Qualcomm CEO Says Smartphones Will Run 100B-Parameter AI Models by 2028

In early October, Qualcomm chief executive Cristiano Amon made a striking claim: leading artificial intelligence companies are already asking his chip company to prepare smartphones capable of running AI models with 100 billion parameters — continuously — by 2028.

Speaking in an October 8 interview, Amon said the request is coming from “some of the companies at the forefront of AI.” He declined to name them, but he was specific about what they want: “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 to keep that model running all day, not just for one-off tasks.

An AI model’s “parameters” are the values it learned during training that determine how it thinks and responds. More parameters generally mean more capable AI — better reasoning, better language understanding, better ability to handle complex requests. Today’s flagship phones can run only a fraction of that scale. For context, Gemini Nano-2, which powers AI features on select Android flagships, has around 3.25 billion parameters — a 100-billion-parameter model would be roughly thirty times larger.

Qualcomm’s newest flagship platforms, the Snapdragon 8 Elite Gen 6 and Snapdragon 8 Elite Extreme Gen 6, already show how fast the ceiling is rising. The Extreme chip can run mixture-of-experts (MoE) models above 30 billion parameters directly on the device, with a context window of roughly 32,000 tokens, plus tiny always-on sensing models of about 200 million parameters that run in the background.

The 100-billion target is therefore more than three times the scale Qualcomm advertises today. And the request is not simply to load such a model for a single question and shut it down. The AI companies want the model resident — always in memory, always reasoning, always available.

The “Battle of the Agents” Behind the Number

Why do AI companies want a giant model living in your phone instead of a data centre? Amon’s answer is about agents. Rather than today’s chatbot-style tools that wait for you to open an app and type a prompt, the next generation of AI is agentic: software that watches, reasons and acts on your behalf — checking your calendar, reading your messages, anticipating needs and taking actions across your installed apps.

Amon argued that the phone is the natural home for such agents because it holds information “uniquely tied to its owner”: personal context, app data and sensor information. Processing that locally instead of shipping it to the cloud could mean faster responses, better privacy and features that keep working when connectivity drops.

The Three Hard Problems: Memory, Battery and Heat

The engineering reality is daunting. Industry reporting points to three unresolved barriers:

1. Memory. Even after aggressive 4-bit compression, the model weights alone take up around 50 GB — before the OS, app context and computation. That pushes phones toward 64 GB of RAM. This lands on top of an existing DRAM supply crunch: AI demand has tightened memory supply and raised prices, and some phone makers have recently backed away from higher-RAM variants after Android flagships briefly reached 24 GB of LPDDR5X.

2. Battery. Keeping a 100B-scale model running continuously means sustained inference all day. “We now have to design chips to allow the phone to be working, even when you’re not operating it, and with an all day battery life and a lot of AI at the edge,” Amon said.

3. Heat. Dense, continuous computation generates heat, and phones have nowhere to put it. Thermal throttling — the chip slowing itself down to avoid overheating — would cap the real-world performance of such a model.

How Qualcomm Thinks the Gap Gets Bridged

Qualcomm is pointing to several engineering workarounds that could make 100B-scale on-device AI feasible without stuffing phones full of expensive DRAM:

• Streaming parameters from flash storage instead of holding everything in RAM
• Sparse mixture-of-experts designs that activate only the relevant parts of a model for each task
• Aggressive quantization — compressing weights down to fewer bits per parameter
• Distillation — training smaller models that mimic the behaviour of larger ones
• Expanding NPU shared memory and low-power memory designs

None of these has yet delivered a consumer-ready 100B-parameter phone. Qualcomm has not announced any smartphone chip capable of running such a model continuously.

What This Means for Pakistan

Pakistan is one of the world’s most Snapdragon-heavy markets. Qualcomm-powered phones from Xiaomi, OPPO, vivo, Redmi, realme and OnePlus dominate mid-range and flagship shelves across the country, so any leap in on-device AI capability will land here quickly — likely in the premium segment first.

Three implications stand out for Pakistani readers:

Prices at the top will climb. 100B-capable phones needing 64 GB-class memory and advanced packaging will arrive at flagship price points — beyond most local budgets. The mid-range, where most Pakistanis actually buy, will feel the benefits later through distilled, smaller models.

Privacy gets a boost where it matters. For Pakistani users, where concerns about personal data leaving the country are real and mobile connectivity is uneven, on-device processing is genuinely attractive: AI that understands your messages and documents without sending them to foreign data centres.

Expect the upgrade conversation to change. Amon’s “battle of the agents” framing means the next big reason to replace your phone may not be the camera or the screen — it will be whether your phone can run your personal AI agent locally. When shopping for your next device in 2027 and beyond, the NPU and RAM specs may matter more than megapixels.