Qualcomm Paid $3.9 Billion for Mojo, Then Gave the Compiler Away Twenty Days After Closing. I Called This Wrong in June.
On June 25 I wrote about Qualcomm agreeing to buy Modular for about $3.9 billion, and the argument was that hardware neutrality is a strategy rather than a property of a codebase, and that strategies change hands along with the equity. The implied prediction was that a chip company would eventually tilt the stack toward its own silicon.
The deal closed on July 29. On August 18, at ModCon, Qualcomm put the entire Mojo language under Apache 2.0. Compiler included. Tooling included. Twenty days after taking ownership.
That is not what I expected, and it is worth working out why it happened, because the reasoning generalizes past this one deal.
What was actually released
The specifics matter here more than the headline, because "open source" gets applied loosely to three different things in the same announcement.
Mojo 1.0 is Apache 2.0, unrestricted, compiler and all tooling. That is real open source and it is irrevocable for the code released under it. Modular had been opening this stack progressively for two years already: the standard library in 2024, the MAX kernels in 2025. The compiler was the last piece.
MAX, the inference platform, is a different story. The announcement says the MAX license no longer contains device usage restrictions and that MAX will be "source-available with an open alliance program." Source-available is not open source. You can read it, and depending on the terms you may be able to modify it, but the owner keeps control over what you can do with the result. That distinction is doing a lot of work in a post that uses the word "open" in its title three times.
Modular Cloud, meanwhile, is a commercial product that went generally available the same day, with MiniMax named as a flagship customer running M3 on a dedicated deployment.
So the shape is: give away the language, hold the platform, sell the service.
Why a chip company gives away a compiler
The stated reason in Modular's own post is that the industry needs a common software foundation so developers can write a model once and reach every accelerator. They commit explicitly to supporting hardware that competes with Qualcomm's own.
I think that commitment is credible, and here is the mechanism, which has nothing to do with goodwill. Qualcomm's problem is not that it lacks a compiler. Qualcomm's problem is CUDA. NVIDIA's moat is that every model, kernel, and tutorial assumes NVIDIA hardware, and no amount of silicon performance fixes that. The only thing that erodes it is a portable layer that enough people actually use.
A portable layer that people suspect is tilted toward one vendor gets no adoption, and a portable layer with no adoption is worth nothing to the vendor who owns it. Neutrality is not a favor Qualcomm is doing the ecosystem. It is the only configuration in which the asset works at all. Apache 2.0 on the compiler is Qualcomm buying credibility with the one currency that cannot be faked, which is giving up the ability to take it back.
The supporting evidence is in the hardware list. Modular Platform now targets AWS Trainium, Google TPUs, and Qualcomm's Cloud AI 100 Ultra and Dragonfly, alongside NVIDIA and AMD. Two of those are direct competitors' accelerators built by companies with no reason to help Qualcomm. And HTEC, an outside firm, brought up Google TPU support themselves in a few months with a few engineers and Modular in a supporting role. That is the test of whether a platform is genuinely extensible, and it passed.
Where I was wrong, and where I was not
I was wrong on the prediction. I expected tilt and got the opposite, faster than anyone expected.
I was wrong for a specific reason worth naming: I reasoned from incentive to behavior in one step. Qualcomm sells chips, therefore Qualcomm favors its chips. That skipped the question of what the acquired asset is actually for. When you buy a credibility asset, the incentive runs toward protecting the credibility, and the strongest way to do that is to give up control in a way that is legally hard to reverse.
The part I would still defend: the analysis separated what a company promises from what it has structurally committed to. That framework is what makes the August news readable. Apache 2.0 on the compiler is a structural commitment. The MAX licensing language and the alliance program are promises with governance attached, and governance changes.
What this means if you build on it
For almost everyone reading this, nothing immediate. Mojo is a systems language for AI kernels and high-performance numerics. If you write application code, this is not a dependency you have or want.
The transferable part is a checklist for any acquired dependency, which is a situation you will hit repeatedly. Separate the layers. For each one, ask what license it carries, whether the license is revocable for code already released, and who controls the roadmap. Then check which layer your project actually sits on. A permissive license on the compiler does not protect you if your product depends on the platform above it, and that is exactly where MAX sits.
Native Windows support is also coming to Mojo via a collaboration with the Microsoft Windows team, which is a real signal: Microsoft does not staff work like that for a language it expects to be a Qualcomm silo.
The honest take
If you were avoiding Mojo because of the acquisition, that specific reason is gone. The compiler is Apache 2.0 and that cannot be walked back for the code already published.
If you were avoiding it because it is a young language with a small ecosystem, that reason is untouched. Mojo hit 1.0 with a stability guarantee earlier this month. A stability guarantee is a promise about the future made by a project that has not had one yet. The libraries, the hiring pool, and the answers to your specific problem still mostly do not exist.
Open sourcing a compiler removes a risk. It does not manufacture an ecosystem. Those are different problems and only one of them got solved on August 18.
Author
Lukas
@lukcombinator