Digital Economy

OpenAI launches Jalapeño: why does the future of artificial intelligence lie in chips?

The aim is to reduce costs, energy consumption and dependence on Nvidia. Like Google, Microsoft and Amazon, OpenAI is also aiming to take control of its hardware

by Luca Tremolada

FILE PHOTO: OpenAI logo is seen in this illustration taken June 11, 2026. REUTERS/Dado Ruvic/Illustration/File Photo REUTERS

3' min read

Translated by AI
Versione italiana

3' min read

Translated by AI
Versione italiana

OpenAI has officially joined Silicon Valley’s most exclusive club: that of companies which no longer confine themselves to writing software, but also design the hardware on which that software runs.

The new chip is called Jalapeño, and its mission is to reduce the cost of artificial intelligence and increase the available computing power. The announcement came from OpenAI in partnership with Broadcom, the semiconductor giant which, over the last two years, has become one of the key players in the AI economy. In essence, Jalapeño is OpenAI’s first proprietary chip, built specifically for a very precise function: inference.

Loading...

What is inference?

Translated from technical jargon: it’s not for training models, but for making them work. It’s the difference between building a brain and using it. Training is the phase in which a model consumes vast amounts of data and computational power in order to learn. Inference, on the other hand, is what happens every time a user types a question into ChatGPT, asks Codex to generate code, or activates an AI agent.

The problem: AI is too expensive

That is why inference has become fundamental in the new economy of intelligent agents. Over the last two years, the bottleneck in artificial intelligence has been hardware – or, more precisely, the costs involved in providing sufficient computing power to run neural networks. Servers are expensive, electricity is expensive, and the process needs to be managed. OpenAI, like almost everyone else, has grown by relying on Microsoft’s infrastructure and Nvidia chips. But there is a structural problem: the more ChatGPT grows, the higher the computational costs become. This is why OpenAI has chosen to move towards a ‘full-stack’ model: controlling not only the models and products, but also the underlying hardware. Jalapeño was born out of this strategy.

Why inference matters more than training

Training is extremely expensive, but it takes place periodically. Inference, on the other hand, is continuous. It is industrial-scale. It is a daily occurrence. It is the factory that runs 24 hours a day. With millions — soon to be billions — of users querying AI models every day, inference has become the real cost centre. And also the real centre of power. OpenAI claims that Jalapeño offers superior performance per watt compared to chips currently on the market. If confirmed, this means only one thing: more requests processed using less energy. In a business where every millisecond and every watt counts, this is a huge advantage.

Why are all the Big Tech companies making chips?

OpenAI is no exception. It just arrived later. The Big Tech companies have realised one fundamental thing: in the age of AI, it is not enough to own the software. You also need to own the hardware. Google was the first to realise this with its TPUs (Tensor Processing Units). Today, they are the invisible engine behind Gemini and much of Alphabet’s AI infrastructure. Microsoft is pushing ahead with its Maia chips, designed to reduce its reliance on Nvidia in Azure data centres. Amazon has developed Trainium and Inferentia, with a clear division: training on one side, inference on the other. Meta has been working for some time on custom accelerators for Llama. And even Nvidia, the market leader, is shifting increasingly towards architectures designed for inferential workloads, because it has realised where demand is heading.

OpenAI no longer wants to be just an AI laboratory or a software company. It wants to become a comprehensive technological infrastructure. A bit like Apple, which realised years ago that designing its own chips meant building a lasting competitive advantage. The difference is that we’re not talking about smartphones here, but about the new global cognitive infrastructure.

Copyright reserved ©
  • Luca Tremolada

    Luca TremoladaGiornalista

    Luogo: Milano via Monte Rosa 91

    Lingue parlate: Inglese, Francese

    Argomenti: Tecnologia, scienza, finanza, startup, dati

    Premi: Premio Gabriele Lanfredini sull’informazione; Premio giornalistico State Street, categoria "Innovation"; DStars 2019, categoria journalism

Loading...

Brand connect

Loading...

Newsletter

Notizie e approfondimenti sugli avvenimenti politici, economici e finanziari.

Iscriviti