Nvidia is going after a $200 billion CPU market with a new category it’s calling the “AI agent PC,” and the launch partners are exactly who you’d guess: Microsoft, Dell, HP. The pitch is that agents need a dedicated machine, with the neural engines and memory bandwidth and on-device inference to run them locally instead of pinging a datacenter. I read the coverage and my first thought was that I already have an AI agent. It runs in Chrome. It cost me nothing.

What is this machine actually for?

The agent PC story assumes the bottleneck is compute. That if your laptop had a beefier NPU, the agent would finally be able to do the work. And for a narrow class of things, sure, local inference matters. Running a model offline, keeping the prompt off someone’s servers, getting tokens out fast without a network round trip.

But that’s not where my workday gets stuck. My workday gets stuck inside authenticated web sessions. The CRM I’m logged into. The Gmail tab with the thread I need to reply to. The vendor portal that has no API and a form with thirty fields. None of that is a compute problem. A faster chip does not give an agent permission to act inside the session I already opened. It just makes the wrong thing faster.

The session is the moat, not the silicon

Here’s the part the hardware story skips. When you log into your bank, your Salesforce, your internal admin panel, the thing that proves you’re you isn’t your CPU. It’s a cookie. A session token sitting in the browser, established after you cleared a login wall, maybe an SSO redirect, maybe a 2FA prompt on your phone. That state is fragile, scoped to one browser profile, and impossible to hand to a cloud agent without re-doing the whole authentication dance somewhere that doesn’t have your phone.

A brand-new agent PC inherits none of that. Out of the box it’s a powerful machine that is logged into nothing. To make it useful you’d have to re-authenticate every account on it, and then trust whatever agent runtime ships with the thing to hold those credentials. So you’ve bought silicon to solve a permissions problem, which is a bit like buying a faster car to fix a locked garage.

The agent that wins is the one standing inside the session that already exists. That’s the whole argument, and no amount of TOPS changes it.

A small example, because the abstract version is boring

Last week I needed to pull line items out of a supplier dashboard that, naturally, offers CSV export only on the plan above the one we pay for. The data was right there on the screen, rendered, mine to read. An agent running on a separate machine would’ve hit a login wall and stopped. Dassi, sitting in my browser side panel, just read the rendered page, the same page my eyes were looking at, and handed me the rows. No export button required, no new hardware, no token I had to generate in some developer console and paste into a settings field.

That is the texture of most real agent work. It is not “train a model.” It is “do the annoying thing on the page I’m already on.”

Local execution is the right instinct, wrong layer

I don’t want to be unfair to the agent PC people, because the underlying instinct is correct. Agents should run close to you. Your data should not take a tour of three datacenters before something useful happens. We’ve written before about how desktop AI quietly went local while browser agents were still renting servers, and the direction of travel there is genuinely good.

But Nvidia and Microsoft are putting “local” at the silicon layer when the layer that matters for most knowledge work is the browser. The browser is already local. It already holds your login state, your tabs, your half-finished forms, the precise context an agent needs and a cloud sandbox can never reach. You don’t need a new neural processing unit to capture that. You need an agent that lives where the context already is.

Dassi runs in the Chrome side panel and reads the page you’re looking at, inside the session you already authenticated. Bring your own model key, or log in with the ChatGPT subscription you already pay for. The model can be GPT, Claude, Gemini, whatever you like, because the model was never the hard part. The plumbing was. You can add it from the Chrome Web Store and skip the $2,000 hardware refresh.

So who buys the agent PC?

Some people, genuinely. If you’re running local models for privacy, doing heavy on-device inference, building things, the extra silicon earns its keep. I’m not going to pretend that market is fake.

What I’m skeptical of is the framing that this is where agents become useful for normal work. Because the agent most people need isn’t waiting on a faster chip. It’s been sitting in the tab the whole time, looking at the same authenticated page you are, and we keep forgetting it can already see that you’re logged in. Maybe the $200 billion is better spent on figuring that out. Or maybe just on a side panel extension. One of those is free.