高通押注开源AI软件可打破英伟达的锁定

Qualcomm wants to sell more AI accelerators. Yet one of its biggest software bets is designed to make developers care less about which accelerator they’re using.

That apparent contradiction sits at the center of Qualcomm’s acquisition of Modular, the AI software company co-founded by compiler pioneer Chris Lattner. Modular is building a software stack intended to let the same high-level model code run across Nvidia and AMD GPUs, Google TPUs, AWS Trainium, Qualcomm accelerators, and Apple silicon.

For Qualcomm, a relative newcomer to data center AI, reducing software dependence on any one chip vendor could be an advantage. It would give customers a way to bring Qualcomm hardware into environments built largely around Nvidia without requiring them to rebuild their software stack.

Rashid Attar, Qualcomm senior VP and head of data center engineering, described the challenge facing alternative accelerator vendors as “two mountains to climb”: first, they must deliver differentiated hardware; then, they must make that hardware easy enough to use that customers do not have to assign engineers to port existing workloads.

By Unikey Electronics Pte. Ltd. 08.26.2026

By Zackary Glazewski, Founding AI Engineer, ChipAgents 08.26.2026

By Colin Duggan, CEO and co-founder, BG Networks 08.26.2026

“When we talk to these customers, they say, excellent, I love what you have. I’m going to give you exactly zero engineers from my side,” Attar told EE Times.

Modular is Qualcomm’s attempt to remove the second mountain.

An architectural diagram of Modular’s software stack featuring Modular Cloud, MAX, and Mojo, illustrating cross-platform support across CPUs, GPUs, and custom ASICs from major hardware vendors.

Software as the barrier to entry

Nvidia’s advantage is no longer simply CUDA, said Matt Kimball, VP and principal analyst for data center at Moor Insights & Strategy. CUDA initially helped Nvidia separate itself from competitors, but the company has extended that advantage across networking, systems, libraries, and software.

“Nvidia has been so smart,” Kimball told EE Times. “It leveraged CUDA effectively to drive a moat across the entire value stack.”

That makes replacing CUDA alone insufficient. Nvidia has NVLink for chip-to-chip connectivity, BlueField data-processing units, a broad software portfolio, and tightly integrated rack-scale systems. Kimball described this as a “softer moat”: customers may not be absolutely required to use Nvidia everywhere, but Nvidia-on-Nvidia can remain the easiest route to maximum performance.

Modular is taking a much broader approach than translating CUDA code.

Matt Kimball

“When we run on an Nvidia chip, we replace all of CUDA,” Lattner told EE Times, before qualifying that Modular retains a small portion for device management. “We replace all the math libraries, we replace all the algorithms, all the modeling — we replace that entire stack.”

The goal is to make AI software behave more like conventional software. New CPU architecture requires low-level compiler and operating-system work, but once that foundation exists, much of the software above it can run unchanged. AI accelerators, by contrast, have evolved with different programming environments, libraries, and toolchains.

Modular’s answer starts with Mojo, its recently released 1.0 programming language, and MAX, its modeling and inference framework. Hardware-specific work remains at the bottom of the stack, Lattner said, but higher-level software and many optimizations can then be shared.

“You have to do that initial bring-up,” he said. “Once you do that, you get this massive leverage across everything else that then composes on top of that.”

Can portability preserve performance?

The difficult part is preserving what makes different accelerators different.

An abstraction layer that hides hardware complexity risks hiding the hardware’s advantages, as well. Kimball expects some trade-off at the extreme. Customers for whom every microsecond matter, he said, are likely to continue optimizing directly for individual architectures.

For the much larger population of enterprise users, however, he thinks the trade-off may be worthwhile. Those customers are more interested in getting workloads deployed quickly and choosing among available hardware than in extracting the final increment of performance from one chip.

Lattner goes further. He argues that Modular can preserve hardware-specific optimization underneath the common programming model and, in some cases, improve performance because engineering work can focus on the relatively small part of the stack that actually changed.

