Nine months ago, Modular raised $250 million at a $1.6 billion valuation. Qualcomm is now reportedly ready to pay $4 billion for the company, a 2.5x step-up in three quarters, with no disclosed revenue below it to anchor the math. Bloomberg broke the talks on June 22, Reuters backed the funding detail, and the thing is advanced rather than signed, so it could still fall apart before anyone uncaps a pen. What makes the number worth staring at is not Modular by itself. It is the four other things Qualcomm is buying at the same time, and the fact that Cristiano Amon walks on stage in two days to explain it all.
Modular does not make chips. Chris Lattner and Tim Davis built a compiler platform that optimizes AI workloads on whatever hardware you point it at, plus Mojo, a language that reads like Python and compiles to native code that runs near C or Rust. $380 million raised across its life. Every headline calling it an AI chip startup has it wrong, and the error matters because the compiler is where Nvidia’s moat actually lives, and Qualcomm does not own one.
Write a model in PyTorch or JAX, and something has to lower it into actual hardware instructions, the matrix multiplies and fused kernels that run on one specific accelerator. That lowering of the framework down to silicon is where performance is won or wasted, and it is where Nvidia has spent twenty years digging in. CUDA is not a driver or an API. It is the accumulated tooling and habit of two decades, a whole profession that learned to write for Nvidia hardware before anything else. Switching off Nvidia was never really about the hardware. The re-platforming underneath it has no bottom, and faster silicon does nothing to touch that.
MLIR is how it gets at that. The Multi-Level Intermediate Representation came out of Google; Lattner helped build it before he founded Modular, and it lets a compiler hold a computation at several layers at once, from high-level tensor math down to a single chip’s instructions, optimizing at each level. Compile a model once, and it can target Nvidia, AMD, Qualcomm’s Snapdragon accelerators, or whatever data-center part ships next year. Mojo sits on top. It feels like Python in a field that standardized on Python a decade ago, while still exposing the low-level controls an engineer needs to hand-write a custom attention kernel without having to drop into C++ or CUDA. Hot path in Mojo, near-optimal code on Qualcomm silicon, no CUDA anywhere in it.
Apache TVM, Google’s XLA, OpenAI’s Triton, Qualcomm’s own AI Stack, and Intel’s OpenVINO. Every one of them set out to do roughly what Modular is doing, every one benchmarked well, and not one has pulled enough of the field off CUDA to matter. The space is littered with compilers that died in deployment after looking great on paper. Modular is early and unproven, and sitting in that same pile until it isn’t. The $4 billion price is an escape from it that has not happened yet.
Chris Lattner built LLVM, the compiler infrastructure that underpins most of the industry’s toolchains today, including Apple and Google. Clang came next and shoved GCC out of serious development. Then Swift at Apple, then MLIR at Google. LLVM is the reason most modern compilers are built the way they are. Lattner is also one of the very few compiler people who ever got the industry to adopt his work at scale, the exact part that kills most projects in this space. So a real slice of the $4 billion is an acqui-hire of the most consequential compiler engineer working and the team he built around him. Qualcomm is a wireless-IP and SoC shop. It cannot build software infrastructure like this from scratch on any timeline that matters, so it is worth paying to skip the wait.
The Oryon cores in every flagship Snapdragon were not designed by Qualcomm. They came from Nuvia, the startup Qualcomm bought in 2021 for $1.4 billion, founded by Gerard Williams III after years running Apple’s CPU designs. Qualcomm bought Nuvia for Williams and his team, and the custom-core work it had tried and failed to grow internally for most of a decade, the exact structure of the Modular deal. It worked: Oryon is why Snapdragon finally matches Apple Silicon. It also cost three years in court. Nuvia’s Arm licenses set Arm off, Arm moved to void the designs and pull Qualcomm’s architecture license, and Qualcomm did not get a clean title until a jury, and then a judge ruled for it in late 2025. The IP Qualcomm wanted came attached to someone else’s lawsuit.
Tenstorrent is the next one, and the biggest. Qualcomm is in talks to buy Jim Keller’s AI-chip company at a reported $8 to $10 billion, on top of two it has already closed: Alphawave for high-speed SerDes and connectivity IP, Ventana for RISC-V server CPU cores. Four targets, an estimated $16-$18 billion, all-in. Tenstorrent and Ventana are the silicon. Alphawave and Modular are what wrap around it, the interconnect and the compiler.
