Qualcomm flew members of the IDC team to Maui for this year’s Snapdragon Summit, and the theme of the week has been agentic AI. CEO Cristiano Amon used the day-one keynote to argue that the smartphone is moving from an app-centric model, where the user starts every task, to an agent-centric one, where software acts on the user’s behalf and is instant, dependable, and frictionless. “In the agentic age, the experience begins with intent,” he said. On day two, Qualcomm took that argument well past the phone, to PCs, glasses, watches, earbuds, and emerging form factors such as pendants.
Day one: The smartphone remains central
Amon was clear that the phone isn’t going away. His thesis is that it becomes more important, because it’s the device that knows you best and is always with you. He also set a practical bar for adoption: people will hand tasks to an agent only if they trust it and if it’s easier than doing the task themselves.
Google’s Rick Osterloh joined Amon on stage with numbers that show how much of this currently runs in the cloud. Gemini has reached 1 billion monthly users, and token usage is up 7x year over year. So what’s the value of running AI locally?
Qualcomm’s answer is the personal layer. Amon said an on-device agent needs personal knowledge, secure permissions, the ability to act on your behalf, persistent context, real-time reasoning, and multimodal input. Most of that depends on what the phone knows about you.
Qualcomm also teased a co-processor called High Bandwidth Compute (HBC) for Snapdragon that moves memory and AI compute closer together, headed for smartphones, laptops, glasses, and cars. In its datacenter accelerator cards and other datacenter implementations, the term HBC is a play on High Bandwidth Memory (HBM) for servers and refers to a near-memory implementation for processing AI workloads. Integrating this into its device SoCs could make Qualcomm more competitive on on-device AI workloads, with faster performance and lower power consumption. More details are expected at Mobile World Congress in 2027.
Two new flagship smartphone chips
For the first time, Qualcomm split its top mobile platform in two: the Snapdragon 8 Elite Gen 6 and the 8 Elite Extreme Gen 6. Qualcomm’s thesis is that one part can’t serve a flagship tier that now spans slabs, Pro Max models, foldables, camera phones, and gaming devices. With memory pricing disrupting the status quo, offering a second option gives OEMs more room to manage cost while prioritizing higher efficiency or higher performance depending on the use case.
The Gen 6 is the mainstream flagship part, on a 2nm process with the Oryon CPU and LPDDR5. The Extreme is the more interesting of the pair. Qualcomm says its Oryon CPU is the first in mobile to break 5 GHz. It moves to LPDDR6, adds matrix cores inside the Adreno GPU, and puts a heat slug directly on the die, which Qualcomm says doubles how long the chip can hold peak performance.
Qualcomm says the Extreme’s Hexagon NPU can run 30-billion-parameter mixture-of-experts models directly from flash. Its redesigned Sensing Hub powers Personal Scribe, which builds a personal knowledge graph on the device. That’s the persistent context Amon said agents need.
Xiaomi’s 18 Pro uses the Gen 6, and the 18 Pro Max uses the Extreme, with Motorola, OnePlus, OPPO, vivo, and others in line. Qualcomm says the price gap between the tiers will likely land around $100.
Day two: Agents across more devices
Kedar Kondap, who runs Qualcomm’s compute business, boiled it all down to one question: Where should inference run? He said the cloud brings the largest models and the deepest reasoning, while devices bring low latency, privacy, and personal context. But at the scale Qualcomm is talking about, he argued, the economics of sending every token to the cloud simply don’t work. That’s a more defensible pitch than the AI PC claims of the last few years.
New PCs on Snapdragon X2
Kondap called the PC the anchor node, with the thermal headroom and memory to run larger models locally. More than 7,300 apps now run on Windows on Snapdragon, he said, across 170-plus designs.
Microsoft’s Brett Ostrum introduced the next Surface Pro 12-inch and Surface Laptop 13-inch, both on Snapdragon X2 Plus. Microsoft claims up to 95% faster on-device AI and more than 60% faster graphics gen-over-gen, plus 25% brighter displays. The Pro starts at $1,150 and adds an optional 5G model for business. The Laptop starts at $1,200 and is rated for up to 22.5 hours of video playback. Both ship October 13 in select markets, along with a new $80 haptic Surface Mouse.
Developers have wanted Linux on Snapdragon for years. Qualcomm said it’s now officially supported, starting with the X2 series. Debian arrives at the end of this year. Ubuntu certification from Canonical follows in the first half of next year, along with ASUS Zenbook A14 and A16 models running it. Kondap also pointed to HUMAIN’s Horizon Ultra, an X2 Elite PC unveiled at LEAP that runs a Linux-based agentic OS.
Googlebook: A new laptop category
Googlebook broke as its own announcement on Monday, and Google’s John Solomon filled in the details on stage. It’s a new laptop category that combines Chrome OS and Android, aimed at Android phone owners. Solomon pitched it as the end of the “toggle tax” between phone and laptop, with access to phone files, native desktop Android apps, full desktop Chrome, and Gemini features such as Magic Pointer.
Worth noting: Googlebook runs on the first-generation Snapdragon X Elite, not the new X2. Dell and HP are first, and pre-orders are open. You’ll find IDC’s full coverage on Googlebook here: New Player Has Entered: Googlebook Is a Modern, AI-Forward Take on Notebooks.
Glasses, watches, and pins
On wearables, Qualcomm said Snapdragon AR1 and AR1+ are the first personal AI platforms to support multimodal 1-bit models, built on work with PrismML that cuts memory needs enough to run larger models on glasses.
Google’s Shahram Izadi showed the Snapdragon-based Pixel Watch 5 and Android XR devices from Samsung and XREAL. Google is also working with Gentle Monster and Warby Parker on AR1 glasses due later this year.
Qualcomm’s Snapdragon START program packages an AR1 module, software, and white-label designs to get glasses to market faster. UK eyewear maker Inspecs showed camera-equipped enterprise safety glasses and said it went from kickoff to working device in four months. Awear showed an agentic OS built around eyewear. Qualcomm also previewed a 5G personal AI pin reference design on Snapdragon Wear Elite, and formed a Personal AI Health Alliance around uses such as fall-risk prediction and clinical notes.
Snapdragon Sound Elite Gen 2 introduced
For earbuds, Qualcomm introduced Snapdragon Sound Elite Gen 2, which it calls the first premium audio platform built for personal AI. Versus the S7 Gen 1 Sound platform, Qualcomm claims a 30% smaller footprint, twice the AI performance, and up to 40% lower power. It also adds integrated micro-power Wi-Fi, so earbuds can reach cloud services and agents without a phone.
That enables a Snapdragon Sound Apps and Agents ecosystem, starting with MicroEJ, Cinemo, and Roon. HP, Bose, and Cleer are developing products.
Software: Memory layers and developer tools
Liquid AI, whose models already run in Mercedes-Benz cars on Snapdragon silicon, announced Liquid Context, a memory layer that lets agents share context across devices. And Chris Lattner, who joined Qualcomm through its Modular acquisition, said Modular’s stack, including the open-source Mojo language, is coming to Snapdragon and Qualcomm’s Dragonfly datacenter chips. Mojo is intended to reduce friction for developers and challenge NVIDIA’s CUDA moat.
Across both days, Qualcomm made it clear that it wants to be the agent’s endpoint on every device a person carries, working alongside the cloud rather than against it. The open questions are which of these new form factors people will buy and which developers will prioritize. We’ll be watching.