Qualcomm’s chief executive just put a number and a deadline on something the AI industry has been dancing around for two years: when a phone in your pocket will run a genuinely large language model on its own, all day, without touching a server. Speaking in an interview that circulated widely this week, Cristiano Amon said unnamed AI companies are already asking Qualcomm to make sure smartphones can run a 100 billion-parameter model locally by 2028, and to keep that model running continuously rather than waking it up for single tasks. The remark landed fast across tech and finance outlets, and it reframes a debate that, until now, mostly centered on chatbots and cloud inference rather than silicon sitting inside a phone case.
The comment arrives three weeks after Qualcomm shipped its newest flagship chips, and it gives those launches a longer runway than most coverage assumed. It also raises an uncomfortable question for every phone maker currently shipping 12GB or 16GB of RAM: what happens when “running a model” means hosting tens of gigabytes of weights, continuously, on a battery-powered device that also has to make calls and survive a full day unplugged?
What Qualcomm’s CEO Actually Said
According to the interview excerpts circulating this week, Amon described direct requests from AI companies rather than a Qualcomm product roadmap. “Some AI companies right now, some of the companies at the forefront of AI are telling Qualcomm, by 2028 I need to make sure that I have the ability to run at least a 100 billion parameter model into a phone, and I need to have this thing running kind of all the time,” Amon said, according to a report circulated by Stocktwits. He added that he could not name the companies making the request.
That distinction matters. Amon isn’t announcing a chip, a phone, or even a Qualcomm commitment. He’s describing a demand signal from customers who build frontier AI models and now want their software to live on a phone instead of (or alongside) a data center. In the same conversation he said plainly, “AI companies are starting to do phones,” a line that reads less like speculation and more like an executive relaying what his sales team hears in meetings. The framing echoes something Qualcomm has said in different words for a year: on-device AI improves privacy and cuts latency because data never leaves the handset.
The Unnamed AI Companies Behind the Request
Amon’s refusal to name the companies asking for 100-billion-parameter phones is the detail most likely to generate speculation over the next few months. Every major AI lab has some reason to want a foothold on phones: distribution, data, and a buffer against paying cloud-compute bills for every single query a user sends. None of the labs most commonly guessed at in industry chatter has confirmed a smartphone hardware partnership tied to this specific 2028 target, and nothing in the public record ties a named company to the request.
What is verifiable is the broader pattern: on-device AI vendors have been pushing phone-makers toward “agentic” software, meaning a model that can act across apps instead of answering one prompt at a time. Amon tied the two ideas together directly, saying there’s an understanding that the phone matters for agents and that “there’s going to be the battle of the agents.” Read against Qualcomm’s own September product messaging, that’s a company trying to make sure its chips are the ones running whichever agent wins that fight, regardless of which lab built it.
Why 100 Billion Parameters Is the Number That Matters
Parameter count isn’t a vanity metric. It roughly tracks how much a model can reason about, remember, and generalize across tasks without retraining. Cloud-hosted frontier models today routinely run past the hundred-billion mark, sometimes well past it when including mixture-of-experts (MoE) architectures that activate only a slice of their total weights per query. That MoE trick is exactly what makes a 100-billion-parameter phone model plausible at all: instead of loading every parameter into active memory for every token, the chip activates a smaller expert subset, cutting the real-time compute and memory load dramatically.
Qualcomm’s current flagship silicon already leans on this approach. The Snapdragon 8 Elite Extreme Gen 6 can run MoE models exceeding 30 billion total parameters locally, a figure the company confirmed when it unveiled the chip. Getting from “30 billion, task-triggered” to “100 billion, running continuously” isn’t a single generational leap. It’s roughly tripling the model size while also changing the operating model from on-demand to always-on, which multiplies power draw and thermal output rather than simply adding to it.
Inside Qualcomm’s New Snapdragon 8 Elite Lineup
Qualcomm announced the Snapdragon 8 Elite Gen 6 and the Snapdragon 8 Elite Extreme Gen 6 on September 22, 2026, at its Snapdragon Summit. Both platforms are built on a 2-nanometer manufacturing process and are explicitly positioned around agentic AI rather than the simpler voice-assistant tasks that defined on-device AI a few years ago. We covered the launch in detail when the Snapdragon 8 Elite Gen 6 debuted on its 2nm process, and the Extreme Gen 6’s headline capability, the 30-billion-plus parameter MoE support, is the current ceiling Qualcomm has publicly committed to shipping.
