OpenAI Built Its First AI Chip in Nine Months. Broadcom’s CEO Says It Cuts Inference Costs Nearly in Half.
On June 24, 2026, OpenAI and Broadcom unveiled Jalapeño — OpenAI’s first custom-designed AI chip — at a launch event where Broadcom CEO Hock Tan and Broadcom President Charlie Kawwas physically handed an early unit to OpenAI CEO Sam Altman and President Greg Brockman on stage. Jalapeño is not a general-purpose GPU. It is an application-specific integrated circuit — an ASIC — built for one job: serving already-trained models to users, known as inference, rather than training them.
The unveiling capped a nine-month span from design to tape-out that OpenAI and multiple outlets describe as the fastest ASIC development cycle ever achieved in advanced semiconductors. It followed a framework agreement the two companies struck eight months earlier, on October 13, 2025, to co-develop and deploy 10 gigawatts of custom OpenAI-designed accelerators and networking equipment between the second half of 2026 and the end of 2029. That first announcement had no chip name attached, and it sent Broadcom’s stock up roughly 9-10% in a single session. The June unveiling — arguably the more concrete milestone — moved the stock a comparatively modest 2%.
Tan told CNBC’s David Faber the new chip delivers cost savings of roughly 50% compared with typical AI GPUs on inference workloads. That figure is the entire rationale for the project: Nvidia remains OpenAI’s platform for training frontier models and general-purpose compute, while inference — the recurring, compounding cost of serving a live product at OpenAI’s scale, every day, indefinitely — is where a narrower, purpose-built chip can pay for itself.
- 9 months — design-to-tape-out span for Jalapeño — described by OpenAI and outlets including TechCrunch as the fastest ASIC development cycle ever achieved in advanced semiconductors
- ~50% — reported inference cost savings versus typical AI GPUs, per Broadcom CEO Hock Tan's estimate to CNBC's David Faber
- 10 gigawatts — custom OpenAI-designed accelerators and networking equipment committed under the October 13, 2025 framework deal, deploying H2 2026 through end of 2029
- 9-10% vs. 2% — Broadcom's stock move on the October 2025 framework announcement versus the June 2026 chip unveiling — CNBC, Bloomberg
Jalapeño is a joint-engineering product with a clear division of labor. OpenAI designed the chip’s architecture — the layout choices that determine how data moves through the silicon. Broadcom handled the silicon implementation, packaging, and networking, including the Ethernet switch silicon (Broadcom’s Tomahawk line) that links thousands of chips into a single serving cluster. Celestica, the Canadian electronics manufacturer, builds the boards, racks, and system integration that turn the chip into a deployable server.
Richard Ho, OpenAI’s hardware lead, described the design logic: “We optimized the architecture around the kernels, memory movement, networking, and serving patterns that matter most for frontier AI models.” Ho’s presence on the project is itself notable — he is a former architect of Google’s Tensor Processing Unit program, the industry’s most mature custom-AI-silicon effort, and is now building OpenAI’s answer to it.
Jalapeño is reported to be manufactured at TSMC on a 3-nanometer process node, though that specific detail comes from secondary reporting rather than an official OpenAI or Broadcom specification sheet. By the June 24 unveiling, OpenAI said engineering samples were already running production workloads on GPT-5.3-Codex-Spark, its coding-focused model — evidence the design had moved past simulation and into a real serving environment ahead of the planned 2026-2029 rollout.
It is worth keeping the two OpenAI-Broadcom announcements distinct, since coverage sometimes blurs them together. The October 2025 announcement was a framework deal — a commitment to scale, with no chip design revealed. The June 2026 announcement was the product itself: named, running workloads, and physically handed to OpenAI’s two top executives on stage.
OpenAI and Broadcom announce a strategic collaboration to co-develop and deploy 10 gigawatts of custom OpenAI-designed AI accelerators and networking equipment. No chip name yet. Deployment: H2 2026 through the end of 2029.
OpenAI and Broadcom unveil the chip, named Jalapeño, physically delivered to Sam Altman and Greg Brockman by Hock Tan and Charlie Kawwas. Engineering samples reportedly already running GPT-5.3-Codex-Spark workloads.
