Qualcomm Buys Modular: The Punch Aimed at CUDA
Qualcomm is buying Modular for about $3.9 billion, all stock. On paper it's a phone-chip company grabbing an AI software startup. Underneath, it's the most direct attack on Nvidia's real moat anyone's made this year.
Here's the thing people get wrong about Nvidia. The moat was never just the silicon, it's CUDA, the software layer that's now nearly impossible to leave because everything is written against it. Modular's whole pitch is killing exactly that lock-in. Its Mojo language and MAX inference engine let you run a model across Nvidia, AMD, Intel, and Qualcomm chips without rewriting your code. Buy that, and suddenly Qualcomm's own data-center accelerators have a software stack that speaks everyone's language.
CEO Cristiano Amon called it a pivotal moment for the industry, which is the kind of thing CEOs always say, but here it might actually be true. If switching chips stops meaning rewriting your entire stack, the thing keeping buyers locked to Nvidia softens. That's the whole game.
And it landed the same day OpenAI revealed its Broadcom chip. So in one afternoon, Nvidia got hit on hardware by OpenAI and on software by Qualcomm. Everyone who's been paying Jensen rent is building a way out at the same time. The interesting question for agent builders is downstream: if inference gets cheaper and chip-agnostic, the cost of running agents at scale stops being a Nvidia decision. More at https://www.modular.com/blog/qualcomm-to-acquire-modular
← Back to all articles
Here's the thing people get wrong about Nvidia. The moat was never just the silicon, it's CUDA, the software layer that's now nearly impossible to leave because everything is written against it. Modular's whole pitch is killing exactly that lock-in. Its Mojo language and MAX inference engine let you run a model across Nvidia, AMD, Intel, and Qualcomm chips without rewriting your code. Buy that, and suddenly Qualcomm's own data-center accelerators have a software stack that speaks everyone's language.
CEO Cristiano Amon called it a pivotal moment for the industry, which is the kind of thing CEOs always say, but here it might actually be true. If switching chips stops meaning rewriting your entire stack, the thing keeping buyers locked to Nvidia softens. That's the whole game.
And it landed the same day OpenAI revealed its Broadcom chip. So in one afternoon, Nvidia got hit on hardware by OpenAI and on software by Qualcomm. Everyone who's been paying Jensen rent is building a way out at the same time. The interesting question for agent builders is downstream: if inference gets cheaper and chip-agnostic, the cost of running agents at scale stops being a Nvidia decision. More at https://www.modular.com/blog/qualcomm-to-acquire-modular
Comments