OpenAI unveiled a custom chip called Jalapeño on June 24, built with Broadcom. It is the company's first piece of hardware designed entirely around its own AI workloads. The announcement sounds technical, but the real story is financial and strategic.
To understand why this matters, you need to know the difference between training and inference. Training is when a company builds an AI model by feeding it vast amounts of data, usually over weeks or months. Inference is what happens every time someone uses that model: a question goes in, an answer comes out. Training happens once per model. Inference happens billions of times a day, indefinitely, and it is by far the larger ongoing cost.
OpenAI posted a $20.9 billion operating loss in 2025 on $13 billion in revenue. Inference costs alone reached $8.4 billion last year, and they are on track to hit $14 billion in 2026 as ChatGPT now serves roughly 900 million weekly users. On top of that, OpenAI paid Microsoft $17.2 billion in 2025, mostly for the cloud hardware that keeps ChatGPT running. That is a structural dependency. Jalapeño is a direct attempt to reduce it.
The chip's design logic is straightforward. Nvidia's hardware is powerful but built to handle many different kinds of work. A chip designed only to answer ChatGPT questions, built around the exact patterns of how those requests arrive and are processed, can do that specific job far more cheaply. Google has operated its own purpose-built chips, called TPUs, since 2015. Amazon's comparable custom chips have delivered cost savings of 50 to 90 percent on inference workloads compared to Nvidia hardware, according to multiple reports from organizations that made the switch. OpenAI is entering this game later, but with a precise target in mind.
The chip went from blank design to production-ready in nine months. OpenAI and Broadcom describe this as potentially the fastest development cycle ever recorded for hardware of this complexity. OpenAI used its own AI models to speed up parts of the chip design process, an interesting detail: the same models being served to users helped build the hardware that will run future versions of them. Broadcom handled the silicon manufacturing and networking integration. A third company, Celestica, builds the physical board and rack systems. TSMC manufactures the chip in Taiwan.
Early lab tests show better performance per unit of energy than current alternatives, though OpenAI has not published the full numbers and has not named what it was compared against. A detailed technical report is expected in the coming months. The chip is currently running an unreleased model called GPT-5.3-Codex-Spark in lab conditions. Broadcom's CEO confirmed the deployment plan: small prototype rollout in late 2026, ramping through 2027, and going to full scale in the first half of 2028.
This is part of a planned multi-generation partnership, not a one-off product. OpenAI and Broadcom have both described Jalapeño as the first step in a longer roadmap of custom chips designed around future model requirements. Microsoft is expected to be a major purchaser of the output.
For anyone buying OpenAI products or using its API, there is an indirect but real implication. AI labs have been cutting prices on API access throughout 2025 and into 2026, partly because their ability to serve responses more cheaply creates room to pass savings along or defend market share. A lab that cuts its per-response cost structurally, through hardware it controls rather than rents, has more room to hold or lower prices under competitive pressure. OpenAI is facing exactly that pressure: Anthropic, Google, and others are in a visible price war on API access, and OpenAI's leaked financials show it is pricing models below the true cost of serving them.
OpenAI has filed a confidential application to go public, with revenue now running at roughly $25 billion annually on an annualized basis as of early 2026. The company does not project turning cash-flow positive until 2030. The gap between revenue and costs is the central question any public investor will ask. Jalapeño is one part of the answer OpenAI needs to give.