OpenAI and Broadcom Unveil "Jalapeño" Custom AI Inference Chip

OpenAI's first chip is purpose-built for the age of inference

OpenAI and Broadcom on June 24, 2026 unveiled Jalapeño, OpenAI's first custom AI accelerator — a chip designed from the ground up for large language model inference rather than repurposed from an earlier training architecture.openai The processor was delivered to OpenAI CEO Sam Altman and President Greg Brockman by Broadcom chief Hock Tan, marking what both companies called the opening of a multi-generation compute roadmap.openai Engineering samples are already running ML workloads in the lab, including GPT-5.3-Codex-Spark, at production-target clock speed and power.tomshardware

A blank-slate ASIC built in record time

Jalapeño is a reticle-sized ASIC — its compute chiplet estimated at roughly 840 mm², approaching the EUV lithography limit — surrounded by six HBM memory modules, a physical configuration that prioritizes the low latency essential for interactive reasoning products.tomshardware The chip's architecture was optimized around memory movement, networking, and kernel patterns specific to frontier models, with the stated goal of achieving effective utilization close to theoretical peak performance.openai Early internal tests indicate performance-per-watt "substantially better than current state-of-the-art," though the companies have not yet released hard benchmarks; a detailed technical report is promised in the coming months.tomshardware

The partnership between OpenAI and Broadcom was announced in October 2025, targeting ten gigawatts of custom AI accelerators.cnbc What makes Jalapeño's debut remarkable is its development speed: the chip went from initial design to tape-out in just nine months — a turnaround the companies believe is the fastest ever achieved in high-performance advanced semiconductors.openai +1 That pace was aided partly by OpenAI's own models, which were used to accelerate parts of the chip design and optimization process.venturebeat Celestica handled board, rack, and system integration alongside Broadcom's silicon implementation.openai

Cutting inference costs to reshape AI economics

The strategic motive is straightforward: inference is where AI reaches real users, and its cost shapes everything from ChatGPT pricing to API margins.techcrunch Early reporting places the chip's targeted cost savings at roughly 50% per inference token compared with general-purpose accelerators.memeburn +1 OpenAI president Greg Brockman framed the effort as a full-stack play: "By designing more of the stack ourselves, we can serve more intelligence with greater efficiency and keep pushing advanced AI toward broader access."openai

Jalapeño is slated for deployment at gigawatt-scale data centers — including with Microsoft — by the end of 2026.tomshardware Broadcom's Hock Tan described the collaboration as "a fundamental commitment to scaling the physical infrastructure required for the next decade of AI," and confirmed plans for additional chip generations beyond the first.openai With Google, Amazon, and now OpenAI all fielding custom inference silicon, the era of near-total dependence on Nvidia GPUs for AI workloads is drawing visibly closer to an end.techcrunch +1