Qualcomm is preparing a data-center chip with 250 cores aimed at the fast-growing market for AI inference, according to a Fortune report, a move that pits the mobile chip maker directly against Nvidia’s dominance in the business of running AI models. The company has paired the hardware push with a $3.9 billion acquisition of Modular, a software startup, in an effort to build the full stack it says Nvidia lacks.
The all-stock deal for Modular, announced last week, gives Qualcomm the software foundation for running AI models efficiently across different chips, a capability the company’s chief executive, Cristiano Amon, has called essential to its data-center ambitions. Modular’s founders include Chris Lattner, a former Apple engineer who created the Swift programming language, and its open-source Mojo language and MAX platform are used by developers to optimize models for specific hardware. Under the terms of the deal, Qualcomm will issue about 19.2 million shares, and the transaction is expected to close in the second half of the year. Lattner will join Qualcomm as executive vice president of advanced AI software and platforms.
The strategy is a study in contrast with Nvidia. Nvidia’s dominance rests on its H100 and next-generation accelerators paired with high-bandwidth memory, or HBM, an expensive and supply-constrained component that has become a bottleneck for the entire industry. Qualcomm’s approach is built on commodity memory and a chip design that spreads work across hundreds of smaller cores, an architecture the company argues delivers comparable inference performance at a fraction of the cost. The difference in memory economics is the heart of the pitch: HBM costs several times more per gigabyte than standard DRAM, and for workloads that are memory-bound, that gap can decide who wins the contract.
Inference is where Qualcomm sees its opening. Training models is a concentrated, hyperscaler-dominated business, but running models repeatedly for millions of users is a volume business with different economics. Analysts estimate that inference will account for a growing share of AI computing demand as applications mature, and that cost per query, not raw speed, will decide which chips win. “Nvidia built the engine. The next fight is about the fuel bill,” one semiconductor analyst said.
The 250-core chip is the centerpiece of that argument. Details remain limited, but people familiar with Qualcomm’s plans describe a processor designed for the heavy, continuous workloads of serving large language models, where memory bandwidth and power efficiency matter more than single-thread speed. Qualcomm has spent years adapting its phone chip designs for servers, and it acquired Ventana Micro Systems last year to gain RISC-V server CPU expertise. The company has also described work on dozens of custom designs for AI gadgets, from glasses to earbuds, but the data-center chip is its most direct attack on Nvidia yet.
The bet extends beyond silicon. Modular’s software lets developers write a model once and run it across CPUs, GPUs and custom accelerators without rewriting, an approach that threatens the lock-in Nvidia has built through its CUDA programming environment. If Qualcomm can make its hardware attractive without forcing developers to learn a new stack, it removes the largest practical barrier to switching. That is why the software acquisition is as important as the chip: hardware without a developer base has failed before, and Qualcomm is buying its way past that problem.
The market has taken note. Qualcomm shares have been volatile this year as investors weigh the company’s phone business against its AI ambitions, and the Modular deal was greeted with a mix of skepticism and optimism. The company is also designing custom accelerators for data-center customers, with China’s ByteDance reported to be an early client, a sign that Qualcomm is willing to compete for the same custom-chip contracts that have enriched Broadcom. Each of these bets is small on its own; together they describe a company that has decided the phone business alone is no longer enough.
The challenge is execution. Nvidia’s installed base, software ecosystem and supply-chain relationships are formidable, and its next-generation platforms are already sold out. Qualcomm has been here before: it has tried repeatedly to enter the server market, and each attempt has ended quietly. What is different this time, executives say, is that the industry is young enough that standards have not hardened, and that the economics of inference favor the low-cost challenger.
Whether that argument holds will be tested in the next two years. The 250-core chip needs to ship, the Modular acquisition needs to close, and customers need to buy. If all three happen, Qualcomm will have done something no one else has managed: put real pressure on Nvidia’s richest business. If any one of them slips, the company’s data-center story will look like the last decade’s all over again.
