Breaking

OpenAI and Broadcom Unveil Jalapeño, Their First In-House Inference Chip

June 24, 2026 at 12:11 EDT

Infra & Chips
Foundation Models

OpenAI and Broadcom on June 24 jointly unveiled Jalapeño, OpenAI's first AI accelerator designed in-house. It is a custom ASIC optimized for LLM inference, targeting ChatGPT, Codex, the OpenAI API and future agentic products.

June 24, 2026 · OpenAI × Broadcom

OpenAI Unveils "Jalapeño," Its First Self-Designed AI Chip

An inference-focused ASIC built to power ChatGPT, Codex and the API — and to cut the company's reliance on NVIDIA GPUs. Designed by OpenAI, built by Broadcom, integrated by Celestica.

~50%

cheaper than a typical AI GPU, per Broadcom

9 mo

from design to tape-out, accelerated by OpenAI's own models

10 GW

of OpenAI-designed accelerators targeted for deployment

Cost per chip: GPU vs. Jalapeño

Relative cost — claimed roughly half that of a typical AI GPU

100%

Typical AI GPU

~50%

Jalapeño

Performance said to rival NVIDIA Blackwell and Google TPU, with performance-per-watt substantially above current state-of-the-art hardware.

The full-stack pipeline → rollout

DESIGN

OpenAI leads the chip design

→

MANUFACTURE

Broadcom builds & produces

→

INTEGRATE

Celestica handles racks & networking

Late 2026

First deployment

2027

Ramp-up

Early 2028

Full operation

Developers cheer

Eliminates GPU waste via inference specialization
Striking nine-month development pace
Accelerates OpenAI's vertical integration

Skeptics caution

Custom ASIC vs. NVIDIA's versatility & CUDA ecosystem
No independent benchmarks yet at the sample stage
Core counts, memory and pricing undisclosed

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.