OpenAI Unveils ‘Jalapeño,’ Its First Custom AI Chip — an Inference Accelerator Built with Broadcom

OpenAI now has its own silicon. On June 24, 2026, OpenAI and Broadcom unveiled Jalapeño 🌶️ — OpenAI's first custom AI accelerator and, the companies say, the first chip in a multi-generation compute platform they are building together. It is not a consumer part you drop into a PC. It is a datacenter-class processor designed for one job above all others: inference, the work of actually serving models like ChatGPT, Codex, and the OpenAI API to users in real time.

What Jalapeño actually is

OpenAI is calling Jalapeño its first "Intelligence Processor." Technically it is an ASIC — an application-specific integrated circuit — co-designed by OpenAI and manufactured by Broadcom, with Celestica helping industrialize the surrounding system: boards, rack integration, high-performance networking, and scalable production.

The key word is purpose-built. Unlike a general-purpose GPU that gets adapted to AI work, Jalapeño is described as a blank-slate design for modern LLM inference — shaped by the exact systems OpenAI runs every day across ChatGPT, Codex, the API, and its future agentic products, while still being built to serve current and future large language models across the wider industry.

Why inference is the right thing to optimize

Training a model and serving it are two very different problems, and Jalapeño targets the second one. When you type a prompt into ChatGPT, the datacenter runs a burst of computation that ingests your context and then assembles the answer token by token. That process is bound less by raw arithmetic and more by fast memory access, low latency, and the ability to juggle huge numbers of concurrent requests efficiently.

A chip built specifically for LLM serving can optimize exactly those pressure points — throughput per request, latency, chip-to-chip communication, power draw, and ultimately the cost of every single answer. And that cost matters more every quarter: as models grow and agentic workflows fire many model calls per task, inference — not training — is becoming the dominant, recurring cost of running AI. Owning the inference silicon is how OpenAI tries to bend that curve in its favor.

Designed from a blank slate — in about nine months

The most eye-catching claim is the speed of development. OpenAI and Broadcom say Jalapeño went from initial design to manufacturing tape-out in roughly nine months — potentially the fastest development cycle ever achieved for a high-performance advanced ASIC.

OpenAI President Greg Brockman told CNBC the chip was designed end to end in that window with help from OpenAI's own models, which were used to accelerate parts of the design and optimization process. In other words: OpenAI used AI to help build the chip that will go on to run its AI.

Jalapeño at a glance

Performance, deployment, and scale

Early testing, the companies say, shows Jalapeño delivering performance per watt substantially better than the current state of the art — a metric that matters enormously when you are filling buildings with accelerators and paying the power bill on every one.

Initial deployment is targeted for late 2026, "expanding in the years ahead." It slots into the broader OpenAI–Broadcom collaboration announced earlier to deploy 10 gigawatts of OpenAI-designed accelerators, with gigawatt-scale data centers rolling out alongside Microsoft and other partners beginning in 2026. Jalapeño is explicitly framed as the first generation in a multi-generation platform — not a one-off.

An ASIC bet against Nvidia's pricing power

Industry watchers immediately read Jalapeño as a shot at Nvidia. An ASIC is less flexible than a GPU — you can't easily repurpose it for arbitrary workloads — but it is cheaper to run and can be tuned tightly for a specific task, here LLM serving. If OpenAI can serve a meaningful share of its inference on its own silicon, it reduces its dependence on scarce, expensive GPUs and gains real leverage on cost. That economic logic is why the announcement put a spotlight on the balance of power in AI infrastructure.

What it means for ChatGPT, Codex, and AI agents

For end users, nothing changes overnight. But over time, cheaper and more efficient inference tends to show up as faster responses, higher rate limits, lower API prices, and the headroom to run more capable agents that make many model calls per task.

For the chatbot and agent ecosystem built on top of OpenAI, the takeaway is structural: the cost and availability of inference — the thing every chatbot pays for on every message — is increasingly something OpenAI controls down at the silicon level. That has long-term implications for anyone whose product is essentially a wrapper around someone else's model.

FAQ

Is Jalapeño for training or inference? Inference — serving already-trained models to users. It is not pitched as a training chip.

Can I buy one? No. It is datacenter hardware for OpenAI's own infrastructure, not a consumer or retail product.

Does it replace Nvidia GPUs? Not entirely. It complements them and reduces OpenAI's reliance on them for inference, rather than wiping GPUs off the menu.

When does it ship? Initial deployment is planned for late 2026, scaling in the following years.

Who built it? OpenAI designed the chip; Broadcom builds the silicon; Celestica helps with board, rack, and system integration.

Bottom line

Jalapeño marks OpenAI's shift from buying compute to designing it. A purpose-built inference ASIC, taped out in roughly nine months with help from OpenAI's own models, aimed squarely at the single biggest recurring cost in modern AI. If the performance-per-watt claims hold up in production, it reshapes both OpenAI's own economics and the competitive map of AI hardware — and turns up the heat, fittingly, on everyone else in the chip race.