The AI models that run on today’s flagship phones hold somewhere in the tens of billions of parameters. Cristiano Amon believes that number is about to grow by a factor of ten, and he is building the chip business around that bet.
Qualcomm’s chief executive said October 11 that AI companies want to run 100-billion-parameter models locally on smartphones by 2028. That would push models now served from data centers onto the silicon inside a pocket, and it would demand far more on-device compute and memory than any phone now delivers.
The gap is real and measurable. Today’s leading phones run models in the single-digit-billions to low-tens-of-billions range, roughly an order of magnitude below the 100-billion mark Amon describes. Closing it means solving for memory bandwidth, power draw, and heat, three constraints that have always defined what a phone can do.
Qualcomm has been building AI into its chips for years, long before the current wave of large models. The neural processing units inside Snapdragon were initially aimed at camera effects, voice recognition, and battery management. The arrival of generative AI shifted the question from whether a phone could run a small model to how large a model it could run, and on-device AI became a marketing battleground as much as a technical one.
For Qualcomm, the leap is an opportunity to redefine what it sells. The company built its name on the Snapdragon processors inside most Android phones, and it has spent years adding neural processing units to those chips. A market where the AI workload runs locally, Amon argues, turns that silicon into the scarce resource, and the phone chip maker into an edge-computing supplier.
The company put money behind the argument in June, when it agreed to buy Modular, an AI software startup, for about $3.9 billion. The all-stock deal, completed in July, brought Chris Lattner, Modular’s co-founder and a compiler engineer known for creating LLVM and Swift, into Qualcomm as an executive vice-president, and tied data-center inference software to the on-device NPU. The message was that Qualcomm would sell the software layer as well as the chip.
Amon’s 2028 timeline is a forecast, not a commitment, and plenty of engineers think it optimistic. Running a 100-billion-parameter model on a phone means holding tens of gigabytes of weights in memory fast enough to generate a token in real time, and keeping the device cool while it does so. Phones have made those trade-offs before, camera processing and gaming pushed silicon harder each year, but the AI jump is larger.
The 100-billion figure is a round number with a specific meaning. Models of that scale, roughly the size of some frontier systems, would currently require far more memory than a phone carries. Compressing them into a handset means quantization, the art of shrinking a model’s precision to fit into less memory, plus faster memory interfaces and better heat management. Each of those is an engineering problem Qualcomm can sell.
The push toward local AI is driven by cost and control as much as performance. Every query that stays on the device is a query a cloud provider does not bill for, and one that never leaves the user’s hand. Privacy-sensitive applications, from health to finance, favor the on-device route, and latency drops when the model lives beside the sensor.
Competition is already moving. Apple has built its own silicon path for on-device intelligence, MediaTek is pressing into the same market, and cloud model-makers are racing to compress their largest models into forms a phone can digest. Qualcomm’s advantage is the breadth of the Android market and the software foundation it just acquired, including Modular’s Mojo language.
The ambition reaches beyond phones. Qualcomm pushed Snapdragon into Windows laptops in 2024 and has pitched the same silicon as the foundation for on-device AI in cars, headsets, and industrial equipment. Amon, chief executive since 2021, has argued that the companies that own the silicon closest to the user will capture the value as AI moves off the cloud.
The memory question is the sharpest constraint. A 100-billion-parameter model stored at full precision needs hundreds of gigabytes, more than any phone ships with, so the industry’s path runs through lower-precision formats and faster memory. Qualcomm has been pushing on-device memory and NPU throughput for several product cycles, and Amon’s remarks position those quiet engineering gains as the foundation of a larger commercial story.
The question Amon has not answered is whether the models or the chips arrive first. If the industry reaches 100 billion parameters on-device by 2028, Qualcomm wants to be the one selling the silicon underneath. If the models stall short of that number, the company has still moved closer to the edge than most rivals. Either way, the phone is becoming the next front in a fight over where AI runs.
