OpenAI is no longer just a company that makes models. As of yesterday, it makes the silicon that runs them. OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom-designed AI chip, and the headline number is the part that made the whole industry sit up: it went from initial design to manufacturing tape-out in just nine months, which the two companies say may be the fastest high-performance ASIC development cycle ever achieved. For context, custom chips like this normally take two years or more. OpenAI did it in nine months, and it used its own AI models to help do it. The chip is an inference processor, built from scratch for one job: running large language models like ChatGPT faster and cheaper. Early lab samples are already humming along running one of OpenAI's own coding models. The deeper story underneath the engineering flex is strategic, and it points straight at Nvidia, the company OpenAI has depended on for every model it has ever trained or served. Here is what Jalapeño actually is, why nine months matters, and what it means for the AI arms race.

Jalapeño is a custom inference chip, and the distinction matters. There are two phases in an AI model's life: training, the massive one-time job of building the model, and inference, the constant job of running it every time someone sends a prompt. Jalapeño is built only for inference, the everyday work of answering billions of ChatGPT and Codex requests. OpenAI is calling it an "Intelligence Processor."

It is what is known as an ASIC, an application-specific chip, designed from a blank slate around exactly how OpenAI's models behave, rather than a general-purpose GPU adapted to the task. That focus is the point. By stripping it down to inference and architecting it around its own workloads, OpenAI says Jalapeño reduces the wasteful data movement that bottlenecks conventional GPUs and runs much closer to a chip's theoretical peak performance. Early engineering samples are already running ML workloads in the lab at production targets, including one of OpenAI's own models, GPT-5.3-Codex-Spark, and OpenAI says early testing shows performance-per-watt substantially better than the current state of the art. A full technical report is coming in the months ahead.

It is the first chip in a multi-generation platform built with Broadcom and manufacturing partner Celestica. First deployment comes at the end of 2026 at gigawatt scale, with Microsoft confirmed as a primary partner, reportedly required to commit to buying 40% of the first production run.

The speed is not just a brag. It signals something bigger about how fast AI hardware can now move, and it ties directly to a theme we covered earlier this month.

OpenAI President Greg Brockman told CNBC the chip was designed end to end in nine months, and credited the company's own AI models with the acceleration. His words: "the degree to which our models have been able to accelerate it was very surprising to us." Read that again, because it is the genuinely remarkable part. OpenAI used AI to help design the chip that will run its AI. This is the same recursive loop Anthropic described in its "AI builds itself" piece this month, where Claude now writes most of Anthropic's code. Here it is again, in silicon. AI is now compressing the development of the very hardware it depends on.

The implication is large. The entire industry runs on Nvidia's roughly two-year hardware release cadence. If AI-assisted design can collapse a two-year-plus chip cycle into nine months, the pace at which new custom silicon arrives could accelerate dramatically, and the companies that own that feedback loop, AI helping design the next chip that trains the next AI, start compounding an advantage that pure hardware buyers cannot match. Nine months is not a one-time stunt. It is a preview of a faster clock for the whole field.

OpenAI has been one of Nvidia's single largest customers since it kicked off the AI boom in 2022. That dependence is both expensive and risky: you pay Nvidia's premium margins, and you are exposed to its supply constraints. Jalapeño is OpenAI's move to own that part of its destiny. By designing its own inference chip, OpenAI can target the cost of running its models, the single biggest recurring expense in its business, and stop routing all of that spend and supply risk through one vendor.

This also matters enormously for OpenAI's finances right as it heads toward a public offering. OpenAI is losing money on the sheer cost of compute. If Jalapeño meaningfully lowers the cost of inference, it is a direct path toward better unit economics and eventually profitability, which is exactly the reassurance public-market investors will want to see in the S-1. As one analysis put it, OpenAI is trying to rewrite the future unit economics of AI by controlling the physics of its own inference pipeline.

The honest caveat: this is not OpenAI dumping Nvidia. Jalapeño is inference-only. The heavy lifting of training next-generation models will still lean on Nvidia, AMD, and others, and OpenAI just took a $30 billion investment from Nvidia in February as part of a deal to deploy 10 gigawatts of its systems. So the real picture is OpenAI diversifying and clawing back leverage and cost control, not severing the relationship. It is a strike at Nvidia's pricing power, not at Nvidia's existence. And nothing ships at scale until late 2026, so the impact is a 2027 story.

If you build on AI, the thing to watch is inference cost. If OpenAI's custom silicon drives down the price of running models, that pressure eventually flows to everyone through cheaper API pricing, as the whole industry races to match its unit economics. The era of inference being dominated by one expensive vendor is starting to crack, and that is good for anyone whose product runs on tokens.

For the strategic game, this confirms the defining move of 2026: every major AI company is racing to own its full stack. SpaceX owns its compute through Colossus. Microsoft built its own MAI models. Anthropic uses Claude to write Claude. Now OpenAI designs its own chips. The companies that control models, software, data centers, and silicon together are building moats that pure-play rivals cannot dig. Vertical integration is the strategy of the year, and Jalapeño is its sharpest expression yet.

And the big picture is the loop at the center of this. OpenAI used its own AI to help design a chip in record time, a chip whose entire purpose is to run more AI, faster and cheaper, so it can build better AI. That flywheel, AI accelerating the creation of the infrastructure for more AI, is the quiet engine under everything happening this year. Nine months is the first time we have seen it turn the crank on physical hardware. It will not be the last.

OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom-built AI chip, a purpose-built inference processor for running large language models like ChatGPT faster and cheaper. The standout: it went from design to manufacturing tape-out in just nine months, potentially the fastest high-performance ASIC development cycle ever, and OpenAI used its own AI models to accelerate the design, with President Greg Brockman saying the speedup surprised even them. Early lab samples already run OpenAI's GPT-5.3-Codex-Spark with substantially better performance-per-watt than current hardware. Deployment begins late 2026 at gigawatt scale, with Microsoft committing to a reported 40% of the first production run. Strategically, it is OpenAI's move to cut its dependence on Nvidia, control the cost of inference, and improve its unit economics ahead of its IPO, though the honest caveat is that it is inference-only, training still leans on Nvidia, and nothing ships at scale until late 2026. The bigger pattern: every major AI company is racing to own its full stack, from models to data centers to silicon, and OpenAI using its own AI to design its own chips in record time is the clearest sign yet of the flywheel driving this entire era. AI is now building the hardware that builds AI.

Stay building. 🤖