MODULAR

Chris Lattner has one of the best engineering résumés in tech. He created LLVM, the compiler tech that runs under huge chunks of modern software.

He built Swift, Apple's programming language. He ran teams at Google on TensorFlow, the machine learning toolkit.

Everywhere he looked, he saw the same problem. AI software was a tangled mess, and almost all of it was chained to Nvidia's chips through Nvidia's CUDA software.

In January 2022 he teamed up with Tim Davis, who had led machine learning interfaces at Google, to fix it. They called the company Modular.

The plan was to rebuild the whole AI software stack from scratch.

Modular sells the plumbing that runs AI. Companies training and running AI models need software to squeeze the most out of their chips.

Today that mostly means Nvidia's software, which only works well on Nvidia hardware. Modular built MAX, a platform that runs AI fast on both Nvidia and AMD chips from one setup.

It also built Mojo, a programming language that looks like Python but runs far faster. The money comes from selling and licensing this software to companies that want better AI performance without being locked to a single chipmaker.

Two things. MAX is the platform that runs AI models fast across Nvidia and AMD hardware from a single, unified setup.

Mojo is the programming language, built to feel familiar to Python developers while running dramatically faster, aimed at the heavy number-crunching that AI needs. Together they are pitched as the unified compute layer for AI.

Modular went straight at the pain point everyone in AI feels but few say out loud. Nvidia chips are scarce and expensive, and CUDA makes it painful to switch.

Modular's hook is freedom. Write your AI software once, run it on whatever chip you can actually get your hands on.

The company opened Mojo and parts of the stack to developers to build a community around it. The bet is that if enough engineers adopt the tools, the big AI companies follow, because nobody wants to be trapped buying one company's chips forever.

Nvidia is the challenge. Its CUDA software has a 15-plus-year head start and millions of developers already using it.

Convincing the AI world to switch to a younger stack is brutally hard, even when the alternative is genuinely good. Modular also has to deliver on a huge promise, top performance across different chipmakers' hardware, which is easy to claim and hard to prove at scale.

And it is burning serious money to do it. That $380 million raised is also $380 million in expectations.

Modular's backer list reads like an AI who's who. The 2022 seed was led by GV, Google's venture arm.

General Catalyst led the $100 million Series A in 2023. The big $250 million round in 2025 was led by US Innovative Technology, the fund run by Thomas Tull, with DFJ Growth joining and earlier backers like Greylock returning.

The money signals serious belief that Nvidia's software grip can actually be loosened.