Qualcomm says the chip includes Adreno matrix cores built specifically to accelerate local model execution, and it’s lining up a wide partner base: HONOR, iQOO, Motorola, OnePlus, OPPO, REDMI, RedMagic, vivo, and Xiaomi have all been named as flagship partners. Motorola already confirmed its upcoming Motorola Signature 27 will use the Extreme Gen 6, with general availability expected later in 2026, well ahead of the broader wave of 2027 flagships that CNET reported will carry the same silicon. iQOO’s early benchmark numbers for a Snapdragon 8 Elite Extreme Gen 6 device, which we tracked when the iQOO 16 Global surfaced on Geekbench, give an early read on how the chip performs before review units land broadly.
Put simply: the hardware Qualcomm is shipping right now tops out at roughly a third of the parameter count Amon says AI companies are requesting for 2028. That gap is the real story here, not the 2028 date itself.
The Memory and Battery Math Nobody Has Solved
Running a model “all the time,” as Amon put it, is a fundamentally different engineering problem than running one occasionally. A continuously resident 100-billion-parameter model, even quantized and sparsely activated, demands memory capacity and bandwidth that current flagship phones don’t carry, plus a power budget that has to coexist with everything else a phone does in a day. DRAM pricing is already working against phone-makers here: memory now costs more relative to leading-edge logic than it did a year ago, a dynamic we examined when DRAM pricing outpaced TSMC’s 2nm silicon costs. Stacking more RAM into a phone to hold a persistent large model isn’t a free decision; it’s a bill-of-materials fight that OEMs have been losing ground on for over a year.
Battery chemistry hasn’t moved at the same pace as silicon efficiency gains, either. Qualcomm’s answer so far has been architectural efficiency rather than brute force, which is why MoE sparsity and dedicated matrix cores matter more to this roadmap than raw transistor count. Whether that’s enough to get from occasional 30-billion-parameter inference to continuous 100-billion-parameter inference within roughly two years is an open engineering question that Amon’s comments don’t answer. He described a target AI companies are asking for, not a solution Qualcomm has already built.
MediaTek’s Dimensity 9600 Pro Is Chasing the Same Target
Qualcomm isn’t alone in this race, and the timing of competing announcements suggests the whole flagship chip industry sees the same demand curve. MediaTek launched its Dimensity 9600 Pro on September 15, 2026, a week before Qualcomm’s event, and positioned it squarely around on-device agentic AI. We broke down the chip’s power profile in our report on the Dimensity 9600 Pro’s 2nm launch and power cuts, and the headline specs track closely with Qualcomm’s own direction: support for on-device MoE models up to 30 billion parameters, paired with a second-generation NPU that MediaTek says cuts power consumption for always-on AI tasks by 40%.
MediaTek also claims its NPU 1090 delivers 51% faster LLM prefill performance and 55% more token generation per watt compared with its prior generation. Those efficiency gains matter more than raw model size for the “always-on” piece of Amon’s comments, since a model that’s resident but idle most of the time still has to sip power rather than drain a battery in hours. The fact that MediaTek shipped an always-on efficiency story in the same month Qualcomm shipped an MoE-scale story suggests the entire chip industry is converging on the same two-part problem: get the model bigger, and get it to idle cheaply.
Where Apple and Samsung Stand
Apple’s newest silicon, the A20 Pro, is also built on a 2-nanometer process, putting all three of the industry’s major mobile chip designers on the same leading-edge node within weeks of each other this fall. What’s missing from the public record is a clear Apple statement on local model-parameter ceilings or a commitment to continuous on-device execution comparable to what Amon described. Apple has historically been quieter about raw parameter counts than Qualcomm or MediaTek, preferring to talk about feature-level capabilities rather than model size, so the absence of a comparable claim doesn’t necessarily mean Apple is behind. It means Apple hasn’t chosen to compete on this specific metric publicly.
Samsung’s position is tied more directly to Qualcomm’s roadmap than Apple’s. Reports point to Qualcomm’s new Snapdragon silicon appearing in Samsung’s Galaxy S27 series, expected in early 2027, which would put Samsung’s flagship line a full generation behind the 2028 target Amon mentioned, at least on paper. There’s no confirmed Samsung-specific AI chip announcement from the past one to three months that changes that picture, so any comparison beyond “Samsung will likely use Qualcomm’s silicon on a similar timeline to other Android OEMs” would be getting ahead of what’s actually been confirmed.
