Meta has built a new piece of hardware for Agents—Muse Charm, an AI smart key fob. Just two weeks earlier, Muse hit over 2.5 million downloads. Meta has models, a social ecosystem, and users, yet it still gives Agents a dedicated hardware platform—effectively pushing the question back to hardware startups: what AI scenarios are actually worth building a separate hardware device for?

The answer is becoming clearer. On the downstream side, intelligent phone-core modules from Huaqiangbei have dropped to dozens of yuan; in Dongguan and Foshan, factories can drive parts of the AI hardware cost down to within a hundred yuan. Upstream, Baidu is embedding large models into Xiaodu, Tencent is opening Agents to AI glasses, earbuds, and recording devices, and Alibaba and ByteDance are also extending into terminals. As both the supply chain and AI capabilities get cheaper, startups in the middle have more and more room to leverage resources.

The real test of who’s ahead is this: “product as the first principle.” Are there real, high-frequency scenarios where it’s impossible without dedicated independent hardware? Can AI capabilities be turned into a great user experience? Will users still be using it after 30 days, or even six months? Competencies and models can be borrowed, but the authority to define the product, the relationship with users, and the long-term moat—these must grow from the hardware company itself.

When all products revert to first principles, what’s left for your hardware—this is the question you should answer first in this wave of AI hardware hype.