Mojo's compiler went Apache 2.0 fifty-five days after Qualcomm's $3.92bn deal

The licence question is settled, and it was the least interesting one on the table. Apache 2.0 with LLVM exceptions is open source whether or not outsiders can push to main, which is the correction that landed on the top Hacker News comment in that thread ^[11], and SQLite has held the same line for years without losing the label ^[12]. So Modular has granted exit. You can clone the compiler and tooling and build the language yourself from one command ^[5]. What it has not granted is voice: contributions to the compiler and tooling stay closed, with a stated aim of opening them by the end of the year, and a stated reason involving caution about AI-generated code ^[6].

That closure is a choice about timing rather than capacity. The standard library has taken outside patches since 2024 and Modular counts more than 1,100 of them ^[7], so the machinery for reviewing external work exists and has been exercised.

Exit also has a price, and the source material accidentally quotes it. One developer's build of the compiler held every core of a Threadripper 9970X for roughly fifteen minutes ^[16]. A fork of something that heavy is a standing staffing line, not a hedge you keep in a drawer. The right to leave is real; the cost of using it means most shops will not.

Which puts the weight on the other announcements from the same day. Modular Cloud went generally available, the MAX runtime added AWS Trainium, Google TPUs and Qualcomm's own datacentre accelerators to its existing Nvidia, AMD and Apple Silicon support, and a Windows port is coming with Microsoft's help, according to Forbes coverage of ModCon ^[8]. The write-up is blunt about what that stack is for: six chip vendors under one programming model is the product, and the language is the way in ^[9]. The compiler now sits upstream of a runtime whose backend list includes silicon sold by the company that just bought the compiler.

Nothing in the material says Qualcomm intends to starve the other five backends. The point is narrower and harder to argue with: for three years the complaint about Mojo was that the toolchain was a binary you downloaded on faith ^[1], and the fix arrived 55 days after the acquisition was announced ^[17] and about three weeks after it closed ^[18]. The dev.to account puts the sequence plainly: a chip company bought the language, and only then did the source drop ^[19].

The 409-point thread splits into people relieved they can finally try it and people who think the moment passed ^[10], and both camps are arguing about the wrong axis. The specific objections that remain are not licence objections. One developer expects Mojo to stay inside niche and senior circles until a Windows release exists, and it does not yet exist ^[13]. Others predict acqui-death, or Chris Lattner moving on inside a year and the language folding into Qualcomm's internal stack ^[14]. Against that, a financial engineer with no GPU background reports getting performant kernels running easily ^[15], which is the actual argument for adoption and has nothing to do with who owns the repo.

So the trade is legible now. You get source you can audit and, in extremis, fork at the cost of a serious build farm, in exchange for a roadmap set by a vendor with one horse in a six-horse race.