Why does a local AI voice agent running on a super cheap SoC matter?

Most recent news about AI seems to involve staggering amounts of money. OpenAI and Nvidia sign a $100b data center contract. Meta offers researchers $100m salaries. VCs invested almost $200b in AI startups in the first half of 2025.

Frankly, I think we’re in a massive bubble that dwarfs the dot-com boom, and we’ll look back on these as crazy decisions. One of the reasons I believe this is because I’ve seen how much is possible running AI locally, with no internet connection, on low-cost hardware. The video above is one of my favourite recent examples. It comes from a commercial contract we received to help add a voice assistant to appliances. The idea is that when a consumer runs into a problem with their dishwasher, they can press a help button and talk to get answers to common questions.

What I’m most proud of here is that this is cutting-edge AI actually helping out with a common issue that many of us run into in our daily lives. This isn’t speculative, it’s real and running, and it doesn’t pose a lot of the ethical dilemmas other AI applications face. Here’s why I think this matters:

• The consumer doesn’t have to do anything beyond pressing a button to use it. There’s no phone app to download, no new account to create, and no Wifi to set up. The solution works as soon as they plug the appliance in. This is important because less than half of all smart appliances ever get connected to the internet.
• It’s using Moonshine and an LLM to do a much better job of understanding natural speech than traditional voice assistants. The questions I asked in the demo were off-the-cuff, I deliberately used vague and informal language, and it still understood me.
• It addresses a genuine problem that manufacturers are already paying money to solve. They are currently spending a lot on call centers and truck rolls to help consumers. This solution has the potential to reduce those costs, and increase consumer satisfaction, by offering quick answers in an easy way.
• Running locally means that audio recordings never have to go to the cloud, increasing privacy.
• Local also means fast. The response times in the video are real, this is running on actual hardware.
• This doesn’t require a GPU or expensive hardware. It runs on a Synaptics chip that has just launched, and will be available in bulk for low-single-digit dollars. This means it can be added to mass-market equipment like appliances, and even toys. Since it’s also able to run all the regular appliance control functions,  it can replace similarly-priced existing SoCs in those products without raising the price.
• More functionality, like voice-driven controls, can easily be added incrementally through software changes. This can be a gateway to much richer voice interactions, all running locally and privately.

All these properties give local AI a much better chance to change our daily lives in the long term, compared to a chat bot that you access through a text box on a web page. AI belongs out in the world, not in a data center! If you agree, I’d love to hear from you.