SAN FRANCISCO, CALIFORNIA – JUNE 02: Open AI CEO Sam Altman speaks during Snowflake Summit 2025 at Moscone Center on June 02, 2025 in San Francisco, California. Snowflake Summit 2025 runs through June 5th. (Photo by Justin Sullivan/Getty Images)
Getty Images
OpenAI just changed how it plans to power itself. This week it unveiled Jalapeño, its first proprietary chip, an application-specific processor built with Broadcom and aimed at a single job: inference, the work of running a trained model rather than teaching it. The move is a deliberate step away from leaning on general-purpose hardware and toward a vertically integrated stack OpenAI controls from the model down to the silicon.
That is the strategy. The investable part of it sits one layer beneath OpenAI. A custom chip is not something a software company builds on its own. It needs a partner who holds the design blueprints, the packaging, and the manufacturing relationships, and OpenAI found that partner in Broadcom. The company that stamps its name on a custom chip and the company that captures the durable profit from it are rarely the same.
Meet Jalapeño
Jalapeño is an ASIC, a chip designed for one narrow purpose rather than the general-purpose flexibility of a graphics processor. Its target is the memory bottleneck that slows large language models down. The physical design pairs a large compute section with six stacks of high-bandwidth memory to move data through faster, and by optimizing the data paths directly on the silicon, the chip cuts the power needed to run the complex, multi-step workloads that AI agents create.
Broadcom chief executive Hock Tan says Jalapeño runs inference for roughly half the cost of a typical AI GPU. Hold that figure at arm’s length. It is Broadcom’s own number, measured on workloads OpenAI chose, with no independent benchmark behind it, so treat the exact percentage as marketing. The direction is the real point. A chip designed for a single task can usually beat a general one at that task, and inference is the bill that never stops arriving.
The reason it targets inference specifically comes down to where the money goes. Training a model is a one-time, enormous expense. Inference is the recurring one, every answer, every agent step, billions of times a day, and it is the cost that grows with success. A chip tuned only for inference can strip out the machinery a general processor has to carry and do that one job for less. For a company serving models at OpenAI’s volume, shaving the cost of inference is the difference between economics that work and economics that do not.
There is a second motive beneath the cost math, and it is the one OpenAI is really after: control. Renting general-purpose hardware means building a business on a component someone else designs, prices, and decides how to allocate. Owning the design changes that equation. The silicon gets shaped around how OpenAI’s own models actually behave, and the company controls a slice of its own supply instead of bidding for all of it on the open market. That is what a vertically integrated stack is meant to buy, from the model at the top down to the processor at the bottom.
Not A One-Time Fix
Owning a chip is a road, not a switch. Building proprietary silicon takes years of engineering iteration before a company reaches anything close to self-sufficiency, and the clearest proof is the company furthest along. Google began designing its own AI chips roughly a decade ago and only this year reached its seventh generation, Ironwood, its first TPU built specifically for inference. Custom silicon is a continuous program, not a single release that erases a hardware shortage overnight. Jalapeño is generation one.
So the merchant suppliers do not disappear. First production runs rarely cover a company’s full demand, which means OpenAI keeps buying inference chips from outside vendors while Jalapeño ramps. Proprietary silicon supplements the hardware OpenAI already runs rather than replacing it, which keeps demand alive across the broader market of merchant chips even as the custom program grows. That detail cuts against the simplest version of the story. A custom chip reads like OpenAI walking away from its suppliers, when the near-term reality is a company that needs both its own silicon and theirs to keep pace with demand.
Why Broadcom Is The Toll Booth
Here is the part worth holding onto. The AI software race is loud and genuinely hard to call. Which frontier model wins, which lab pulls ahead, which assistant the market finally settles on, none of it is easy to predict. Owning the company that designs the custom silicon underneath all of them sidesteps the guessing game entirely.
These are also not relationships a customer walks away from casually. Co-designing a chip means years of shared engineering, intellectual property, and hardware roadmaps that tie the lab and the designer together long after the first part ships. The longer a partnership runs, the harder it is to unwind, which is what turns a single design win into recurring revenue.
Google, Meta, and OpenAI all build their custom chips on Broadcom, which turns a noisy contest between AI models into steady, predictable revenue for the company sitting beneath every one of them. That is what a toll booth looks like. It does not matter which car is fastest, because they all pay to use the road. The order book already shows it. Broadcom reported $8.4 billion in AI chip revenue in the first quarter of fiscal 2026, up 106% from a year earlier, and it sits on a $73 billion backlog of committed orders, close to half the company’s total, that management ties to a path toward $100 billion in annual AI chip revenue by 2027.
The Factory Sets The Pace
The constraint on all of this is not design. It is manufacturing. Every one of these chips, Jalapeño included, depends on Taiwan Semiconductor for advanced fabrication and the specialized packaging that bonds compute and memory into a single working part. That packaging capacity is sold out through 2026, and demand across the industry runs far ahead of supply. OpenAI does not get to skip the line. It competes for finite allocation alongside every major technology company on earth, which is why a finished design does not become chips in racks on any timeline OpenAI controls by itself.
That bottleneck is also the quiet reason the thesis holds. When the scarce resource is the physical capacity to build and package advanced silicon, the companies that hold the designs and the manufacturing relationships are sitting exactly where the scarcity is.
The Picks-And-Shovels Thesis
The headlines will score this as OpenAI versus its chip suppliers, a software company going independent. The more durable read runs the other way. Secular shifts like this one create volatility in the names everyone watches, while the infrastructure underneath stays necessary and largely unglamorous.
The companies that gain the most from a wave of custom AI chips are not the labs racing to put their names on silicon. They are the ones that own what every chip needs: the design blueprints, the packaging relationships, the manufacturing access. Broadcom is the clearest example, the firm whose architecture sits inside the custom programs of three of the largest AI companies at once. The names on the chips will keep changing. The company that designs them stays the same.