Qualcomm in AI: Snapdragon, On-Device Models & Data Centers

From CDMA patents to the chip in most Android phones

Qualcomm was founded in San Diego in July 1985 to commercialize CDMA, a wireless technology that became the basis for much of mobile communications, and it still earns a large share of its revenue from licensing patents on the wireless standards that grew out of that work. Its better-known business today is Snapdragon, the family of mobile processors that ships inside most premium Android phones, plus modems and, more recently, laptop chips. Cristiano Amon is chief executive. [1]

Every current Snapdragon 8-series chip, including the Snapdragon 8 Elite Gen 5, ships with a Hexagon neural processing unit as a dedicated block alongside the CPU and GPU, marketed under Qualcomm's AI Engine branding. The pitch is running AI workloads, voice processing and vision features on the device itself rather than sending them to a server, which Qualcomm frames as the way phones and laptops will keep adding AI features without depending on a network connection or a subscription to someone else's cloud. [2]

A license pulled over an AI laptop

In May 2024, the US Commerce Department revoked the export licenses that let Qualcomm and Intel ship chips to Huawei for laptops and handsets, effective immediately. Huawei had launched an AI-branded laptop running an Intel chip weeks earlier, and members of Congress pressed the administration on how a company under export restrictions was still getting American chips at all; the revocation followed. The episode shows Qualcomm on the receiving end of AI-era export control rather than shaping it, since the trigger was a rival's laptop and not anything Qualcomm itself shipped. [3]

Qualcomm goes after the data center

On October 27, 2025, Qualcomm announced its first data center AI accelerators, the AI200 and AI250, aimed squarely at the AI inference market Nvidia and AMD have dominated. The AI200 is expected in 2026 and supports 768 GB of memory per card; the AI250 follows in 2027 with a near-memory computing design Qualcomm says delivers more than ten times the effective memory bandwidth of conventional approaches. Both ship in full, liquid-cooled server racks drawing 160 kilowatts, with Qualcomm committing to an annual product cadence afterward. Qualcomm's stock rose about 20 percent the day of the announcement. [4][5]

The same day, Qualcomm and HUMAIN, the Saudi state-backed AI company, announced a plan for HUMAIN to deploy 200 megawatts of Qualcomm AI200 and AI250 racks starting in 2026, part of a broader partnership the two companies had previewed at the Saudi-US Investment Forum in May 2025. It is a commitment with a capacity figure and a start date attached, not a stated intention to explore data center chips at some point. [6]

Atlas interpretation: Qualcomm's two AI stories run in different directions. On phones and laptops it already ships silicon at a scale few AI chip companies can match, and its argument is that keeping inference local is a genuine advantage, not just a mobile constraint dressed up as a feature. In the data center, it is a first-generation entrant competing against Nvidia's years of software lock-in with CUDA and AMD's own accelerator line, and the AI200 and AI250 are Qualcomm's attempt to buy a seat at that table with a memory-bandwidth argument rather than a raw compute one. [4][6]