OpenAI and Broadcom Unveil 'Jalapeño,' a Custom LLM-Inference Chip Designed in Nine Months and Targeting Deployment by the End of 2026
OpenAI's first custom silicon, the Broadcom-built Jalapeño inference accelerator, went from design to tape-out in nine months and targets deployment by the end of 2026.
Overview
OpenAI unveiled its first piece of custom silicon on June 24, a large-language-model inference chip named Jalapeño that was designed and manufactured in collaboration with Broadcom, according to TechCrunch. The chip went from initial design to manufacturing tape-out in nine months, a pace MLQ News describes as “the fastest ASIC development cycle for high-performance semiconductors.” The companies are targeting initial deployment by the end of 2026.
The move pushes OpenAI deeper into its own infrastructure stack, building the hardware beneath the models rather than renting it. The partnership behind the chip was officially announced in October 2025, as TechCrunch noted.
What We Know
Jalapeño is what OpenAI calls its first “Intelligence Processor,” an AI accelerator designed for large language model inference, according to Pulse 2.0. It is an inference part — the kind of silicon that runs already-trained models in response to user prompts — rather than a training chip. The Decoder reports the chip was designed from scratch around OpenAI’s understanding of LLM inference rather than adapted from a general-purpose design, and that a technical report has been promised but not yet released.
The development timeline is the headline claim. MLQ News reports that the chip “went from initial design to manufacturing tape-out in nine months,” and that engineering samples are “already running machine learning workloads in the lab at production target frequency and power.” Constellation Research notes the lab workloads include OpenAI’s own GPT-5.3-Codex-Spark model — the coding model OpenAI first deployed on Cerebras hardware earlier this year.
The two companies split the work along their strengths. OpenAI handled the chip design, while Broadcom contributed the silicon implementation and networking technology, including its Tomahawk networking chips, according to The Decoder. Constellation Research adds that Broadcom and Celestica will integrate the chip with board, rack systems, networking and production systems. OpenAI also used its own models to accelerate parts of the design and optimization process, Constellation Research reports.
On performance, the claims so far are self-reported. Pulse 2.0 reports the chip is “expected to deliver performance per watt substantially better than current state-of-the-art systems,” and Constellation Research says the architecture “reduces data movement and balances compute, memory and networking.” The chip is designed to work with current and future LLMs across the industry, not just OpenAI’s own, according to Pulse 2.0. MLQ News reports the program “targets roughly 50% cost savings compared with typical AI GPUs for inference workloads.”
Jalapeño is the first part of a much larger program: MLQ News reports the broader collaboration covers “10 gigawatts of OpenAI-designed AI accelerators,” with “racks of accelerator and network systems” targeted to begin shipping in the second half of 2026 and the full buildout completing by the end of 2029.
Executives framed the launch as a long-term bet on owning more of the stack. Greg Brockman, OpenAI’s president and co-founder, said “Jalapeño is part of our long-term full-stack infrastructure strategy,” according to Pulse 2.0. In separate remarks reported by MLQ News, Brockman added: “By designing more of the stack ourselves, we can serve more intelligence with greater efficiency.” Broadcom President and CEO Hock Tan called the chip “just the beginning of a multi-generation roadmap,” per Pulse 2.0, which also quotes OpenAI hardware program lead Richard Ho saying “Jalapeño will efficiently execute our most important workloads.”
What We Don’t Know
OpenAI has not released a technical report, so independently verifiable specifications are scarce. The Decoder cautions that the performance-per-watt claims are self-reported and lack independent verification. Process node, die size, memory capacity, and clock figures were not disclosed in the materials reviewed.
The customer split is also only partly clear. The Decoder reports that Microsoft is expected to buy 40 percent of the chips, but that figure was not confirmed across the other sources reviewed.
Analysis
Jalapeño slots OpenAI into a now-familiar pattern among hyperscale AI buyers: designing custom inference silicon to reduce dependence on merchant GPUs. The emphasis on performance per watt and inference cost, rather than raw training throughput, signals where the economics now bind — serving models at scale, not training them. Whether the nine-month design-to-tape-out claim translates into competitive deployed hardware will depend on the technical details OpenAI has yet to publish and on how the 10-gigawatt buildout holds up against a 2029 completion target.