For years, the AI industry ran on Nvidia. The Santa Clara chip giant’s H100 and B200 GPUs became the de facto currency of the intelligence age — so essential that access to them was rationed like a scarce resource, with companies like OpenAI spending billions to secure allocations. That era may now be ending, or at least fracturing.
On June 24, 2026, OpenAI and Broadcom jointly unveiled Jalapeño — a custom AI inference chip developed over just nine months with substantial help from OpenAI’s own AI models. The announcement is more than a product reveal. It’s a declaration that the world’s most prominent AI company intends to own its destiny at every layer of the stack, from models to software to the silicon underneath.
What Is Jalapeño, Exactly?
Jalapeño is an inference chip, meaning it’s designed to run AI models already trained — not to train new ones from scratch. When you send a message to ChatGPT or ask OpenAI’s coding agent Codex to fix a bug, the compute that powers the response is called inference. It’s fast, repetitive, and extremely expensive at scale.
Training — the process of building a frontier model from scratch — remains GPU-intensive work that will likely continue to rely on Nvidia’s hardware for the foreseeable future. But inference is where the operational costs pile up, day after day, request after request. OpenAI processes hundreds of millions of interactions daily. Even modest efficiency gains at the inference layer translate to enormous savings.
That’s where Jalapeño fits. OpenAI says early testing shows the chip delivers significantly better performance per watt than current state-of-the-art alternatives. The company was deliberate about keeping the exact benchmarks vague, but the emphasis on power efficiency is telling — energy costs are one of the fastest-growing line items for AI infrastructure operations.
AI Designed With AI
One of the more striking details in the announcement is how the chip was built. OpenAI’s models played a direct role in the chip’s development process. The company says Jalapeño went from concept to finished design in roughly nine months — an unusually compressed timeline for semiconductor development, where tape-out cycles typically span years.
This isn’t mere marketing. It reflects a broader industry trend of applying machine learning tools to hardware design itself. Google has used AI-assisted chip design for some of its tensor processing unit (TPU) layouts. Nvidia has experimented with AI in its verification pipelines. OpenAI appears to have accelerated this approach significantly, using its own frontier models to compress a design cycle that has historically been one of the most labor-intensive endeavors in engineering.
OpenAI president Greg Brockman previewed the company’s approach to chip development in the months after the Broadcom partnership was announced. “We have a deep understanding of the workload,” Brockman explained. “We’ve really been looking for specific workloads that are underserved — how can we build something that will be able to accelerate what’s possible?” That philosophy — designing silicon around a deeply known use case rather than building general-purpose hardware — is exactly the playbook Google and Amazon have followed with their custom AI accelerators.
Broadcom as the Manufacturing Partner
Broadcom brings critical capabilities to the partnership. While the company is perhaps best known for networking chips and its recent acquisition of VMware, it has a substantial custom silicon business — helping hyperscalers design and produce application-specific integrated circuits (ASICs) tailored to their exact needs. Google’s TPU line is manufactured with Broadcom’s assistance, making the Jalapeño partnership a logical extension of an established model.
This structure lets OpenAI focus on what it knows best — understanding the AI workload — while Broadcom handles the manufacturing complexity. It also avoids the extraordinary capital expense of building semiconductor fabrication capabilities in-house, which even the largest tech companies have generally declined to do (the notable exception being Intel, whose foundry ambitions have faced serious headwinds).
Jalapeño chips are expected to enter initial deployment by the end of 2026, with the partnership expanding capacity in subsequent years.
The Nvidia Question
The obvious subtext in any custom AI chip announcement is Nvidia. The GPU giant has benefited enormously from the AI boom — its market capitalization has at times exceeded $3 trillion, driven almost entirely by demand for AI training and inference hardware. Nvidia’s CUDA software ecosystem, built over two decades, creates significant switching costs for AI developers who have optimized their code around it.
But the big AI players have strong incentives to reduce that dependence. Nvidia charges premium prices, supply remains constrained, and every dollar spent on Nvidia hardware is a dollar not recaptured by the company doing the spending. Google and Amazon both built custom chips to address this — Google’s TPUs now power much of its AI workload, and Amazon’s Trainium chips are increasingly available to AWS customers building their own AI applications.
OpenAI’s Jalapeño play follows the same logic. The company was explicit: “OpenAI is not only developing frontier models or building products on top of them; it is designing the infrastructure underneath them: chip architecture, kernels, memory systems, networking, scheduling, deployment systems, and product experience. Because OpenAI operates across the stack, each layer can be optimized around the same goal: making its models faster, more reliable, and more affordable for users.”
That’s not subtle. OpenAI is describing a vertically integrated AI operation — one where the model, the software, and the silicon are all tuned to work together. It’s the same logic Apple applied to its M-series chips: when you control the hardware and the software, you can extract efficiencies that general-purpose solutions can’t match.
A Broader Industry Shift
The Jalapeño announcement comes alongside another major silicon story this week: Qualcomm announced it is acquiring AI software platform developer Modular for under $4 billion, as part of a broader push into AI infrastructure. Qualcomm — long known as the dominant chip supplier for smartphones — is simultaneously unveiling data center products and signing a multi-generation agreement with Meta. The message from both companies, in different ways, is the same: the AI chip market is no longer Nvidia’s alone to define.
The implications for the industry are significant. As more AI workloads shift from pure training runs to inference at scale, the calculus for which hardware wins changes. Nvidia’s GPUs are extraordinarily flexible and powerful, but flexibility has a cost — both in price and in power draw. Purpose-built inference chips like Jalapeño can sacrifice that flexibility in exchange for efficiency, because the workload is already well-understood.
If Jalapeño delivers on its performance claims at scale, it will accelerate a trend already underway: the fragmentation of AI hardware from a Nvidia-dominated monoculture toward a more competitive landscape where the hyperscalers and large AI labs run on silicon they helped design. Nvidia will likely remain essential for frontier model training for years to come. But inference — the layer that touches every user query, every AI agent interaction, every generated response — may increasingly belong to chips like this one.
OpenAI has spent years as a software and model company. With Jalapeño, it’s becoming something more. Whether that’s a chip company is almost beside the point. It’s a company that understands, more deeply than almost anyone, what its models actually need — and it’s building the hardware to give it to them.
