AINEWS
The math behind OpenAI's $1.4 trillion chip bet
Jalapeño, the company's first custom AI processor built with Broadcom to serve ChatGPT queries, is a vertical integration move aimed at a 33 cents on the dollar margin gap with Nvidia.
Jalapeño, the company's first custom AI processor built with Broadcom to serve ChatGPT queries, is a vertical integration move aimed at a 33 cents on the dollar margin gap with Nvidia.
OpenAI keeps roughly 33 cents of every revenue dollar after operating costs, per reporting on the company's inference economics. Nvidia, the dominant supplier of the high-end AI processors that train and run frontier models, is widely estimated to keep about 75 cents of every dollar on its top chips. That gap is the structural pressure behind every AI capex headline in 2026, and it is the reason OpenAI built Jalapeño, the company's first custom AI chip, announced this week with Broadcom.
Jalapeño is not a general-purpose AI processor. It is an application-specific chip built only to run inference, the work of generating answers from a trained large language model, rather than to train those models in the first place. OpenAI calls it its first "Intelligence Processor", and its hardware lead Richard Ho has framed the design around minimizing the data movement that eats into inference cost per token. Broadcom designed the silicon; TSMC manufactures it in Taiwan; Celestica handles the boards and rack systems. The point is not bragging rights. The point is closing the spread between what OpenAI charges for a ChatGPT response and what it pays the merchant-silicon vendor that delivers that response.
The arithmetic explains why. OpenAI spent roughly $8.4 billion serving ChatGPT last year while running about 900 million weekly active users, and it projects about $14 billion in inference operating expense this year. Against that, the company has committed roughly $1.4 trillion to compute over the next eight years, while current annual revenue sits near $25 billion. When the largest single line item in your cost structure is silicon rented from one supplier, vertical integration stops being a vanity project. It becomes the only durable answer.
Jalapeño's target, per OpenAI and echoed in VentureBeat and TechTimes coverage, is roughly 50% lower inference cost versus OpenAI's prior serving stack. That is a company-stated design target, not an independently benchmarked result, and the baseline (per token, per request, or blended) has not been publicly detailed. Even so, a 50% cut on a $14 billion annual inference bill is the kind of number that justifies a multi-year chip program on its own. The Register's framing is blunt: this is OpenAI putting Nvidia's pricing power on notice.
A quieter story sits inside the program. According to VentureBeat, OpenAI used its own models to speed up Jalapeño's design and verification, an early case of AI tools designing the silicon that will run AI. That is a smaller claim than the headline margin number, but it points to a future in which the cost of building custom chips falls for the labs that already own the best models. The competitive moat shifts from "we can rent the most GPUs" to "we can design, verify, and iterate our own inference substrate faster than our competitors can."
Two caveats keep the announcement from being a clean Nvidia-killer story. First, Jalapeño is inference-only. Training the next generation of frontier models still rides on Nvidia and AMD capacity, so Nvidia remains strategically relevant to OpenAI even as serving share shifts. Second, the publicly available Jalapeño silicon today is "early lab samples running an unreleased model (GPT-5.3-Codex-Spark)", not shipped production hardware. The 50% number is a design target; the production cost curve will be set by yield, volume, and how Broadcom's networking stacks against Nvidia's NVLink.
What to watch next is straightforward. The first production deployment of Jalapeño in real ChatGPT traffic will turn the 50% target into either a real margin lever or a slide-deck number. Broadcom's next earnings call should disclose whether the OpenAI program moves the needle on its AI-revenue line, or whether it remains one custom-ASIC customer among several. And the rest of the frontier-model industry, from Anthropic to Google to Meta to xAI, now has a working template for the question every CFO in AI is quietly asking: if your biggest supplier takes roughly 75 cents on the dollar, how long do you keep renting instead of building?