On-Device Model Size: Qualcomm’s Climb From 200 Million to 100 Billion
Lined up chronologically, Qualcomm’s own disclosures show a parameter curve that’s been climbing fast even before Amon’s 2028 comments. The table below tracks publicly confirmed figures against the still-unconfirmed 2028 target.
| Generation / Product | Max Local Parameters | Operating Mode | Status |
|---|---|---|---|
| Qualcomm Sensing Hub (always-listening tier) | Up to 200 million | Low-power, continuous listening | Shipping today |
| Snapdragon 8 Elite Extreme Gen 6 | 30 billion+ (MoE, sparse activation) | Task-triggered, agentic | Announced Sept. 22, 2026 |
| MediaTek Dimensity 9600 Pro | Up to 30 billion (MoE) | Task-triggered, efficient always-on NPU | Announced Sept. 15, 2026 |
| Reported 2028 AI-company request | 100 billion+ | Continuous, always running | Requested target, no confirmed product |
The jump from a 200-million-parameter sensing model to a 30-billion-parameter MoE model already represents a roughly 150x increase in local capability inside two to three product cycles. Getting to 100 billion, running continuously rather than on-demand, is a smaller multiple on paper but a much harder systems problem, because “continuous” changes the power and thermal equation in ways that parameter count alone doesn’t capture.
Flagship AI Chip Comparison, Fall 2026
Here’s how the three major mobile chip designers compare on what’s actually been confirmed publicly as of this month. Cells marked “not disclosed” reflect gaps in the public record, not assumptions.
| Chip | Maker | Process Node | Announced | Max Local Model (disclosed) | Key AI Feature |
|---|---|---|---|---|---|
| Snapdragon 8 Elite Extreme Gen 6 | Qualcomm | 2nm | Sept. 22, 2026 | 30 billion+ params (MoE) | Adreno matrix cores, agentic AI |
| Snapdragon 8 Elite Gen 6 | Qualcomm | 2nm | Sept. 22, 2026 | Not separately disclosed | Agentic AI, standard flagship tier |
| Dimensity 9600 Pro | MediaTek | 2nm | Sept. 15, 2026 | Up to 30 billion params (MoE) | NPU 1090, 40% lower always-on power draw |
| A20 Pro | Apple | 2nm | Fall 2026 | Not disclosed | Not disclosed publicly for on-device LLM ceiling |
What stands out is how tightly bunched the three companies are on process node, even as they diverge sharply on how openly they talk about model-size ceilings. Qualcomm and MediaTek are both happy to put a parameter number in a press release. Apple, so far, isn’t.
From Siri to Agents: How We Got Here
It’s worth remembering how recently “on-device AI” meant something much smaller than what’s being discussed now. For most of the smartphone era, local AI meant wake-word detection, basic photo tagging, or predictive text, tasks that fit comfortably inside tens of millions of parameters. The generative AI boom that started in the early 2020s pushed phone-makers to chase cloud-scale capability on local hardware, but the first few years of that push were mostly marketing: a phone running a “local AI feature” usually meant a small distilled model handling one narrow job.
What’s changed over the past year is the shift toward agentic AI, meaning software that can take multi-step actions across apps rather than answering isolated prompts. Qualcomm’s own September announcement framed both new Snapdragon chips around exactly that shift, describing AI agents that operate in the background, interact with multiple apps, and anticipate user needs. Nvidia’s open-weight model push, which we covered when Nvidia’s Nemotron 3.5 shipped at 30 billion parameters while eyeing a trillion-parameter ceiling, shows the same scaling instinct playing out on the cloud side of the industry at the same time phone chips are scaling locally. Amon’s 2028 comments are the first time a chip CEO has put a specific, large parameter number and a continuous-operation requirement into the same sentence about phones specifically.
What a “Battle of the Agents” Means for the Smartphone Market
If AI companies are genuinely asking chipmakers to support always-running, 100-billion-parameter models by 2028, the smartphone market’s competitive axis shifts. Camera megapixels and screen refresh rates have driven flagship marketing for a decade. A phone that can host a persistent large model without draining its battery by lunchtime would be a different kind of flagship pitch entirely, one closer to a subscription service running on hardware than a one-time purchase.
That also changes who phone-makers need to negotiate with. If AI labs are the ones pushing chipmakers toward this target, rather than the chipmakers pushing it on their own, phone OEMs could find themselves negotiating software deals with AI companies as a precondition for which silicon they buy, not the other way around. Amon’s phrase, “AI companies are starting to do phones,” hints at exactly that kind of leverage shift, where model providers start dictating hardware requirements instead of simply optimizing software to fit whatever hardware ships.
