Qualcomm on Wednesday said that it will spend $3.9 billion to purchase AI-native software platform developer Modular Inc., a move that Qualcomm says will allow it to level the playing field on data centers by creating “a silicon-agnostic compute layer.”

The stock-based acquisition “further enables Qualcomm Technologies to deliver a silicon-agnostic compute layer across devices, edge, and data centers, improving performance-per-watt, increasing hardware flexibility, and expanding an open developer ecosystem so customers can deploy AI more efficiently across heterogeneous platforms globally,” the company said in a statement. 

Qualcomm’s position is that enterprises need far more flexibility in their data center strategies, especially given how fluid the AI space is today. When CIOs need to make bets on data centers without knowing what the field will look like in two years, it can be challenging.

Chris Lattner, CEO of Modular, posted on LinkedIn that this leveling of the data center playing field was one of the company’s key early goals.

“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. This gap holds back innovation and choice and makes development painful,” Lattner wrote in his LinkedIn post.

“Modular was founded 4.5 years ago to solve this problem,” he wrote. “We’ve already integrated support for several hyperscale datacenter silicon providers, but we’re not stopping with what’s publicly announced. We’ve built an open platform and are continuing to open it further.”

Lattner added that the Qualcomm acquisition “will accelerate our progress and path” by “spanning edge to cloud, CPU, GPU, NPU, and custom ASICs and perhaps more.”

Addresses a pain point

Analysts, although skeptical of the probability of success in taking meaningful market share away from Nvidia, said that Qualcomm has focused on a true sore point for enterprises struggling with data center approaches. 

Matt Kimball, VP and principal analyst with Moor Insights & Strategy, said, “the argument that Modular can make datacenters cost-effective is directionally correct. As enterprise AI actually hits velocity, heterogeneity is almost an understatement. Different accelerators are required for different use cases across different deployment scenarios.”

To date, it’s been a challenge for organizations to manage AI in this environment, he noted. “And when enterprise AI takes off, this challenge will be fully exposed.”

Kimball said that Modular “can be extremely valuable in achieving two things that will vex most organizations: abstracting complexity and delivering significantly more flexibility. And this would certainly lead to TCO advantages. I think the per-watt performance claim can be challenging to validate across every and any deployment scenario, but I understand the spirit behind it.”

Yuri Goryunov, CIO of consulting firm Acceligence, also applauded the Qualcomm move, but he stressed that the deal’s value is not in the technology as much as in the talent.

“The key is what Qualcomm actually bought: not silicon, but the software layer, meaning Chris Lattner’s team plus Mojo and the MAX engine. That’s the right place to apply pressure. Nvidia’s real moat has never been the GPUs,” he said. “It’s CUDA and the rewrite cost that keeps workloads pinned to their hardware. A credible ‘write once, run across CPU/GPU/NPU/ASIC without rewrites’ layer is exactly what lowers the switching cost and makes non-Nvidia silicon a safer bet.”

But Goryunov said that the data center “democratization” argument also is powerful.

“Anything that pushes toward democratization of compute and better routing of tasks to best-fit capacity adds real flexibility to the ecosystem,” he noted. “If workloads can be matched to the right compute instead of defaulting to one vendor, everyone gets more efficiency on performance-per-watt and TCO and customers get real choice. That’s the part of this I find most compelling.”

Still some obstacles

That said, none of this will be easy, he pointed out.

“Does it change the competitive position versus Nvidia? Directionally, yes. It opens a credible second front at the exact point where Nvidia is stickiest. I’d stop short of saying it shifts the balance overnight. CUDA’s moat is a decade deep and this is a multi-year execution play,” Goryunov said. “But the attack is aimed at the right wall and the team they bought is about as serious as it gets for this fight.”

But he stressed that much of Qualcomm’s strategy with this acquisition relies on an uncertain assumption: That Nvidia won’t counterattack by opening its architectures to various others. Or, at the very least, that Nvidia won’t do so quickly enough.

“That’s the barrier to entry, which is that Nvidia will focus on their stickiness,” Goryunov said.

Kimball added that, from a competitive perspective, Qualcomm has various obstacles to overcome. “Part of this acquisition goes directly to the Nvidia challenge” of finding a way to “make it easier for customers to deploy heterogeneous silicon without software getting in the way.”

John Annand, senior technical counselor at Info-Tech Research Group, is more skeptical of Qualcomm’s ability to do serious damage to Nvidia.

“Nvidia has something like 85% of the AI accelerator chip market,” he pointed out. “Sure, they have nowhere to go but down, but that’s still going to take them a while. More importantly, they have literally spent decades working with practitioners in AI and ML and compute-intensive fields, indoctrinating them into their CUDA software ecosystem. Rewriting that tool chain will take institutional change at most organizations, which means years, if not decades, to uncouple.”

“Organizations that think they’ve achieved agnosticism because they’re using high-level abstractions like PyTorch, well,  they have come closest,” he observed. “But just cutting and pasting the same code into AMD Instinct can lead to memory and dependency errors. It’s like VM lift and shifts to the public cloud 10 years ago. Easier, but still possible to screw up.”

Nonetheless, Annand said that the deal, if it goes through, is still good news for enterprises. 

“What it means for enterprise IT is that the vendors we currently rely on to deliver AI have another potential building block. Because enterprise IT accesses AI via an API call, it’s operationally irrelevant to us if Claude runs on Nvidia, AMD ROCm, or Modular,” he said.

“Now, because of the commercial and stock agreements, OpenAI and Anthropic aren’t going to jump ship anytime soon. But if your enterprise is looking for more boutique offerings, like those from Cohere, or is looking to build its own models and tools from scratch, this is an exciting announcement.”

Goal: build once, run anywhere

Shashi Bellamkonda, principal research director at Info-Tech Research Group, looks at the potential acquisition, while it will potentially deliver benefits, as suffering from many practical roadblocks.  

“Qualcomm is chasing what you might call model democracy,” he said, noting that today, AI deployment teams are locked to whatever accelerator they trained on, and moving a model to different hardware means re-engineering, not just configuration changes.

“Modular’s pitch is that this goes away: build once, run across CPU, GPU, NPU, whatever the infrastructure calls for,” Bellamkonda said. “That’s a credible goal. The catch is that democracy and portability aren’t the same thing. Qualcomm will tune hardest for Qualcomm silicon. Every hardware company does. Vendor-neutral software foundations have a habit of developing hardware preferences once their acquirers need to differentiate silicon.”

Flavio Villanustre, CISO for the LexisNexis Risk Solutions Group, provided a different perspective. 

“I think it’s important to clarify that Modular is behind the Mojo programming language, which provides an abstraction layer for AI models, enabling them to run across different hardware architectures,” he said. “In the traditional approach, if you code an AI stack on Python or C and target a particular hardware architecture, such as X86, Nvidia GPU, AMD GPU, or TPU, you will need to rewrite a significant portion of that to run it on a different architecture. With Mojo, you code it once and it runs everywhere, even on hybrid systems composed of different hardware architectures.”

And, he said, “if you now consider the fact that Qualcomm owns intellectual property and manufacturing across different hardware architectures, both CPU and GPU, this acquisition could offer their customers significant lift. I see this as Qualcomm buying abstraction that allows them to provide diverse hardware offerings and still offer their customers full code reuse across their entire CPU/GPU/TPU/NPU portfolio.”