News · 2026-06-25
Qualcomm buys the software that lets AI run anywhere
Qualcomm, best known for the chips inside your phone, has agreed to buy a software company called Modular for about $3.9 billion in stock, a deal expected to close in the second half of this year. The announcement comes from Modular's own blog. On paper Modular is small. In practice Qualcomm is buying three things that are worth a lot more than they look: a programming language called Mojo, a piece of software called a compiler, and the engineer who built it.
Start with the problem Modular exists to solve. When you run an AI model, the model has to be translated into instructions a particular chip understands. That translation layer is the compiler. For years, one chipmaker has dominated AI not just because its hardware is good, but because its software, the layer that turns models into chip instructions, is so entrenched that almost everything is written for it. Switching to a different chip means rewriting your software, and that switching cost is the real moat. People talk about the hardware lead; the lock-in is mostly in the software.
Modular's pitch was to break that lock by building a translation layer that doesn't care which chip is underneath. Write your AI program once, and it runs efficiently on whatever hardware you have, this chipmaker's, that one's, a phone, a data center. Mojo is the language they built for it, designed to feel as friendly as Python while running as fast as the low-level code underneath. The goal is a world where the chip is a swappable part rather than a lifetime commitment.
That is exactly why Qualcomm wants it. Qualcomm's ambitions now stretch from the tiny AI processors in handsets all the way to chips for data centers. To compete there, it needs developers to be able to run their models on Qualcomm hardware without rewriting everything, which is precisely what a chip-agnostic software layer provides. Buying Modular is Qualcomm attacking the software moat from the side, rather than trying to out-spend the leader on raw hardware.
And then there is the person. Modular was co-founded by Chris Lattner, who is something close to a household name among engineers. He created LLVM and Clang, the compiler technology underneath a huge fraction of modern software, the Swift language that powers iPhone apps, and MLIR, a framework now central to AI compilers. Acquiring his team is, in effect, acquiring one of the deepest benches of compiler talent in the industry. In a field where the bottleneck is increasingly software, not silicon, that is the real prize.
Why it matters: this is the second "own the whole stack" move in a single week, arriving alongside OpenAI's own custom inference chip. The message is that the AI value chain, from silicon to compiler to runtime, is being carved up and bought by the giants. Whoever controls the portable software layer controls a slice of everyone's compute bill, which is why a piece of developer tooling is worth nearly four billion dollars.
The honest caveat sits right at the heart of the deal. Modular's entire promise was independence, software that doesn't play favorites among chips. Now it will be owned by a chipmaker. The developer community will watch closely to see whether Mojo and the compiler stay genuinely neutral, or whether, over time, they quietly run best on Qualcomm's own hardware. "Open and silicon-agnostic, owned by a silicon vendor" is a tension that doesn't resolve itself on the day the press release goes out. It will be judged over the next few years by whether the software still treats a rival's chip as a first-class citizen. For more on why portable, open AI matters, see our explainer on open-weight models.