ANALYSIS Two developments this week point in the same direction: the smartphone is being repositioned as a persistent AI inference engine, not merely a thin client for cloud models. Qualcomm's CEO says AI companies are designing handsets around always-on large language models, while a startup shipping ambient-listening software on Apple devices shows what those models would actually do.
Why it matters
The phone has been a passive relay for cloud AI. If chipmakers and app developers converge on continuous, on-device inference, the economics of AI deployment shift from data-center scale to silicon-per-handset scale, and the privacy calculus changes with it.
The big picture
Qualcomm CEO Cristiano Amon said AI companies are "starting to do phones" and are asking Qualcomm to support running 100-billion-parameter local LLMs continuously by 20283. The phrasing is notable: not "using phones" but "doing phones," implying that AI-native companies are moving toward designing or specifying handset hardware around inference workloads rather than adapting models to existing phone constraints. Amon discussed how AI is changing smartphones and wearables in the same remarks.
Separately, Mistlabs Ltd. brought its Ambi AI system out of beta on Apple devices, backed by $26 million in funding led by Luminous Ventures and ODM Huaqin1,2. Ambi is releasing apps for iPhone, Mac, and Apple Watch that act as meeting note-takers but can also be set to listen continuously, capturing context throughout the day. CEO Peter Mo said most AI platforms are built to complete tasks, and Mistlabs wants to see what becomes possible once software can "understand, reason and act in real time". Ambi's actions "can sometimes come before the user even thinks to ask," Mo said.
Between the lines
The Qualcomm and Mistlabs stories describe two halves of the same architecture. Mistlabs demonstrates the application layer: agents that fact-check claims during a live meeting, compare rival products, and draft proposals on a user's MacBook before the meeting ends. Those agents work from everything Ambi has picked up about a person, and the company says the software's understanding of how that person likes to work sharpens over time. None of it is locked to a single device; whatever Ambi is working on follows the user from Mac to phone to watch.
ANALYSIS That kind of persistent, cross-device ambient reasoning is exactly the workload Amon describes AI companies asking Qualcomm to enable at the silicon level by 2028. A 100-billion-parameter model running continuously on-device would supply the local inference backbone such applications need without round-tripping every utterance to the cloud.
The privacy tension is explicit even in the coverage sympathetic to the product. "There's obviously privacy and data security concerns with anything that's collecting data about you continuously," TechCrunch noted, adding that startups are building such apps "undeterred anyway, arming themselves with the belief that the market for personalized assistance is big enough to swallow such concerns". ◆ On-device inference could, in principle, ease that tension: if the model runs locally, raw audio need not leave the handset. Ambi's cross-device sync, where context follows the user from Mac to phone to watch, raises the question of how much processing stays local versus how much travels between devices today, a design choice the company has not detailed.
Mistlabs' funding sources are worth noting. Huaqin, an original design manufacturer, co-led the round alongside Luminous Ventures. ◆ An ODM investing in an ambient-AI software company suggests hardware partners see value in coupling always-on inference software with device design, a pattern consistent with Amon's observation that AI companies are now "starting to do phones".
What's next
Amon's 2028 target for continuous 100-billion-parameter on-device models sets a concrete hardware milestone. Mistlabs, meanwhile, is shipping today on existing Apple silicon with a smaller footprint, using its Routine feature to let users hand recurring jobs to multiple agents at once. ◆ The gap between what ships now on current hardware and what Qualcomm says AI companies are requesting for 2028 defines the near-term competitive space: startups building ambient AI must architect for cloud-assisted inference today while positioning for a local-first future. The $26 million Mistlabs raised will fund that bridge.