Qualcomm announced two flagship smartphone processors, the Snapdragon 8 Elite Gen 6 and Snapdragon 8 Elite Extreme Gen 6, at its annual Snapdragon Summit in Hawaii. Both are built on a 2nm process and are designed around on-device AI, with the Extreme variant capable of running a 30-billion-parameter mixture-of-experts model entirely on the phone.
The chips carry new sensing hubs that can run small models of up to 200 million parameters continuously in the background. Qualcomm says that lets a phone run a personal scribe locally, tell speakers apart in a conversation, and build a memory of user habits to suggest task automations. The company also claims the new silicon can run a complete voice-in, voice-out agent without sending audio to a data center, which is the capability the whole announcement is built around.
The on-device bet
Qualcomm’s pitch is that the next wave of AI features will live on the phone rather than in the cloud. The argument has three parts: latency, privacy and cost. A voice agent that processes audio locally responds faster than one that round-trips to a server. Personal data that never leaves the device is easier to defend to regulators and users. And inference on the device costs the chipmaker’s customer nothing per query, while cloud inference is a recurring bill that has weighed on every AI feature launch so far.
The mixture-of-experts architecture is what makes the local model practical. A 30B-parameter MoE model activates only a fraction of its parameters for any given task, so the memory and compute footprint stays within what a phone battery can tolerate. For comparison, Apple’s most advanced on-device foundation model, shown at WWDC in June, was a 20-billion-parameter MoE. Qualcomm is claiming the lead in what fits inside a handset, at least on paper.
There is a catch the company did not emphasize. A 30B model running locally still needs memory bandwidth and thermal headroom that only the largest, most expensive phones have. Whether the feature reaches mid-range devices, where most of the market actually sits, will depend on future generations and on how aggressively phone makers cut memory to hit price points. The history of on-device AI so far has been flagship-first and mid-range-never.
Motorola signs on first
Motorola used the event to announce the Signature 27, a smartphone built on the Extreme chip, with general availability later this year. Samsung and other Android flagship makers are expected to adopt the Gen 6 family for their 2027 devices, which will put the chips in competition with Apple’s A-series and MediaTek’s Dimensity flagships.
The timing is tight for Xiaomi, which launched its 18 Pro tonight on the previous generation Snapdragon 8 Elite Gen 5, meaning the phone announced hours ago already sits one chip generation behind Qualcomm’s newest silicon. That pattern, launches shipping on last year’s flagship chip, has become normal as Qualcomm’s cadence has accelerated. Xiaomi’s launch leaned on other differentiators, a 2nm Qualcomm chip of its own naming, dual 200MP Leica cameras and 100W charging, rather than raw AI capability.
What the sensing hub changes
The sensing hub is the quieter but possibly more consequential piece. By dedicating a low-power block to always-on small models, Qualcomm is pushing toward phones that listen, watch and learn continuously without draining the battery. A hub running 200M-parameter models can handle wake-word detection, speaker identification and context sensing at milliwatt power levels, waking the main CPU only when a task needs it.
That architecture also changes the privacy conversation. Contextual awareness requires processing audio, location and usage patterns around the clock, and doing it on-device is the only version of that story most users and regulators will accept. Qualcomm’s design essentially assumes that agentic AI features will need local processing to pass review, and builds the hardware for it.
The practical payoff, according to Qualcomm’s demos, is an agent that knows enough context to be useful: it can draft a reply in your voice because it has seen your messages, schedule around your habits because it has learned them, and transcribe a meeting while distinguishing who said what. None of that requires a cloud round-trip in the new architecture, and all of it would have been impractical on last year’s silicon.
Market context
Qualcomm has said it is working on more than 40 AI devices beyond phones, including glasses, vehicles and robotics, but the company still makes most of its money from handsets. The Snapdragon Summit framing reflects a belief inside the company that phones remain where most people will meet AI agents, even as dedicated AI gadgets multiply.
The competition is not standing still. MediaTek launched a 3nm laptop chip with a 55 TOPS NPU this week and is pushing into custom AI server chips, a market where Qualcomm does not compete. Apple’s M5 and M6 silicon, now shipping in refreshed Macs, sets the on-device AI benchmark in laptops. In phones, though, Qualcomm’s 2nm lead and its claim of the largest local model give it a clear talking point for the 2027 flagship season.
Whether users care is the open question. AI agent features have shipped on phones for two years now, and consumer surveys keep showing modest enthusiasm. A chip that runs a 30B model locally is impressive engineering, but the products that make people actually use an agent on their phone every day have not arrived yet. Qualcomm is betting the hardware will be ready when they do, and the 2027 flagship season will show whether that bet lands.