Industry Reaction
Coverage of Amon’s comments spread quickly across financial and tech trade outlets this week, with most framing it as a notable but unverified forward claim rather than a settled roadmap. Analysis from Wccftech pushed back on the practicality of the target, arguing that streaming model weights through flash memory or leaning harder on sparse MoE architectures are the only realistic paths to running a 100-billion-parameter model on phone-class hardware by 2028, given current DRAM supply and cost constraints. That skepticism lines up with the memory-cost pressures already squeezing phone bills of materials industry-wide.
Other outlets framed the comments as part of a bigger story about AI companies moving into hardware generally. A write-up from Sources described the moment as an early signal of a coming “AI phone supercycle,” where upgrade cycles get driven by model capability rather than camera or display improvements. Coverage from TokenPost and Startup Fortune both repeated Amon’s “battle of the agents” framing as the clearest summary of why phone hardware suddenly matters to AI labs that have, until now, mostly cared about data-center compute.
Amon’s own closing line in the interview, “There’s more coming,” suggests he expects this isn’t the last chip-industry statement on the topic this year. Given how quickly MediaTek and Qualcomm both shipped 2nm, agentic-focused chips within a week of each other in September, that’s a reasonable expectation.
Five Predictions for the Road to 2028
- The 30-billion mark becomes the new flagship floor within a year. Expect Qualcomm, MediaTek, and likely Apple to treat 30 billion-parameter local MoE support as a baseline flagship feature by late 2027, not a differentiator.
- Memory, not compute, decides who gets to 100 billion first. DRAM capacity and pricing will matter more than NPU performance in determining which phone-makers can realistically ship a 2028-class device, given current memory cost pressure on handset bills of materials.
- At least one unnamed AI company gets identified before 2028. Given how specific Amon’s description was, expect a leak, partnership filing, or chip-supply report to surface naming one of the AI companies behind the request well before any product ships.
- Battery marketing shifts toward “agent uptime.” If continuous model operation becomes a real feature, expect phone-makers to start marketing battery life in terms of hours of active agent operation rather than screen-on-time, similar to how laptop makers already market “AI PC” battery claims.
- Samsung follows rather than leads. Given its reported Snapdragon dependency for the Galaxy S27 series, expect Samsung to confirm any 100-billion-parameter local AI capability at a Galaxy Unpacked event only after Qualcomm has already shipped the underlying silicon, not ahead of it.
Frequently Asked Questions
Did Qualcomm announce a phone that runs 100 billion-parameter models?
No. Cristiano Amon described requests from AI companies for that capability by 2028. Qualcomm has not announced a chip or product confirmed to run 100-billion-parameter models continuously.
Which AI companies asked Qualcomm for this capability?
Amon said he could not disclose the names. No company has publicly confirmed making this specific request.
What’s the biggest model a phone can run locally today?
Qualcomm’s Snapdragon 8 Elite Extreme Gen 6, announced September 22, 2026, supports mixture-of-experts models exceeding 30 billion total parameters locally. MediaTek’s Dimensity 9600 Pro, announced a week earlier, supports on-device models up to 30 billion parameters as well.
What does “running continuously” mean for an AI model on a phone?
It means the model stays loaded and active in the background rather than being loaded only when a user triggers a specific task, similar to how a background app stays resident in memory. That requires sustained power and memory bandwidth well beyond what task-triggered AI features need.
Is memory the main obstacle to 100-billion-parameter phones?
It’s one of the two most cited obstacles, alongside thermal and battery limits. DRAM has become more expensive relative to leading-edge chip logic, and continuously hosting a large model requires more memory capacity and bandwidth than current flagship phones carry.
How does this compare to Apple’s approach?
Apple’s newest chip, the A20 Pro, also uses a 2-nanometer process, but Apple has not publicly disclosed a comparable local model-parameter ceiling or committed to continuous on-device execution at the scale Amon described.
Will Samsung’s Galaxy S27 support this capability?
Reports indicate the Galaxy S27 series, expected in early 2027, will use Qualcomm’s new Snapdragon silicon, which currently tops out at 30 billion-plus parameters, not 100 billion. No Samsung-specific confirmation beyond that has been reported.
Is the 2028 date a confirmed Qualcomm product launch?
No. It’s a capability target that AI companies reportedly requested, as relayed by Amon. It is not an announced Qualcomm product, launch date, or delivery commitment.
Related
- Snapdragon 8 Elite Gen 6 Hits 5.11GHz on 2nm
- Qualcomm’s 30B AI Chip Challenges Apple A20 Pro
- MediaTek Dimensity 9600 Pro Hits 2nm, Cuts Power 61%
- iQOO 16 Global Spotted on Geekbench With 10,029 Score
- DRAM Now Costs 54% More Than TSMC 2nm Silicon
