OpenAI's first homegrown inference chip is less a clean break from Nvidia than a warning shot: the biggest AI customers now want leverage over the hardware bill.
OpenAI has started testing Jalapeño, its first in-house AI chip, and you should read that as a very practical move. According to Axios, the chip was developed with help from Broadcom, is aimed at inference rather than training, and is already being used in OpenAI labs for work similar to answering Codex queries. Customer use is planned for later this year.
That is the story. Not that Nvidia suddenly loses its grip on AI computing. It does not. The company still sits at the center of model training and high-end accelerator supply. The point is narrower and more interesting: OpenAI is no longer willing to treat Nvidia hardware as the only serious option for serving its products at scale.
Inference is where the daily bill lives. Every ChatGPT answer, every Codex response, every low-latency customer query has to run somewhere, and that workload is different from training a frontier model from scratch. Training rewards flexible, general-purpose systems that can handle enormous experimentation. Serving billions of requests rewards lower power use, predictable latency, and hardware shaped around the model you actually run.
Jalapeño is built for that second job. Axios reported that this first generation is meant for inference, not training, and OpenAI is still considering whether its homegrown chip work should eventually move into training as well. That distinction matters because it keeps the announcement in the real world. OpenAI is not replacing Nvidia tomorrow. It is attacking the part of the cost structure where a custom chip can make the most immediate sense.
Broadcom's role is just as important as OpenAI's name on the chip. The company has become the favored custom silicon partner for hyperscalers that know their workloads well enough to stop buying only off-the-shelf accelerators. Google has its TPUs. Amazon has Trainium and Inferentia. Meta has MTIA. Microsoft has Maia. Now OpenAI has Jalapeño, with Broadcom helping turn the design into something that can actually be manufactured and deployed.
Look at the timing. In October 2025, OpenAI and Broadcom announced a multi-year agreement to co-develop and deploy 10 gigawatts of custom AI accelerators and rack systems, with deployments expected to begin in the second half of 2026 and continue through 2029, as Tom's Hardware reported at the time. That deal was already a serious infrastructure commitment. Wednesday's Jalapeño update gives it a more concrete face.
Broadcom also gets something cleaner than a one-off win. It gets to sit underneath the custom silicon plans of companies that are spending at a scale most chip buyers will never reach. Business Insider reported after the October agreement that Broadcom shares jumped as much as 9% on the OpenAI deal, and Barron's recently described Broadcom's ASIC business as one of the strongest AI positions outside Nvidia. You do not need to love stock-market reaction pieces to see the underlying point. Broadcom is becoming the workshop for AI companies that want chips built around their own math.
Nvidia is still powerful, but less alone
Frankly, the risk for Nvidia is not that Jalapeño beats the H100, B200, or whatever comes next in a neat benchmark chart. That is the wrong comparison. The risk is that the biggest customers stop accepting one default architecture for every job. Once Google, Amazon, Microsoft, Meta, and OpenAI all run serious custom silicon programs, Nvidia's pricing power has to answer a harder question: why should every inference dollar flow through a general-purpose GPU stack?
Nvidia still has the best full platform in the market for many buyers. It has chips, networking, software, developer familiarity, and a supply chain that enterprises trust. A mid-sized company is not going to design an ASIC because its chatbot bill looks ugly. You need immense volume, stable workloads, and engineering teams that can live with a long hardware cycle. OpenAI has those things. Most customers do not.
That is why Jalapeño matters most as a signal. It tells you where the largest AI operators think the economics are heading. They will keep buying Nvidia where Nvidia is the best answer. They will build or co-build their own chips where the workload is repetitive enough and expensive enough to justify it.
OpenAI has already been diversifying. Axios noted that the company recently began using Cerebras chips for inference, while Nvidia remains central to its training and broader compute stack. That is the pattern to watch. Not a dramatic divorce. A supplier map with more names on it.
If Jalapeño works in customer-facing deployment later this year, OpenAI gets more than a chip. It gets bargaining power, a tighter handle on inference costs, and a path to hardware tuned for the products it actually sells. Nvidia will still be hard to displace. But the days when the largest AI companies simply waited in line for the next GPU allocation are ending.
Also read: Vinod Khosla wanted every dollar of Runlayer's $30 million round and that tells you everything about where enterprise AI is heading • Seltz raises $12.5 million to build the search layer that agentic AI actually needs • Meta and Microsoft just pre-bought the AI future, and the landlords are the ones with real leverage
