For years, OpenAI has been Nvidia's most famous customer and most obvious vulnerability. Now it's trying to become something else entirely: a chipmaker.
The company unveiled its first custom AI chip this week, designed in partnership with Broadcom and manufactured by TSMC. The move transforms OpenAI from a pure software play into something closer to a vertically integrated technology conglomerate — think Apple, but for artificial intelligence. It's a declaration that the company believes controlling the entire stack, from model weights to transistors, is the only sustainable path forward.
The Nvidia problem
OpenAI's dependency on Nvidia has always been an awkward arrangement. The GPU giant supplies the vast majority of chips powering AI training runs, which means OpenAI's costs, timelines, and ultimately its competitive position have been subject to Nvidia's production schedules and pricing power. When H100s were scarce in 2024, OpenAI scrambled alongside everyone else. When Nvidia raised prices, OpenAI paid.
Custom silicon changes the calculus. A chip designed specifically for transformer inference — the computationally intensive work of running a model like GPT-5 — can be dramatically more efficient than a general-purpose GPU. Google proved this years ago with its TPUs. Amazon followed with Trainium and Inferentia. The question was never whether OpenAI would pursue custom chips, but when it could afford the billion-dollar R&D commitment.
Why Broadcom, why now
Broadcom is an unusual partner for a company of OpenAI's ambition. The semiconductor giant specialises in designing chips for other companies — it built Google's TPUs and has similar arrangements with Meta and ByteDance. Choosing Broadcom over building an in-house design team signals that OpenAI values speed over total control. The company reportedly wants chips in production by late 2027, an aggressive timeline that would be impossible starting from scratch.
The timing also reflects OpenAI's evolving business model. As the company pushes deeper into enterprise software and consumer subscriptions, inference costs matter more than training costs. Every ChatGPT query, every API call, every agentic workflow burns silicon. At scale, even marginal efficiency gains translate into hundreds of millions in savings — or, alternatively, the ability to offer capabilities competitors cannot match at the same price point.
Our take
This is less a pivot than an inevitability finally arriving. OpenAI has raised over $20 billion and commands a valuation that demands it become a durable technology platform, not merely a research lab licensing compute from others. Custom chips are the toll road to that future. The risk is distraction: semiconductor development is unforgiving, and OpenAI's core advantage remains its model research, not its hardware engineering. But the alternative — remaining permanently dependent on Nvidia's roadmap — was never really sustainable for a company that wants to define the next era of computing. The chip announcement isn't a guarantee of success. It's an admission of what success now requires.
