Modular Opens Its AI Stack Across Trainium, TPUs and Qualcomm Silicon

Modular spent ModCon 2026 arguing that the AI software stack should stop being a per-vendor concern, and it backed the argument with licence changes rather than roadmap slides. The company removed device usage restrictions from MAX, expanded source access to the platform, and placed the entire Mojo compiler under Apache 2.0.

The news carries new corporate letterhead. Qualcomm completed its acquisition of Modular ahead of the event, and the announcements were issued jointly.

One platform, four new silicon targets

The substantive change for infrastructure teams is hardware coverage. Modular Platform previously served production workloads on NVIDIA and AMD GPUs. It now adds AWS Trainium, Google TPUs, and Qualcomm's Cloud AI 100 Ultra and Dragonfly accelerators.

Custom accelerators are a far harder portability test than a second GPU vendor, because memory hierarchies and execution models diverge much more sharply. Modular's claim is that all four targets run behind identical modelling APIs, serving workflows and language, so a model authored once moves between architectures without a rewrite underneath it.

The company also says each bring-up consumed more than ten times less engineering effort than enabling new AI hardware conventionally demands. One outside result supports that: engineers at HTEC stood up Google TPU support themselves within a few months using a small team, with Modular in a supporting role rather than leading the integration.

Why the compiler had to open

Full Apache 2.0 licensing for Mojo is the piece that makes the rest credible. Modular had already opened its standard library in 2024 and the MAX kernels in 2025, but the compiler stayed closed, which meant nobody outside the company could port the language to hardware Modular had not prioritised.

That constraint is gone. Third parties can extend the language, target new silicon and build commercial products without negotiating terms first. Contributors are being pointed to mojolang.org.

Windows is the first visible beneficiary. Mojo has run on macOS and Linux for years while Windows developers were routed through WSL, and native support is now being built with Microsoft's Windows team. Logan Iyer, corporate vice president for Windows Platform and Developer, said Microsoft wants developers to reach the tools they need on whichever platform they choose.

The commercial half

Modular Cloud reached general availability at console.modular.com, offering OpenAI-compatible shared endpoints billed per token alongside dedicated deployments running on Modular's compute or a customer's own.

It has been operating quietly for months. The service carried OpenRouter traffic under the name ModelRun, where Modular says its endpoints sat at or near the top of the platform for latency and throughput on live traffic.

MiniMax is the reference customer, serving M3 at billions of tokens per minute on a dedicated deployment. That model is unusually demanding to host, combining a one-million-token context window, native multimodality and a sparse attention design that attends selectively to relevant key-value blocks. Modular implemented it natively in MAX and wrote specialised kernels for the attention scheme.

An alliance, and a test of neutrality

Modular says it is assembling an industry alliance spanning hardware vendors, model providers, cloud companies and datacenter operators, with partners taking a direct role in integrating and optimising MAX for their platforms.

The obvious question is whether a Qualcomm subsidiary can credibly maintain a neutral platform. Modular's answer is that it will keep optimising for silicon competing directly with its parent's products. Until this week that was a promise. With the compiler open and MAX source-available, it becomes something the ecosystem can check for itself.