Tenstorrent and Ventana are both RISC-V, the open instruction set that owes Arm nothing, not in license fees and not in court. A company three years out of Arm litigation is rebuilding its compute stack on an architecture Arm has no power to revoke. Tenstorrent’s accelerators run inference, and Modular’s compiler programs them, with Alphawave moving data between them, and nothing in that chain touches CUDA or Arm. Keller is the draw on the silicon. He did AMD’s Zen and Apple’s early A-series, then the Tesla self-driving chip, and his designs have long been known for delivering more performance per watt than the field. Intel was circling Tenstorrent too, per Bloomberg’s May reporting, before Qualcomm’s number cleared the prior $3.2 billion fundraising talk and kept going. Deep pockets in the room drive up the price.
| Target | Reported Value | Status (as of June 22, 2026) | Core Asset |
|---|---|---|---|
| Modular Inc. | ~$4B | Advanced talks | AI compiler platform, Mojo language |
| Tenstorrent | $8B–$10B | Negotiations | Efficient AI accelerator chips |
| Alphawave Semiconductor | Undisclosed | Completed | High-speed SerDes, connectivity IP |
| Ventana Microsystems | Undisclosed | Completed | RISC-V server CPU cores |
Qualcomm is worth around $237 billion and is still the biggest wireless-chip vendor on earth by revenue, and that dominance is also the cage. Smartphone volumes move in cycles, the Android market grinds prices down every year, and Apple, the largest single buyer of Qualcomm modems, has been building its own to replace them for years. Amon has said outright that the company has to move into automotive, the industrial edge, and, most of all, data-center AI. On June 24, two days out, he takes the stage at Qualcomm’s Investor Day to sell exactly that, growth and diversification, and the stock has already run up about six percent into the date. The acquisitions are the spending side of that pitch. And the order is backward: Amon is committing $16 to $18 billion to a strategy he has not yet proven on stage, while the bears note that the multiple is already priced in for a data-center business with almost no revenue.
NVIDIA owns somewhere between 70 and 90 percent of AI-accelerator revenue, depending on how you draw the lines. AMD has clawed out a foothold with the MI300X and its successors. The hyperscalers are all building their own, Google’s TPUs furthest along, Amazon and Microsoft behind with Trainium and Maia, each able to match Nvidia’s raw compute on the workloads they actually care about, much of it on the same TSMC leading-edge silicon built for this moment. None of it has dented the lock-in. Matching the chip prevents you from running the existing universe of CUDA-tuned models and libraries without a brutal port. That port is the whole problem, and Modular exists to delete it.
If Mojo and the Modular compiler become the easiest way to get a model onto Qualcomm silicon, the developer writes it once, ships it to the Qualcomm part, and never has a reason to open CUDA again. That is the entire play. You do not have to beat CUDA if developers never have to choose it. On paper, it is clean. What stands in the way is not technical.
Most production AI is built with PyTorch, which is tightly integrated with CUDA. Triton compiles to CUDA-compatible hardware first, and the research that becomes next year’s production ships CUDA by default. That is the inertia Modular has to overcome, and it does not move toward a better tool. It moves for years of unglamorous integration work, and Lattner’s name buys attention for that without buying the result.
Nine months, $1.6 billion to $4 billion, no revenue in between. Compiler engineers who can ship something that survives production barely exist, so Lattner’s team is a concentration nobody reassembles by posting job openings. There is an auction underway, with Qualcomm, Intel, and probably others bidding over a tiny pool of assets that might end up mattering. Under all of it sits a wager about where value in the AI stack pools over the next ten years. Make the compiler layer a real chokepoint, and $4 billion looks cheap. If it never becomes one, if CUDA simply keeps the grip it has held for twenty years, the number is a write-down on a timer. Modular’s investors, in for $380 million all in, walk away fine either way. Qualcomm is the one left holding the variance.
Four acquisitions at once, $16 to $18 billion, folded in while the wireless business keeps paying for everything. That is the part that should worry people, well past the price. Modular is a developer-tools shop built around a charismatic founder. Tenstorrent is a chip house with loud architectural opinions and a founder who has walked out of big companies more than once. Folding both into an outfit that has been handset SoCs and licensing for forty years is the most operationally aggressive thing Qualcomm has tried, and big semiconductor deals usually end as a list of cultures that never merged and engineers who left once the earnout cleared. None of that means the deal is wrong. A company serious about AI infrastructure has no safe option here, and Qualcomm took the unsafe one on purpose, with Nvidia, AMD, and the hyperscalers waiting to split the decade if it stumbles. Watch whether Tenstorrent closes alongside Modular. If it doesn’t, Qualcomm owns the best compiler team in the world, and no one owns fab-made silicon of its own to run it on.