On 29 July, Qualcomm closed a $3.92 billion all-stock deal for a startup called Modular. Qualcomm makes chips, but it did not buy Modular for silicon. It bought a software layer, one designed to let AI models run across processors from different vendors without anyone rewriting their code. That detail is the whole story. It is the clearest admission yet, backed by nearly four billion dollars, that the industry now believes Nvidia's dominance rests on software, and that the way to attack it is not with a faster chip but with a bridge around the one nobody wants to rebuild.
The moat in question is CUDA, the programming framework Nvidia launched back in 2007. Over nearly two decades, AI researchers built their tools, libraries and habits on top of it, and CUDA runs only on Nvidia hardware. That is the trap. A team that has wired its entire workflow into CUDA faces brutal switching costs to move elsewhere: code rewritten, tooling replaced, engineers retrained. Even when a rival accelerator looks cheaper on paper, the migration bill often wipes out the saving. Analysts credit that friction, more than raw silicon speed, with keeping Nvidia's share of the AI accelerator market somewhere between 70 and 80 percent.
Modular's pitch is to dissolve that friction. If the same code runs on any accelerator, then hardware stops being a lock-in and becomes a commodity, bought on price, availability and power efficiency. Those happen to be exactly the terms Nvidia's rivals would prefer to fight on. Qualcomm, a marginal player in the data centre against Nvidia, AMD and the hyperscalers' in-house chips, does not need to win a benchmark war if portability lets customers move existing workloads onto Qualcomm silicon wherever the economics favour it, especially in efficient, power-constrained inference. Notably, Qualcomm paid roughly two and a half times Modular's valuation from just nine months earlier, a measure of how urgent the prize has become.
Qualcomm is not attacking alone, and that is what turns a single acquisition into a trend. AMD spent its Advancing AI 2026 keynote making the same argument from a different direction, pushing its open-source ROCm stack as a CUDA alternative and unveiling Helios, a rack-scale system it claims runs 10 to 15 percent faster than Nvidia's comparable hardware, while securing gigawatt-scale accelerator commitments from OpenAI and Meta. Independent comparisons still hand CUDA a 10 to 30 percent performance edge and a far broader library ecosystem, so the moat has not been drained. But ROCm now costs meaningfully less and gained official PyTorch support, and the gap is narrowing rather than widening. Chinese labs, cut off from top Nvidia parts by export controls, have been quietly proving the same point by training and serving competitive models on other silicon out of necessity.
Nvidia's response reveals how it sees the threat. Rather than defend the software end, it is reportedly locking in its largest customers through capital, in talks to guarantee around $250 billion of financing for OpenAI's Ohio data centre campus plus a further sum toward chip purchases. Binding buyers with money rather than code is a rational move, but it is also a tacit concession that the code moat alone may no longer hold. The strategic question the Modular deal poses is the uncomfortable one for a company worth trillions: if a portability layer can make AI workloads hardware-agnostic, the most valuable asset in the industry may turn out to have been the software nobody bothered to rewrite, right up until someone finally did.