He pointed to Modular’s work on AMD’s MI355X accelerator. In a case study published by the company, Modular said two engineers brought MAX up on MI355 in 14 days. The data showed 99.9% of its stack was architecture-agnostic, leaving engineers to concentrate on changes such as new BF16 conversion instructions, larger tensor-core tiles, and increased shared memory.

Modular also claimed MAX ultimately outperformed AMD’s optimized vLLM fork by as much as 2.2× across the workloads it tested. Those are Modular’s own results, but the example illustrates the company’s argument: portability does not mean eliminating hardware-specific optimization. It means limiting how much of the software stack has to be rewritten to exploit new hardware.

Attar said that distinction is critical for Qualcomm. Its forthcoming accelerator generations must bring enough hardware differentiation to win workloads, while Modular is intended to make those differences accessible without forcing customers to adopt another vendor-specific programming environment.

Qualcomm’s neutrality problem

Another question remains unresolved. Modular’s usefulness depends partly on developers and chip vendors believing that it can extract competitive performance from hardware supplied by multiple vendors. Modular is now owned by one of those vendors.

Lattner acknowledged the problem directly.

“That’s a hard question,” he said when asked how Modular could remain neutral after joining Qualcomm. “There’s not really a perfect answer to that.”

His answer rests on several mechanisms. Qualcomm has made public commitments to keep Modular multi-vendor. Mojo’s compiler and toolchain have been released under the Apache 2.0 license, creating what Lattner called an “irrevocable grant of IP to the world.” Modular also plans to launch an alliance later this year that will give hardware companies and other ecosystem participants a role in the platform.

But, Lattner added, “you have to walk the walk.”

Attar said Qualcomm will maintain organizational firewalls. Modular engineers may need confidential information about forthcoming AMD hardware to optimize for it, for example, but Qualcomm’s accelerator engineers would not receive that information. The same separation would apply in the other direction.

Kimball believes neutrality clearly serves Qualcomm today. The company is trying to break into an Nvidia-dominated market, so it benefits from a software layer that lets its accelerators coexist with Nvidia, AMD, Trainium, and TPU deployments.

The harder test could come later. If Qualcomm becomes a much larger data-center AI supplier, the incentives to remain equally supportive of competing silicon might look different.

Opening the door for smaller accelerators

The bigger impact may extend beyond Qualcomm.

Kimball argues that AMD has enough resources to build a competitive software stack of its own. Smaller accelerator companies do not. A common software foundation could therefore matter most to companies with differentiated silicon but neither the engineering scale nor developer ecosystem required to challenge Nvidia.

Lattner makes a similar argument. Modular initially concentrated on large vendors because, as a startup, it lacked the resources to support every new accelerator. With Qualcomm’s backing and the planned alliance, he wants the platform to expand further.

“A lot of the startups have the most to win,” Lattner said. “They’re innovating really quickly. They don’t have existing software platforms.”

If Modular can extend that common software foundation to smaller accelerator vendors, Qualcomm’s acquisition could do more than strengthen its own data-center ambitions. It could lower one of the largest barriers to entry in AI silicon.

For Attar, the measure of success is straightforward: new Qualcomm hardware should deliver close to its theoretical performance from day one instead of requiring months of software optimization.

For Lattner, success is broader. He expects AI infrastructure to become more heterogeneous, with workloads moving among different accelerators from the edge to the cloud.

If Modular can make that diversity usable, the competitive question in AI silicon could shift. Instead of asking which chip has the software ecosystem developers cannot leave, customers could increasingly ask which chip delivers the best performance, power efficiency, availability, or economics for the job.

For Qualcomm—and almost every Nvidia challenger—that is precisely the competition they would rather have.

Read also:

IBM Makes Quantum Cryogenics Modular, but Scaling Problems Remain

AI Hardware’s Next Frontier Is Integration

GlobalFoundries’ Growth Makes the Case for a U.S. Photonics Buildout

AI AND BIG DATA, BIG DATA

AMD, MODULAR, NVIDIA, QUALCOMM