Until now, using AI always meant the same thing: you type something, it travels to a server in another country, and an answer comes back. Every word you paste into the model leaves your office.
On August 25, Perplexity and Nvidia launched Portable Computer, an agent that flips that logic. The model, the conversation memory and the engine that carries out tasks run on the user’s computer. Local is the norm; the cloud is the exception, and the system asks permission before sending any step outside.
It runs open models with 27 billion parameters. They’re not the most powerful in the world, but according to the tests the company published, they match or beat alternatives on real office work tasks.
Why this matters to you even if you can’t buy it
And you can’t buy it. The hardware it needs costs about $4,700, it only runs on Linux for now, and on top of that it requires a paid subscription. For a Chilean small business that’s out of the question.
What matters is the direction.
Two years ago, a model capable of doing useful work needed a data center. Today it fits in a desktop device. Next year it will fit in a decent laptop. That curve isn’t slowing down.
And there’s a second, more urgent reason: on December 1, Law 21.719 takes effect, and it regulates what you do with your partners’ personal data. Every time you paste a customer database into a chat so it can draft something for you, that data has left your control. A model that runs on your machine makes that question disappear, because nothing leaves.
It’s no coincidence that Perplexity is aiming the product at law firms, healthcare and finance. Those are the industries where sending data out is expensive.
What you can do today
Separate the tasks. Not everything you do with AI involves sensitive data. Drafting an Instagram post isn’t the same as analyzing your customer portfolio. Treat those two things differently, with different tools.
Find out what happens to what you write. Serious tools state whether they use your conversations for training and whether you can turn that off. If a tool doesn’t say anywhere, that silence is the answer.
Anonymize before pasting. If you need a model to analyze sales, change the names. “Customer 1, Customer 2” works just as well for the analysis and takes the legal problem off the table.
Don’t buy hardware yet. The device that costs $4,700 today will cost a fraction of that in two years, and it will run better models. Buying now means paying the price of being first without getting the advantage.
The bigger picture
We’re an AI-based lab. The cloud serves us well, we use it every day, and it will remain the default for a good while.
But the conversation about AI has been held almost entirely in terms of which model is smarter. Almost nobody is asking where the computation happens, and that question is going to matter more than the other one once the law starts to bite.
It’s worth knowing, for every tool you use, where what you write ends up. Most small businesses we know don’t know. Not out of carelessness: because nobody told them.
This note is part of a series on what AI regulation means for a small company in Chile. The previous one: The law that’s coming for you isn’t the one you’re watching.