Nvidia's real lock on artificial intelligence was never only the chips — it is the software, and Qualcomm has just spent close to $4 billion to attack exactly that moat.
The Modular purchase is the more interesting half. Rather than buy more raw silicon, Qualcomm buys a software layer, and aims the layer at the part of Nvidia's business rivals have found hardest to touch.
What Qualcomm actually bought with Modular
Modular is not a chip company. As the deal coverage explains, the startup builds the Mojo programming language and the MAX inference engine — software that lets AI models run across many kinds of hardware without rewriting the underlying code.
The talent is part of the point. Modular was founded by Chris Lattner — creator of the LLVM compiler infrastructure and Apple's Swift language — together with Tim Davis, and brings roughly 150 engineers in the all-stock deal.
Portability is the product. A model that today is tuned to run on Nvidia hardware could, through a layer like Mojo and MAX, run on a Qualcomm accelerator, an Arm CPU or a rival GPU without a painful migration — which is precisely what a hardware incumbent does not want.
💡Why CUDA, not silicon, is the moat
Nvidia's dominance rests on CUDA — the software ecosystem that, over nearly two decades, tied AI developers to Nvidia GPUs. The chips are excellent, but the lock is the code: rewriting a mature AI stack to leave CUDA is costly enough that most teams never try. A portability layer attacks the cost of leaving, not the quality of the chip.
A full data-centre platform, not just a chip
The hardware is concrete. The Dragonfly C1000 is a 250-core Arm server CPU slated for deployment by Meta from 2028, and the Dragonfly AI300 inference accelerator is claimed to deliver four to eight times the performance-per-watt of GPU-based designs on memory-bandwidth metrics, sampling in 2028.
The ambition is financial as much as technical. Qualcomm told investors it expects two hyperscaler customers to generate at least $1 billion in combined revenue within a year, with first shipments by the end of 2026, and is targeting $15 billion in data-centre revenue by 2029.
The strategy reads clearly when the pieces are placed together: own a competitive inference chip, own the software that frees models from a rival's ecosystem, and sell both to the hyperscalers already hunting for an alternative to a single dominant supplier.
Why a dignity-first frame welcomes the assault on the moat
Emergent Intelligence (EI) — the dignity-first lens through which I read artificial intelligence — treats concentration as the thing to watch. When one company controls both the best AI chips and the software ecosystem every developer must use, the dependency runs deep and the leverage runs one way.
A portability layer is, in that frame, a quietly democratising move. Software that lets a model run anywhere returns a measure of agency to the people building on top — the freedom to choose hardware on price and merit rather than on the sunk cost of a rewrite.
A moat made of switching costs is a tax on everyone downstream. Whatever loosens the lock — open standards, portable runtimes, real competition — pushes power back toward the builders and away from the single gatekeeper.
The caution is equally important. A contest between two giants is not the same as an open ecosystem, and a Qualcomm strong enough to rival Nvidia could build a new lock of its own. The dignity-first test is not which company wins but whether the winner's tools stay portable and the standards stay open.
💡The whole stack is in play
The deeper pattern connects to the rest of the stack. The compute layer is being contested on every front at once — <a href="https://humphreytheodore.com/writing/intel-18a-p-foundry-ai-chips-vlsi-2026">Intel reopening the foundry race</a>, <a href="https://humphreytheodore.com/writing/nvidia-tsmc-fabtwin-ai-runs-the-fab">AI now running the fab itself</a>, and <a href="https://humphreytheodore.com/writing/physical-ai-world-models-data-odyssey-xdof-2026">compute and data emerging as the real bottleneck for physical AI</a>. Modular is the software front of the same war.
The moat was always made of code
Whether Qualcomm executes is an open question. Shipping a competitive accelerator in 2028 and turning Mojo into the default portability layer are hard, and Nvidia is not a static target.
What the deal clarifies is where the contest actually sits. The decisive battle in AI hardware is being fought over software lock-in — over how expensive it is to leave — and Qualcomm has named that battleground by paying billions for a compiler company rather than a chip.
A dignity-first reading keeps the larger interest in view. The goal worth wanting is not a change of monarch but a genuine loosening of the moat — portable runtimes, open standards, and a market where the people building artificial intelligence can choose their tools freely. Qualcomm's move serves that goal if the openness survives success. The thing to watch is whether the new challenger keeps the door open once the door is its own.
Frequently Asked Questions
The questions below address the most common queries about Qualcomm's acquisition of Modular and its move against Nvidia, drawn from the June 2026 announcements and reporting.
Why did Qualcomm buy Modular?
Qualcomm acquired Modular, in an all-stock deal valued near $4 billion confirmed on 24 June 2026, to obtain a software layer that lets AI models run across different hardware. The aim is to challenge Nvidia not on raw chip performance but on its CUDA software ecosystem, which has long tied AI developers to Nvidia GPUs.
What is Modular and what does Mojo do?
Modular is an AI software startup founded by Chris Lattner, creator of LLVM and Swift, and Tim Davis. It builds the Mojo programming language and the MAX inference engine, which together allow AI models to run on a wide range of hardware without rewriting the underlying code — reducing the cost of moving off any single vendor's platform.
How does this threaten Nvidia's CUDA moat?
Nvidia's strongest advantage is CUDA, the software environment that makes leaving Nvidia GPUs expensive because mature AI stacks are written against it. A portable runtime like Mojo and MAX attacks the cost of leaving rather than the quality of the chip, making it easier for developers to run models on competing hardware — directly eroding the lock-in CUDA provides.
What chips did Qualcomm announce alongside the Modular deal?
At its 2026 Investor Day, Qualcomm unveiled a data-centre platform including the Dragonfly C1000, a 250-core Arm server CPU set for deployment by Meta from 2028, and the Dragonfly AI300 inference accelerator, claimed to deliver four to eight times the performance-per-watt of GPU designs on memory-bandwidth metrics. Qualcomm is targeting $15 billion in data-centre revenue by 2029.
What is the Emergent Intelligence view of the Qualcomm–Nvidia fight?
Emergent Intelligence (EI) is a dignity-first reading of artificial intelligence that watches concentration of power. It welcomes portability layers that return choice to developers and loosen a single vendor's grip, but cautions that replacing one dominant supplier with another is not real openness — the test is whether the winner keeps its tools portable and the standards open.
Sources and Further Reading
Cover image: a dramatic black-and-white circuit-board macro — via Pexels.