We've designed and built our first AI chip: Jalapeño. Designed from the ground up by OpenAI and brought to production with @Broadcom, Jalapeño is purpose-built for the LLM workloads powering ChatGPT, Codex, the API, and future agentic products.
The smaller stock reaction to the more concrete announcement is arguably the more informative one: investors had already priced in the collaboration eight months earlier, and Wall Street’s muted response in June suggests the chip’s existence, while a genuine engineering milestone, changed few near-term financial assumptions about Broadcom’s custom-silicon backlog.
The recurring framing across coverage of Jalapeño is “escape route, not replacement.” Nvidia remains the platform OpenAI uses for training frontier models and for general-purpose compute; Jalapeño targets inference specifically — the cost of serving ChatGPT, the API, and agentic products like Codex to users every day, at OpenAI’s scale, indefinitely. That recurring bill, not the one-time cost of training, is the target.
“We like to think we can do better because there is a lot of demand.”
Hock Tan, President & CEO, Broadcom — CNBC, June 24, 2026
Tan framed the tradeoff in straightforwardly economic terms, telling CNBC’s David Faber the chip’s roughly 50% cost advantage over typical AI GPUs comes from stripping out general-purpose flexibility that inference workloads simply don’t need. Altman and Brockman made a version of the same point in less numeric terms: “By designing more of the stack ourselves, we can serve more intelligence with greater efficiency and keep pushing advanced AI toward broader access.”
Introducing Jalapeño — designed from scratch for LLM inference over nine months, accelerated by our models. Perf per watt looking incredible.
OpenAI is not the first frontier lab to build its own chip — it is the last of the major ones to join an already-established trend. Google’s Tensor Processing Units have been in production since 2015 and now power both Gemini and external Google Cloud customers. Amazon has shipped two custom AI chip families, Trainium for training and Inferentia for inference, inside AWS. Microsoft has its own accelerator, Maia, for Azure workloads. Meta has MTIA, its in-house inference chip — also built in partnership with Broadcom, making Broadcom the common silicon partner across three of the industry’s frontier labs: Google, Meta, and now OpenAI.
None of this threatens Nvidia’s core moat in the near term. Nvidia’s advantage isn’t only hardware — it’s CUDA, the software ecosystem developers have built around for more than a decade, which gives Nvidia’s GPUs a flexibility that a purpose-built ASIC doesn’t try to match. Jalapeño and its counterparts at Google, Amazon, Microsoft, and Meta compete on cost-per-token for narrow, predictable inference workloads — not on Nvidia’s general-purpose flexibility.
Notably absent from both announcements: a dollar figure. Neither OpenAI nor Broadcom has disclosed the value of the chip deal itself — not the October framework agreement, not the June unveiling. That is a contrast with the headline valuations attached to some of OpenAI’s other 2026 announcements, and it leaves outside observers to gauge the deal’s scale from the 10-gigawatt capacity figure and Broadcom’s own capital-expenditure disclosures, rather than from a stated contract value.
Confirmed: The October 13, 2025 framework deal (10GW of custom accelerators, deployment H2 2026–2029) and the June 24, 2026 Jalapeño unveiling — both announced jointly by OpenAI and Broadcom, with the chip physically delivered to Altman and Brockman on stage. The nine-month design-to-tape-out span and the GPT-5.3-Codex-Spark engineering-sample workload are OpenAI’s own claims.
Reported, not officially specified: The TSMC 3-nanometer manufacturing node, sourced from secondary reporting rather than an OpenAI or Broadcom spec sheet.
Not disclosed: A dollar figure for the deal. Hock Tan’s roughly 50% cost-savings estimate is his own on-air comparison to CNBC, not an independently audited benchmark against a named competing chip.
Jalapeño is not an attempt to replace Nvidia — it is OpenAI’s attempt to control the recurring cost of running its own products at scale. Built in nine months, the fastest ASIC cycle either company has publicly claimed, and reportedly cutting inference costs roughly in half, the chip joins Google, Amazon, Microsoft, and Meta’s own custom-silicon programs in a race that targets cost-per-token, not Nvidia’s software moat. No dollar figure has been disclosed for the deal. The rollout runs through 2029.
