-optimized inference chip" — 24 June 2026 (name, nine-month tape-out, lab samples, efficiency claim, deployment target, quotes).
CNBC, "OpenAI unveils first chip as part of Broadcom deal" — 24 June 2026 (ASIC framing, Tan's "insatiable" demand, ramp timeline, October 2025 deal).
TechCrunch, "OpenAI unveils its first custom chip, built by Broadcom" — 24 June 2026 (inference focus, Nvidia-dependence context).
CNN Business, "OpenAI just announced its first custom chip" — 24 June 2026 (IPO context, infrastructure strategy).

Last updated: 26 June 2026 · First published: 26 June 2026