Qualcomm today revealed it has acquired Modular, a provider of a platform that enables application developers to run artificial intelligence (AI) models on top of any class of processors.
Additionally, Qualcomm announced that a Dragonfly C1000 family of CPUs will be made available in 2028 to extend its reach into the data center, with Meta signing up to be one of the first customers.
Speaking at a Qualcomm Investor Day event, Qualcomm CEO Cristiano Amon said the acquisition of Modular is part of a larger five-year plan to expand the scope of its reach into the data center using Dragonfly AI accelerators, CPUs and a high-bandwidth compute (HBC) memory architecture that partners such as Microsoft have committed to deploy in Azure cloud services in 2027.
While rivals such as Intel, AMD and NVIDIA are much more established in the data center, the pace of innovation creates an opportunity for Qualcomm to invest in processors that, for example, process tokens much more effectively to reduce costs, said Amon. “It’s never too late for Qualcomm,” he added.
The Modular framework, meanwhile, adds a missing framework for building software in a way that makes it possible for application developers to build an AI application once and then deploy it anywhere, regardless of whether the platform is based on CPUs, graphics processing units (GPUs), neural processor units (NPUs) or custom ASICs. It provides an alternative to the NVIDIA software stack that doesn’t lock organizations into a specific class of processors.
Previously, Qualcomm has acquired Arduino, Edge Impulse and Foundries.io as part of an effort to build an ecosystem of application developers. In general, Qualcomm is now working toward becoming a “developer-first” company, said Amon.
It’s not clear to what degree application developers are prepared to embrace another software development framework, but pressure to reduce costs in the age of agentic AI increases with each passing day. Many organizations are already discovering that deploying AI agents at scale on existing compute platforms is simply cost prohibitive. The issue is that most of the AI applications that have been deployed thus far are dependent on the CUDA framework that NVIDIA provides.
Qualcomm, however, is betting on a full stack approach to AI that includes robotics and other physical AI applications, many of which will be based on the processors it has historically provided for any number of edge computing applications, including automotive vehicles and any number of smartphones.
Regardless of the platform or software framework employed, it’s clear there is now a lot more focus on costs as organizations move beyond the experimentation phase of agentic AI. Every agentic AI input and output requires access to processor and memory horsepower that, at this point, is becoming increasingly scarce. As such, a Qualcomm effort to expand the total number of processor options for running agentic AI applications will, at the very least, force its rivals to focus more on building platforms that ultimately serve to make AI much more affordable for all.
