Qualcomm CEO says AI labs want 100B-parameter phones by 2028. He named none

Amon, Qualcomm's president and chief executive, says frontier AI companies have told him they need a phone that can run a model of at least 100 billion parameters by 2028, kept resident continuously rather than loaded for a single task. He said so in a Sources podcast interview recorded around 8 October. No company was named, and no Qualcomm product carries that date.

What Qualcomm ships today

Qualcomm unveiled the Snapdragon 8 Elite Gen 6 and Snapdragon 8 Elite Extreme Gen 6 at its Snapdragon Summit on 22 September, both on TSMC's 2-nanometre process. The Extreme part's Hexagon NPU adds a new Element Accelerator and a larger shared memory pool.

• On-device ceiling today: mixture-of-experts models above 30 billion parameters
• Context held there: roughly 32,000 tokens

The 100-billion-parameter target is more than three times that ceiling, and the labs asking for it want the model resident rather than spun up. Qualcomm's route there is streaming parameters from flash instead of holding the model in RAM, plus a bigger NPU-addressable memory pool. Its always-on hubs already run 200-million-parameter models continuously; the new chips process about a million tokens a day.

Why the battery question stays open

None of that answers the question a phone poses. Amon describes a target the industry is pushing toward, not a commitment Qualcomm has made. His verbatim words, via Stocktwits: "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." Qualcomm has not said which labs are asking, and several are reportedly building their own phones, which reads as much like a warning as an order. A permanently resident 100B model is a power and thermal problem before it is a memory problem, and the memory problem is not small: at four bits per parameter the weights alone need roughly 50 GB before any KV cache, on a phone that shares its memory with the operating system.

What would make 2028 real

Amon argues agentic AI belongs on the phone because the user's personal context, a memory graph and sensor data, exists only there. He calls cloud-versus-edge a false dichotomy, pointing to Nvidia's roughly 40/40 split between hyperscaler revenue and on-premise, neocloud and sovereign buyers. Memory prices five to six times higher have cut the smartphone market by about 20% this year, even as roughly 1.2 billion phones sell annually, a dynamic we flagged when memory supply tightened into 2027. The nearest named customer is an estimate: analyst Ming-Chi Kuo reports OpenAI worked on an AI-agent phone program with MediaTek and Qualcomm, possible production in early 2027, shipments near 30 million units across 2027-2028. What would settle the 2028 date is a named customer, a shipping part, and a runtime budget measured on a battery.