HWNEWS
Modular, the startup behind the Mojo language and Max platform, will stay independent inside Qualcomm in a deal that tests whether an open AI software stack can stay neutral inside a chip company.
Qualcomm sells proprietary chips. Modular, the startup it just bought for close to $4 billion, builds AI software designed to run on any chip, including ones Qualcomm does not sell. The deal rests on a premise the rest of the AI industry is still arguing about: the next platform fight will be fought in the software layer between the trained model and the underlying silicon.
That layer has a name. It is the runtime: the compiler, the kernel library, and the deployment system that turns a model into a running application on real hardware. Today that layer is dominated by NVIDIA's CUDA, which is why almost every AI model performs best on NVIDIA GPUs. Modular is an attempt to build a runtime that is genuinely hardware-agnostic, so the same model can run on a Qualcomm Hexagon NPU, an AMD GPU, or a data-center accelerator without a rewrite.
Modular's two named products are Mojo, a programming language designed for AI kernels that need to run portably across chip types, and Max, a graph and deployment platform that schedules model execution across heterogeneous hardware. The company was founded by engineers from Apple and Google, including cofounders Tim Davis, its president, and others. Davis, in an interview with EE Times, frames the acquisition as the validation of four ideas.
The first is that AI compute will be heterogeneous. Not every workload will run on the same chip. Training large frontier models will keep living in data centers, but inference, the act of running an already-trained model, will spread across phones, PCs, cars, and dedicated accelerators. The second is that inference, not training, will dominate the volume of compute the industry consumes. Most AI compute already is inference, and that share is going up as models ship into devices. The third is that any serious path to "superintelligence" requires both the data center and the edge. A model that lives in the cloud but cannot run on the device is incomplete. The fourth is that software is the bottleneck. The hardware is already there. The cross-silicon runtime that lets developers target it portably is what is missing.
Davis names three reasons Qualcomm was the buyer. The first is footprint: Qualcomm's NPU, CPU, and GPU silicon already ships in phones, PCs, automotive platforms, and IoT devices, so the company has a route into all of those categories. The second is distribution: tens of millions of devices a year, which gives a runtime an immediate install base. The third is CEO Cristiano Amon, who has personally pitched the AI software strategy and was willing to spend close to $4 billion on it. NVIDIA owns the data center. Qualcomm is one of the few companies with reach into the device, and the device is where inference is going.
The single most consequential term of the deal, by Davis's account, is that Modular remains an independent entity inside Qualcomm. The team stays. The open-source mission stays. The roadmap, on paper, stays neutral across all silicon, including NVIDIA. That neutrality is the only reason a developer outside Qualcomm would ever adopt Modular's stack. If the runtime drifts toward favoring Hexagon NPUs in benchmarks, in scheduling defaults, or in optimization paths, the developer base has no reason to stay. The independence guarantee is what makes the deal legible. It is also the clause most likely to be tested in practice.
The structural question behind the deal is who controls the software layer AI runs on. NVIDIA controls it today through CUDA lock-in. Modular wants to make that layer neutral the way Linux is neutral across CPU vendors. Whether a chip company can own the steward of cross-silicon neutrality without steering it is the open question, and it is the only question that matters for whether the $4 billion pays off.
The first concrete test will be Modular's release cadence and commit history over the next year, and whether the project's major contributions keep coming from engineers outside Qualcomm. Davis says the deal closes in the third quarter.