Qualcomm has open-sourced Mojo as part of a cunning plan to loosen Nvidia’s grip on AI coding.
According to Jon Peddie Research, Qualcomm closed its acquisition of Modular on 29 July 2026, after announcing the deal in June at roughly $3.1 billion to $3.9 billion in stock. Its first move as owner was to push the Mojo compiler to GitHub under the Apache 2.0 license, with LLVM exceptions.
Qualcomm executive vice president for advanced AI software, Chris Lattner, built LLVM, Clang, MLIR, and Swift before Mojo became his latest systems-language punt. It looks like Python on the surface, but underneath it compiles a single codebase to GPUs, NPUs, and whatever accelerator turns up next.
The pitch is to write the kernel once, rather than writing it in CUDA and then rewriting it for everything else. Every hour a developer spends trapped inside CUDA is time Qualcomm’s silicon is not in the running.
Modular engineering leader Mostafa Hagog, a former Nvidia engineering director, ran teams behind cuDNN and CUTLASS. The code sits in the modular/modular repository. One Bazel command builds the compiler and standard library from scratch, while a prebuilt flag pulls the latest nightly binary for those who prefer not to wait.
Apache 2.0 with LLVM exceptions includes a patent grant and lets Mojo binaries ship inside commercial software without needing legal permission.
Modular is not taking compiler pull requests yet, although the standard library has accepted outside contributions since 2024 and compiler patches are expected by year’s end. ModCon brought the rest of the pitch. Mojo compiles across Nvidia, AMD, the Fruity Cargo Cult Apple’s Silicon GPUs, AWS Trainium, Google TPUs and Qualcomm’s own accelerators from a single codebase.
MAX, Modular’s inference framework, moved to broader source access and dropped device-usage restrictions. Qualcomm talks about an “open, heterogeneous ecosystem”, although Modular had already been opening parts of the stack before Qualcomm turned up with its chequebook.
While it looks great on paper, the cunning plan still has to compete with Nvidia’s 15 years of libraries, tools and developer habits. Compiler portability gives developers a path away from CUDA, but the next 18 months of library work will decide if it will be the Cuda killer after all.
