Modular held its ModCon 2026 developer conference in San Francisco this week, its first developer conference since Qualcomm completed its acquisition of the company earlier this summer.

The event was both a coming-out party for the newly combined organization and an early proof point for Qualcomm’s bet on buying Modular, namely that ceding control of a piece of its AI software stack to the open-source community is the fastest way to challenge NVIDIA’s CUDA ecosystem.

At the event:

• Modular open-sourced the full Mojo 1.0 compiler and toolchain under an Apache 2.0 license
• Rewrote the MAX license to remove device usage restrictions and move toward a source-available model
• Extended hardware support to a broader set of accelerators,
• Modular Cloud to general availability.

The announcements matter because they establish, at least on paper, a hardware-neutral software layer that spans Qualcomm’s Cloud AI100 and Dragonfly accelerators, as well as AMD, NVIDIA, Apple, AWS Trainium, and Google TPU silicon.

AMD‘s on-stage participation is a key detail, suggesting that at least one major accelerator vendor is willing to build on infrastructure now owned by a chip company it competes with directly in PC and data center processors.

For Qualcomm, ModCon is the first real test of whether the acquisition, reported at roughly $3.9 billion (though Qualcomm’s official release did not disclose terms), delivers genuine influence over how AI workloads are written and deployed.

For the broader accelerator market, the announcements add momentum to a multi-year push, alongside efforts such as OpenAI’s Triton, Meta’s PyTorch compiler stack, and the UXL Foundation‘s oneAPI initiative, to loosen NVIDIA hold on the software layer that has historically locked developers into NVIDIA’s hardware.

Announcement Details

At ModCon 2026, Modular’s product and engineering teams outlined a set of platform changes spanning licensing, compiler tooling, hardware integration, and cloud deployment.

Taken together, these changes remove the commercial and technical friction that Modular’s leadership has said kept some developers and hardware partners from adopting the platform before the Qualcomm acquisition.

Here’s what the company announced:

• Mojo 1.0, the full compiler and toolchain, is now released under the Apache 2.0 license. This follows the 2024 open-sourcing of the Mojo standard library and the 2025 release of the MAX kernels, completing a three-year path toward a fully open Mojo stack.
• The MAX license no longer includes device usage restrictions. Modular said MAX will become source available, paired with a new open alliance program intended to enable hardware vendors, model providers, and cloud companies to build on and extend the platform directly.
• Broad Platform Support: Platform hardware support was expanded to include AWS Trainium and Google TPUs, in addition to existing support for NVIDIA GPUs and DGX Spark, AMD Ryzen AI Halo and AMD GPUs, Apple silicon (including Mac Mini), and Qualcomm’s Cloud AI100 and Dragonfly accelerators. Modular said that integrating a new accelerator into the platform now requires more than 10 times less engineering effort than earlier versions,
• Native Windows support for Mojo is under development through a collaboration with Microsoft‘s Windows engineering team.
• Modular Cloud, the company’s hosted inference service, reached GA, offering both OpenAI-compatible shared endpoints and dedicated deployment options. Modular cites MiniMax as a production customer running a deployment that serves billions of tokens per minute and is tuned for MiniMax’s 1M-token context window and sparse-attention architecture.

Analysis

For ML engineers and platform teams, an open-source Mojo compiler and a source-available MAX runtime lower legal and procurement barriers that previously slowed adoption, particularly for organizations unwilling to build production workloads on a platform with device-restricted licensing terms.

Beyond this:

• Removing device restrictions and moving toward source availability simplifies legal review for enterprises and reduces the risk of vendor lock-in driven by licensing terms rather than technical merit.
• Broader accelerator support gives platform teams more flexibility to mix hardware, in theory allowing a single Mojo or MAX codebase to target Qualcomm, AMD, Nvidia, or cloud-specific silicon.
• Native Windows support expands desktop and workstation development to a much larger pool of developers who previously needed Linux or WSL to work with Mojo, potentially broadening the contributor base for the open-source project.

Competitive Landscape

Modular’s stack competes most directly with NVIDIA’ CUDA and TensorRT ecosystem, but the ModCon announcements also place it in the same conversation as other efforts to build hardware-neutral AI software layers.

The comparison below summarizes how Modular’s platform compares with the primary alternatives:

Modular’s strongest differentiation is the breadth of accelerator support announced in one place and the speed with which it can move now that Qualcomm is funding it directly.

Its weakest point is close to its biggest asset, a dependence on a single well-funded parent whose neutrality the rest of the hardware ecosystem has not yet had time to test.

Final Thoughts

The ModCon announcements extend the strategy Qualcomm laid out when it completed the Modular acquisition in July, naming Modular founder Chris Lattner Executive Vice President of Advanced AI Software and Platforms and keeping Mojo, MAX, and Modular Cloud as distinct brands.

Qualcomm CEO Cristiano Amon has described the deal as delivering “high-performance, energy-efficient AI solutions from edge to cloud.” Open-sourcing the software stack is the clearest evidence that this is happening.  

Overall:

• The strategy positions Qualcomm as a software-enabled silicon vendor competing on more than raw accelerator performance, an approach closer to how Nvidia built CUDA into a moat than to Qualcomm’s historical approach to mobile and PC chips.
• Open-sourcing Mojo and loosening the MAX license mean Qualcomm is giving up direct licensing control over intellectual property it just paid for, a bet that ecosystem adoption and influence over how AI software is written are worth more than the platform’s short-term monetization.
• The approach carries some execution risk. Open-source projects succeed as much on contributor momentum and governance credibility as on code quality, and Qualcomm will need to demonstrate that Modular operates with sufficient independence from its new parent to attract contributions from competing hardware vendors over time.

ModCon 2026 gave Qualcomm and Modular a credible, concrete story just three weeks after the acquisition closed, built on real licensing changes and a broader hardware roadmap. Mojo 1.0’s move to Apache 2.0, the MAX license rewrite, and AMD‘s willingness to appear alongside Qualcomm on stage are all meaningful early signals of momentum.

What the event did not resolve is whether developers and rival hardware vendors will build on infrastructure controlled by one of the accelerator makers the event is meant to serve neutrally. Nvidia’s CUDA advantage rests on a decade of library depth and developer habit that an open license does not erase overnight. Modular’s own production track record, outside of showcase customers like MiniMax, remains thin relative to the incumbents it is positioning against.

The clearest takeaway from ModCon 2026, however, is that Qualcomm has moved quickly to turn the Modular acquisition into an open-infrastructure play rather than, as many feared, a walled-garden extension of its silicon business.

That choice, more than any single feature announced onstage, is what the rest of the AI accelerator market will respond to for the next several quarters.