Qualcomm’s $4 Billion Modular Deal Targets the Software That Runs AI
Qualcomm Inc. on Wednesday agreed to acquire Modular, an artificial-intelligence software startup, for roughly $4 billion in stock, one of the largest deals in the history of the San Diego chipmaker and its most direct assault yet on Nvidia Corp.’s grip over AI software.
The deal values Modular, co-founded by Chris Lattner, the creator of Apple Inc.’s Swift programming language and the LLVM compiler project, at nearly four times the valuation it commanded nine months ago. Modular raised $250 million at a $1.6 billion valuation late last year; under the acquisition, Qualcomm will issue up to 19.2 million shares, including about $300 million earmarked for Modular employees.
Lattner is one of the most recognizable engineers in Silicon Valley. He created LLVM, the compiler infrastructure that underpins much of modern software development, built Swift at Apple, and briefly ran Tesla Inc.’s Autopilot software before founding Modular in 2023 with Tim Davis, a veteran of Google’s TensorFlow Lite team. The company’s roughly 150 employees, nearly all of them compiler and systems engineers, will join Qualcomm.
Modular builds software that lets developers write an AI model once and run it on almost any chip—Nvidia GPUs, AMD accelerators, Google’s TPUs, Amazon’s Trainium, Qualcomm’s own silicon. The company’s core product, the MAX platform, replaces the entire stack of libraries and tools that developers normally need for each hardware vendor, a pitch that directly challenges CUDA, the two-decade-old software ecosystem that locks most AI developers to Nvidia.
The strategic logic is straightforward. Nvidia’s dominance rests as much on CUDA as on its hardware; developers have spent years optimizing for it, and switching costs are enormous. Qualcomm, which makes chips for phones, cars, PCs and increasingly data centers, has struggled to persuade developers to write for its accelerators. Modular gives it a software layer that makes its chips compatible with the broader AI ecosystem without asking developers to learn a new toolchain.
“When we run on an Nvidia chip, we replace all of CUDA,” Lattner told a trade publication this year, describing how Modular’s runtime supplants Nvidia’s math libraries and algorithms. That ambition is why Qualcomm is paying a price that one analyst called “a panic buy wrapped in a vision statement.” Supporters counter that Qualcomm is buying the scarcest asset in AI: a team that can make software portable across hardware, at a moment when every chipmaker except Nvidia needs exactly that capability.
The deal is the boldest move yet by Qualcomm Chief Executive Cristiano Amon, who has spent two years positioning the company as an AI player beyond mobile phones. Qualcomm has said it is working on dozens of chip designs for AI devices ranging from smart glasses to earbuds, and it is developing custom data-center accelerators, with ByteDance among the early customers reported for that effort. Modular’s cross-platform software is meant to tie that hardware together: write once, run on every Qualcomm chip from a watch to a server rack.
Qualcomm’s breadth is the point. The company’s Snapdragon line powers most Android phones, its chips run in millions of cars through the Snapdragon Digital Chassis, and its PC processors have gained ground against Intel. Each of those markets is being reshaped by AI, and each needs software that lets developers deploy models without hand-tuning for every chip. Modular is the bridge Qualcomm could not build itself.
The CUDA moat is real but not unbreakable, analysts said. Matt Kimball of Moor Insights & Strategy has described Nvidia’s advantage as a “softer moat”—customers are not required to use Nvidia everywhere, but Nvidia-on-Nvidia remains the easiest route to maximum performance. Modular attacks that ease-of-use advantage directly, and if its platform works on Qualcomm silicon, developers gain a reason to consider alternatives for the first time in years.
The inference market is where the fight will happen. Training models remains dominated by Nvidia, but running those models in production—inference—is growing faster and is more fragmented across hardware. Qualcomm believes inference workloads will spread across phones, cars, edge devices and modest-sized servers, markets where its silicon is already present. Modular is the software that lets those chips compete.
The transaction is expected to close in the second half of the year, subject to regulatory review. It joins a series of multibillion-dollar bets by chipmakers on AI software, and it raises the stakes for Nvidia, which has responded to portability pushes by deepening CUDA’s integration with its own hardware. Qualcomm’s bet is that the AI industry will not accept a single-vendor lock-in forever—and that the company that owns the neutral software layer will capture the value when it breaks.
