Qualcomm Is Spending $3.9 Billion on Modular. The Target Is CUDA, Not the GPU.

Qualcomm said on Wednesday it will buy Modular, a software company with no chips of its own, for about $3.9 billion in stock. That looks strange for a chipmaker until you notice what Modular actually sells, and what it is pointed at. The target is CUDA, and behind CUDA, Nvidia.

What Qualcomm actually bought

Modular is the company Chris Lattner started four and a half years ago with Tim Davis. Lattner is not a random founder. He built LLVM, the compiler infrastructure that sits under most modern programming languages, and he created Swift at Apple. Modular's two products are Mojo, a programming language aimed at AI work, and MAX, an inference engine (the software that takes a trained model and serves its answers to users). The pitch is easy to say and hard to deliver: write your AI code once, run it across CPUs, GPUs, NPUs (the neural accelerators now baked into phones and laptops) and custom ASICs, without rewriting it for each one.

Around 150 Modular employees, Lattner and Davis included, move to Qualcomm. The deal is all stock. The Wall Street Journal reported Qualcomm expects to issue up to 19.2 million shares. Qualcomm says it expects to close in the second half of 2026, subject to regulatory approval.

In the company's own words, the acquisition "further enables Qualcomm Technologies to deliver a silicon-agnostic compute layer across devices, edge and data centers." Translated: software that does not care whose chip it runs on.

Why a software layer is worth $3.9 billion

Nvidia holds roughly 85% of the AI accelerator market, and not because its GPUs are the only fast ones. The lock-in is CUDA, the software stack developers have written against for close to two decades. Once your models, your kernels (the low-level routines that do the actual math) and your tooling are tuned for CUDA, moving to AMD or anyone else means redoing a lot of that work. That switching cost is the moat. The hardware is replaceable. The habit is not.

Modular attacks the habit. If a portable layer genuinely lets a model run on non-Nvidia silicon without a painful rewrite, then Qualcomm's own chips, and everyone else's, become a safer bet for a buyer. Qualcomm owns CPUs, GPUs and NPUs, so a write-once layer that spans its whole portfolio is worth more to it than it would be to most acquirers.

"In a world with a tremendous amount of innovative heterogenous AI hardware, there has always been a gap: existing fragmented software technologies weren't built to scale effectively across this hardware."

Lattner wrote that on LinkedIn after the deal, adding that Modular was founded to close the gap and that the company has "already integrated support for several hyperscale datacenter silicon providers."

The part to be skeptical about

"Build once, run anywhere" is one of the oldest promises in computing, and the road behind it is littered with attempts that never quite delivered (OpenCL among them). The analysts who cover this market were blunt about the odds.

John Annand at Info-Tech Research Group pointed out that Nvidia has spent decades "indoctrinating them into their CUDA software ecosystem," and that unwinding that toolchain "will take institutional change at most organizations, which means years, if not decades, to uncouple." Even teams that believe they are hardware-neutral because they use a high-level framework like PyTorch are not all the way there. He noted that copying the same code onto AMD's Instinct GPUs "can lead to memory and dependency errors."

There is also a conflict of interest sitting inside the deal. Shashi Bellamkonda, also at Info-Tech, put it plainly: "Qualcomm will tune hardest for Qualcomm silicon. Every hardware company does." A neutral compute layer owned by a chipmaker has an obvious reason to drift toward that chipmaker's hardware over time. Lattner says the platform will stay open and get more open. Whether that holds once Qualcomm needs Modular to make Qualcomm chips look good is the open question.

And the efficiency claim is not yet checked. Qualcomm sells the layer partly on better performance-per-watt. Matt Kimball at Moor Insights & Strategy called the cost argument "directionally correct," then added that the per-watt claim "can be challenging to validate across every and any deployment scenario." Hold that number loosely until someone outside Qualcomm measures it.

The bigger move

Modular did not arrive on its own. Qualcomm used the same Investor Day to lay out a data center roadmap for products it mostly does not sell yet: the Dragonfly line, including a C1000 server CPU (a 250-plus-core chiplet design Qualcomm puts in production around 2028) and the AI200, AI250 and AI300 inference accelerators on a yearly cadence. It announced a multi-generation agreement to supply Meta with data center CPUs. And it roughly doubled its fiscal 2029 target for non-handset revenue to $40 billion, with more than $15 billion of that pinned on the data center.

That is the frame for the Modular buy. Qualcomm is trying to turn itself into a data center company, and you cannot sell inference silicon into a CUDA world without a software answer. Buying Lattner's team is that answer, or the start of one. It is aimed at the right wall. The wall is still a decade thick.