OpenAI Partners with Broadcom to Launch First Custom AI Inference Chip Jalapeño
Alina Collins
OpenAI and Broadcom jointly unveiled Jalapeño, OpenAI's first custom AI inference chip, with early tests showing costs roughly 50% below typical AI GPUs — a move that shifts OpenAI from renting compute to building its own, directly reshaping its cost structure.
What does this chip do, and how much does it save?
Jalapeño is built specifically for AI inference — the stage where a trained model actually answers user queries, not the training itself.
Early tests show ~50% cost savings versus typical AI GPUs, with performance comparable to Nvidia's Blackwell and Google's TPU.
This means → if those numbers hold at scale, OpenAI's per-conversation cost for running ChatGPT could roughly halve — the single biggest variable in its path to profitability.
How was it built, and who is involved?
The chip design took about nine months, partly accelerated by AI-assisted design tools, then fabricated by TSMC.
Hardware lead Richard Ho said Jalapeño is "materially better than the current state of the art" on performance per watt.
Canadian electronics manufacturer Celestica will build the server systems. The finished hardware is exclusive to OpenAI — not for sale.
In plain terms = OpenAI supplies the design intent, Broadcom the chip engineering, TSMC the fabrication line — three parties assembling a dedicated supply chain.
When will it be deployed, and what comes next?
OpenAI plans to deploy Jalapeño in large-scale data centers by year-end, working with Microsoft and other partners.
Broadcom CEO Hock Tan said the two companies expect to exceed the previously projected 1.3 gigawatts of chip deployment by next year — "demand is very strong."
The next generation is planned for 2028, with annual iterations after that. Jalapeño focuses on inference; future generations may expand to other workloads.
Where is the money coming from?
OpenAI earlier this year closed $122 billion in funding earmarked for chips, data centers, and talent.
Broadcom has set up chip-financing vehicles with Apollo Global Management and Blackstone to support OpenAI's orders.
This means → the specific financing for full chip orders is not yet locked in, but the funding infrastructure is already in place — building the tool before filling the order signals both sides expect large-scale production.
What does this mean for the industry?
Hock Tan predicted that other frontier-model developers "will eventually take the custom AI accelerator path as well."
This reflects a broader shift: leading AI companies are moving from buying off-the-shelf Nvidia GPUs to designing their own chips and outsourcing fabrication — the same road Google's TPU and Amazon's Trainium have taken.
In plain terms = whether Jalapeño can deliver its lab-stage numbers at full deployment is the critical proof point for OpenAI's custom-hardware strategy.
Content is for reference only, not financial advice.