The compiler legend who wants to break AI free from a single chip

The most reliable way to make a fortune in artificial intelligence, for a while, was to sell the shovels: NVIDIA's chips, and the CUDA software that made those chips the only ones worth buying. Modular was built on the opposite hunch. What if the software layer everyone leaned on could be pried loose from the hardware underneath it - so a model could run just as happily on an AMD card, an Apple chip, or something nobody has shipped yet?

That is the whole company in one sentence, and it is a strange thing for a startup to attempt. It means competing not with a product but with a habit - the deep, industry-wide assumption that "AI" and "one vendor's stack" are the same thing. Modular's founders thought the assumption was expensive, brittle, and beatable. Enough investors agreed to hand them roughly $380 million, and in 2026 Qualcomm agreed enough to buy the whole thing.

A resume that keeps solving the same problem

Modular's co-founder and CEO is Chris Lattner, and if you have used an Apple device in the last fifteen years, you have run his code. As a graduate student he started LLVM, the compiler infrastructure that now underpins an enormous share of modern software. He built Clang. He created Swift, the language Apple uses across its platforms. Later he co-created MLIR, a compiler framework aimed squarely at machine learning. Between those, he did tours at Apple, Tesla as VP of Autopilot Software, Google on TensorFlow infrastructure, and SiFive.

His co-founder and Modular's president is Tim Davis, whom Lattner met at Google. Davis co-created TensorFlow Lite, the on-device machine-learning runtime that ended up on billions of phones, and led product for a stack of Google's AI infrastructure - TPUs, XLA, MLIR. Read the two resumes side by side and the pattern is hard to miss: both men spent years building the plumbing that lets AI actually run, and both concluded the plumbing was a mess worth leaving Google to fix.

Three layers, one bet

Modular calls its product "AI's unified compute layer," which is marketing shorthand for a real stack of three parts. Each sits a little higher than the last, and each is useless as a lock-in play - the point is that they run anywhere.

Mojo is the piece people fall in love with. The pitch is almost cheeky: keep the Python you already write, delete the part where Python is slow, and let one language target the accelerator directly. Modular has claimed speedups of up to 35,000x over pure Python on certain workloads - the kind of number you treat as a headline, not a guarantee, but one that gets a systems programmer's attention. Its mascot, fittingly for a language meant to run hot, is a flame emoji.

The customer list gives the game away

Modular sells to the people running AI at industrial scale. Its named customers and partners include Inworld, SF Compute, Jane Street, Oracle, AWS, Lambda Labs, TensorWave, AMD and NVIDIA - a list that mixes quant traders, cloud giants, GPU-rental shops, and, tellingly, the chipmakers themselves. The business model is consumption-based: you pay for the compute and the inference requests you push through the platform, cloud-style. The core language and much of the stack are open source, which is the on-ramp; the money is in enterprise deployment, cloud usage, and support.

That NVIDIA and AMD both show up as partners is the neatest trick in the whole story. Modular is, in one reading, trying to erode NVIDIA's software moat. In another, it is the neutral broker every hardware maker outside NVIDIA badly wants to exist - a way for their silicon to feel as easy to use as the incumbent's. Sitting in that seat is what made the company strategically valuable.

From a $30M seed to a Qualcomm exit

Modular emerged from stealth in 2022 with a $30 million seed round led by GV. A year later it raised $100 million, led by General Catalyst, at roughly a $600 million valuation. In September 2025 it raised a $250 million Series C - led by the US Innovative Technology Fund, with DFJ Growth joining - that valued the company at $1.6 billion and introduced Mammoth. Then, in June 2026, Qualcomm announced it was acquiring Modular in an all-stock deal valued at roughly $3.9-4 billion; the deal closed the following month.

One number needs a caveat: the roughly $4 billion is the acquisition value, not the venture funding. As an independent company Modular raised about $380 million in equity. It is worth saying plainly, because the gap between "raised $380M" and "sold for ~$4B" is the entire point of building infrastructure the incumbents cannot ignore.

What a chipmaker actually bought

Qualcomm did not buy Modular for its revenue. It bought a position. Qualcomm wants a foothold in data-center AI, and the hardest part of that fight is not silicon - it is the software that makes silicon usable, the thing NVIDIA spent two decades building with CUDA. Modular arrived with a horizontal, developer-friendly stack that already ran across many kinds of hardware. For a company trying to offer AI customers real choice about where they deploy, that is close to the ideal acquisition: a ready-made answer to the software problem, and the team that built it.

The playbook, if you want to borrow it

There is a repeatable move buried in Modular's story. Pick the layer nobody wants to own because it is genuinely hard - here, the compiler and runtime that sit between models and metal. Build it for real, not as a demo. Open-source the developer-facing piece so adoption compounds, and monetize the operational piece where scale lives. Then the incumbents come to you, because you are sitting in the seat they need and cannot easily build themselves. Modular ran open-source-then-exit as a single strategy, and both halves reinforced each other.

The conditions where it does not work are worth naming too. This only pays off if the middle layer is truly painful and truly portable - if the incumbent's lock-in is loose enough to pry and the alternative hardware is good enough to matter. Modular also had a rare asset most teams cannot summon: a founder whose name alone signals that the compiler will actually be excellent. Absent that credibility, "we're building a neutral AI compute layer" is a much harder sell to both developers and acquirers.

The neutral ground in an AI turf war

Modular's competition is really a set of moats and the tools that reinforce them: NVIDIA's CUDA, cuDNN and TensorRT on one side, AMD's ROCm, OpenAI's Triton, vLLM and TVM-style compilers scattered around. Modular's differentiator was never being the fastest at any single thing on any single chip - it was refusing to pick a chip at all. In a market where nearly every other layer is optimized to keep you on one vendor, a stack whose entire reason for existing is portability occupies ground almost no one else wants to defend. That is an unglamorous place to stand. It also turned out to be worth several billion dollars.

Founded 2022 in Los Altos, California, by Chris Lattner and Tim Davis. Products: Mojo, MAX, Mammoth. ~130+ employees at 2025 Series C. $1.6B valuation (Sept 2025). Acquired by Qualcomm in a ~$4B all-stock deal (2026). Figures for token volume, speedups, and headcount are as reported by the company and press; treat the largest numbers as approximate.