OpenAI and Broadcom Unveil the Jalapeño Chip: The Official Stage, Division of Labor, and What's Not Yet Announced

On June 24, 2026, OpenAI and chip maker Broadcom jointly unveiled a custom chip called Jalapeño. Both companies' official articles call it OpenAI's first “Intelligence Processor,” built for model “inference” (the computation that runs after a user submits a prompt and before the model produces a reply). The two companies' headlines are not quite the same: Broadcom's reads “LLM-Optimized Intelligence Processor,” while OpenAI's own reads “LLM-optimized inference chip.” The word “GPU” does not appear in the body of either article.

This article was checked on September 18, 2026, against the full text of the press release on Broadcom's investor relations site, the full text of OpenAI's own article published the same day, and the full text of Broadcom's earlier collaboration press release from October 13, 2025. This site has not tested any of this itself and does not offer purchase or investment advice. Everything below is a statement from one of these three official articles, and most of it still has no verifiable figures behind it.

The June 24 joint announcement: a chip that isn't called a GPU

What the two companies published is, in fact, largely the same text. Both open by calling it “an accelerator architected around OpenAI's vision for the future of LLM inference,” and both say Jalapeño “is designed with flexibility to work with all LLMs,” guided by OpenAI's insights into the inference needs of current and future AI models across the industry. Neither article says Jalapeño will be sold, rented, or offered to any company besides OpenAI, and neither names any third-party customer.

The official articles describe a handover scene: Jalapeño was delivered to OpenAI CEO Sam Altman and President Greg Brockman by Broadcom President and CEO Hock Tan and another Broadcom executive, Charlie Kawwas. Brockman's words come in two parts: “The world is moving to a compute-powered economy,” and “Jalapeño is part of our long-term full-stack infrastructure strategy to make compute more abundant, resulting in AI which is faster, more reliable, more affordable for people and businesses…”

Who did which part: OpenAI's design, Broadcom's silicon and networking

The division of labor as officially described: OpenAI designed the chip architecture from scratch based on its own understanding of large-language-model computing, and Broadcom and Celestica then helped “industrialize” the platform together. The closing lines of both articles then split the two apart: Broadcom contributed silicon implementation, networking, and connectivity technologies (naming only one product, the Tomahawk networking silicon), while Celestica contributed board, rack, and system expertise. This is the companies' own description of their roles, which this site cannot independently verify.

One thing worth noting: across all three official articles cited in this piece, Celestica's role is described only from Broadcom's and OpenAI's side. None of the three articles quotes Celestica directly, and this article does not cite any statement from Celestica itself as corroboration.

Both official articles also state that the design process used OpenAI's own models to speed up parts of the design and optimization work, saying that “the same models served to users” are helping improve the infrastructure that will run future models. This is the two companies' own account, which outsiders have no way to verify.

What stage things are at now: tape-out done in nine months, still running models in the lab

Both official articles contain a pair of statements that look similar but are not the same: the bullet points say the chip went from design to “production” in just nine months, while the body text says it went “from initial design to manufacturing tape-out in just nine months.” Tape-out is the stage where a chip's design is finalized and handed off for manufacturing; it is not the same as already being in mass production and shipping. Both articles' own subheading also reads “Nine-month tape-out.” Going by the body text and the subheading, what the two companies are saying is that tape-out was completed within nine months, not that mass production was completed within nine months; this figure, too, is the companies' own statement.

The companies say that, for now, engineering samples are running machine-learning workloads, including GPT-5.3-Codex-Spark, in the lab at production-target frequency and power. Both articles stop there — neither says how these samples relate to a shipping product, and neither gives a shipping or launch date. Performance is described only in directional terms so far: OpenAI says it is “still measuring final performance,” and early testing shows performance per watt “substantially better than current state-of-the-art,” but neither article defines what “current state-of-the-art” means or names any vendor; the companies say a detailed technical report will be presented “in the coming months.”

For the same claim about the “fastest ASIC development cycle,” the two companies hedge it differently: Broadcom's own wording is “what may be the fastest ASIC development cycle ever achieved in high-performance advanced semiconductors,” while OpenAI's own article uses “what we believe to be” for the identical sentence. Both versions are hedged, and this article presents them side by side exactly as written rather than merging them.

Deployment timing and an earlier collaboration: two announcements, not to be mixed up

Both official articles describe the deployment scale only as “gigawatt scale,” with no gigawatt figure, chip count, or data-center location given; the official line is that this multi-generation platform is “designed for initial deployment by the end of 2026, and expanding in the years ahead.” Hock Tan names Microsoft as one data-center partner, in his own words: “we are enabling the deployment of gigawatt scale data centers with Microsoft and other partners beginning in 2026.” Microsoft is the only data-center partner named by name in either official article; everyone else is referred to only as “other partners.”

This deployment timeline carries a caveat that is easy to miss: at the end of its press release, Broadcom includes a “forward-looking statements” notice required under U.S. securities law, which explicitly lists “the deployment of gigawatt scale data centers” as one of the forward-looking statements and states that “undue reliance should not be placed on such statements.” In other words, Broadcom itself is acknowledging that this timeline is a plan that has not happened yet and could change.

Also to be kept separate is an earlier announcement from October 13, 2025: the two companies then announced a collaboration to deploy 10 gigawatts of custom AI accelerators, stating that Broadcom's deployment of racks was “targeted to start in the second half of 2026, to complete by end of 2029,” and that the two sides had signed a term sheet to deploy the racks; that announcement also stated that OpenAI had, at that time, “grown to over 800 million weekly active users.” The June 2026 article does not repeat the “10 gigawatts” figure and does not state that Jalapeño is that accelerator — these are two separate announcements, and they should not be treated as the same thing.

How to read this announcement: what is fact and what is the companies' own words

This announcement mixes three different kinds of content: externally observable facts, such as both companies publishing official articles on June 24 and exactly where the wording of the two differs; statements that are visible only inside the companies and cannot be verified from outside, such as tape-out being completed in nine months, engineering samples running in the lab, “performance per watt substantially better than current state-of-the-art,” and “over 800 million weekly active users”; and plans for the future, such as the deployment timeline, which even Broadcom itself warns not to over-rely on.

Both official articles contain one sentence that is often over-read: “If AI can help engineers design better chips faster, it can lower the cost of compute across the industry and help democratize access to advanced AI.” This sentence opens with “if,” is about the industry as a whole, carries no number or percentage, and does not say how much cheaper Jalapeño itself is. As checked through September 18, 2026, neither article contains any cost-savings percentage of any kind.

As checked through September 18, 2026, neither official article states the chip's manufacturing process, which foundry makes it, the memory type and capacity, or its power draw in watts, and neither makes any direct comparison with an NVIDIA chip or any other vendor's chip. OpenAI's article does have one undated, conditional sentence saying that improvements in cost, speed, and reliability “can show up as” a faster ChatGPT reply, an API product that is cheaper to build, and more; beyond that, no sentence states on what date users would notice any difference.

2026 AI News Roundup: Highlights and Daily Life Applications from January to September

OpenAI Discloses Habitat Storage Architecture: The Scale and Limits Behind Over 1 Billion Weekly Users

• AI tools
• Gadgets
• AI news