Hock Tan walked into OpenAI's San Francisco headquarters on June 24 carrying a chip. The Broadcom president and CEO handed the first engineering samples of Jalapeño to Sam Altman and Greg Brockman in person, a piece of silicon that did not exist in any form nine months earlier.
That is the part worth sitting with. Nine months from a blank sheet of paper to a working processor running real machine learning workloads is not how the semiconductor industry operates. A high-performance chip normally takes two to three years from architecture to tape-out, the point where a finished design is sent to a factory. OpenAI and Broadcom say Jalapeño went from initial design to tape-out in nine, and that they believe it is the fastest such cycle ever achieved in advanced semiconductors.
They also say something stranger helped make that possible: OpenAI used its own AI models to accelerate parts of the chip's design and optimization. The models OpenAI serves to customers helped build the hardware that will one day serve those same models.
The Chip Does One Thing on Purpose
Jalapeño is not a general-purpose accelerator. It is built for inference only, the process of running a finished model to answer a user's request, as opposed to training, which is the far more compute-hungry work of building the model in the first place.
That narrow focus is the whole point. Training is where Nvidia's H100 and B200 GPUs are hardest to displace, because training rewards flexible, general-purpose compute. Inference is different. Inference is the bill that arrives every single day, every time someone types into ChatGPT or fires off a Codex task. OpenAI runs that workload at a scale almost no one else does, which means it understands exactly what the silicon underneath needs to do.
"Jalapeño was designed from the ground up for LLM inference using detailed insights from our close collaboration with OpenAI researchers," said Richard Ho, who leads OpenAI's hardware program. "We optimized the architecture around the kernels, memory movement, networking, and serving patterns that matter most for frontier AI models."
In plain terms, the chip is shaped around how large language models actually move data, not how a generic AI workload from five years ago moved data. OpenAI calls Jalapeño a "blank-slate design," informed by the systems it runs across ChatGPT, Codex, and its API. The engineering samples are already running models in the lab at production frequency and power, including an unreleased model called GPT-5.3-Codex-Spark.
The Number Nvidia Will Notice
OpenAI's official language about performance is careful. The company says only that early testing shows Jalapeño will deliver performance per watt substantially better than current state-of-the-art, with a full technical report promised in the coming months.
Broadcom's chief executive was blunter. Hock Tan told Reuters and Bloomberg that the chip can serve tokens at roughly half the cost of current-generation Nvidia GPUs, with raw performance matching Nvidia's Blackwell parts and Google's TPUs. That gap between OpenAI's measured caution and Broadcom's confident framing is worth noting, because one of those statements is a verified lab claim and the other is a sales pitch from the company that builds the silicon.
Either way, the direction is clear. Inference cost is the lever that decides whether a frontier AI company makes money. OpenAI reportedly spends billions of dollars a year serving ChatGPT on rented GPUs. Cutting the per-token cost of that work is not a marginal engineering win. It is the difference between a business that loses money on every query and one that does not.
This is the same logic that pushed Amazon and Anthropic into a $25 billion deal built around Amazon's own Trainium chips, and that has Microsoft pitching its Maia accelerator to external customers. Every lab with a serving bill is trying to own more of its own stack.
How OpenAI Got Here
OpenAI's chip ambitions are not new. The Broadcom partnership was announced in October 2025, and reports of OpenAI finalizing a custom chip design go back further than that. Jalapeño is the moment the plan turned into hardware.
The roles in the project break down cleanly. OpenAI designed the chip architecture. Broadcom handled the silicon implementation and the networking that ties many chips together, including its Tomahawk switching silicon. Celestica builds the boards, racks, and systems that turn loose chips into a deployable data center.
Why This Is About the Full Stack
Greg Brockman framed Jalapeño as one piece of a larger strategy. "Jalapeño is part of our long-term full-stack infrastructure strategy to make compute more abundant, resulting in AI which is faster, more reliable, more affordable for people and businesses," the OpenAI president and co-founder said.
The phrase that matters is full stack. OpenAI already builds the models, the products on top of them like Codex, and increasingly the data centers that run them. Designing the chip underneath is the last layer. When a single company controls the architecture, the kernels, the memory system, the networking, and the product, every layer can be tuned toward the same goal of serving a model cheaply.
That control is also a hedge. OpenAI is preparing for a public offering, having filed confidentially for an IPO earlier this year, and the single largest variable in its financial story is compute cost. A chip that meaningfully cuts that cost is a chip that makes the entire business look more sustainable to investors.
The Skeptics Have a Point
A first chip is not a finished one. Matt Bryson, an analyst at Wedbush Securities, called the announcement a "probable positive" for Broadcom but cautioned that compute chips usually need several generations of refinement before they see wide adoption. Initial shipment volumes, he warned, may be modest. The hardest engineering often comes after the first sample works, not before.
The market reaction reflected that ambivalence. Broadcom's stock moved only slightly on the news, and Nvidia fell just 0.26%, a shrug rather than a sell-off. Investors appear to have decided that Jalapeño is a real but bounded threat. Nvidia still owns training, where flexible GPUs are hard to beat, and OpenAI has confirmed Nvidia remains a key partner for that work. A blank-slate inference chip from a company that has never shipped silicon at scale has to prove itself in production, not in a lab demo.
There is also the question of dependence. Jalapeño does not free OpenAI from outside partners. It trades some reliance on Nvidia for reliance on Broadcom, Celestica, and the foundries that fabricate the chip. The full stack is more OpenAI's than it was, but it is not OpenAI's alone.
The Bottom Line
For working engineers, the relevant fact is not the corporate strategy. It is the trend line. The cost of running a token through a frontier model is becoming the central battle of the AI industry, and the labs are answering it the same way: by designing their own silicon. Google has its TPUs. Amazon has Trainium. Microsoft has Maia. Now OpenAI has Jalapeño, with Nvidia's general-purpose GPUs still dominant in training but facing real competition everywhere inference happens.
The most quietly radical detail in the whole announcement is the one OpenAI mentioned almost in passing. Its models helped design the chip that will run the next generation of its models. If that loop holds, the cost of building better chips falls, which lowers the cost of compute, which trains better models, which design better chips. OpenAI is betting the loop holds. The technical report, due in the coming months, is where that bet gets graded.
Sources
- OpenAI and Broadcom unveil LLM-optimized inference chip (OpenAI) (Jun 24, 2026)
- OpenAI unveils its first custom chip, built by Broadcom (TechCrunch) (Jun 24, 2026)
- OpenAI and Broadcom Unveil LLM-Optimized Intelligence Processor (Broadcom) (Jun 24, 2026)
- OpenAI unveils first chip as part of Broadcom deal in effort to 'build the full stack' (CNBC) (Jun 24, 2026)
- OpenAI and Broadcom announce first custom AI chip, in strike at Nvidia (Reuters via Yahoo Finance) (Jun 24, 2026)