I read the energy-economy modeling paper going around this week, the one that feeds AI productivity gains into a global model and comes out the far side with a net increase in CO₂ rather than the decrease everyone assumes, and the first image in my head wasn’t a data center in Virginia. It was a browser window nobody asked for.

The mechanism is old. Jevons published it in 1865 about coal: make a resource cheaper to burn and people burn enough more of it that total consumption climbs. The AI version says the hours you save get spent on more work, and that work runs on hardware someone has to power, cool, and depreciate.

Economy-wide arguments are slippery, though. You can model them in either direction depending on what you assume about elasticity, and I’m not qualified to referee that fight. So take a version small enough to watch with your own eyes.

Your tab already rendered the page

Every cloud browser agent works the same way underneath. You hand it a task, it allocates a container, it launches a headless Chromium inside that container, and that Chromium wakes up as nobody. No cookies. No session. No extensions, no history, no saved cards, no idea who you are.

Then it navigates to a page your own Chrome, eighteen inches from your face, has open in tab four. Already rendered. Already authenticated. Already past the Cloudflare interstitial it cleared twenty minutes ago.

So the remote machine does all of it again. DNS, TLS handshake, HTML, the 1.8 MB of JavaScript that a modern SaaS app ships before it draws a single pixel, the webfonts, the avatars. And then it hits a login wall, because it always hits a login wall, and now you’re either pasting credentials into a vendor’s secret vault or clicking an OAuth consent screen so a stranger’s container can become a worse copy of the session sitting on your desk.

The number I couldn’t find

I went hunting for a defensible per-task energy figure and came back irritated, because the honest answer is that nobody publishes one and the ones floating around are marketing. What I can say from watching these things run: a cold container start is a second or two of CPU, a headless Chromium holds 300 to 500 MB resident for the life of the task, a heavy app page is a few hundred milliseconds of layout and script execution, and the bytes crossing the network have their own cost that nobody attributes to anybody.

Per task it rounds to zero. That’s exactly the trap. Per task, everything rounds to zero, which is how you end up with a industry-wide accounting problem and no line item to point at.

Multiply instead. One person running forty agent tasks a day is forty cold Chrome boots, forty duplicate renders of pages that were already rendered locally, forty login replays. Scale that to a team, then to the fleet of autoscaling browser-agent startups that seem to launch weekly, and the duplicated work stops being a rounding error. It becomes the product. The compute isn’t doing anything your laptop wasn’t already doing, and your laptop was doing it anyway, with the fans off, because rendering a page you’re looking at is the browser’s entire job.

That’s the part that bugs me. Not the emissions math, which I hold loosely. The waste.

Local execution skips the duplicate machine

Same model. Same task. No second Chrome.

What actually changes

A browser agent that runs inside your Chrome doesn’t provision anything. dassi lives in the side panel and drives the tab that’s already open, with the session you already established, in the profile that already has your extensions and your ad blocker and your damn cookie preferences.

The only thing that leaves your machine is the prompt and the page context going to whichever model you picked. That request happens either way. A cloud agent sends the same inference call, plus it rents a browser to generate the context. The browser is the part you’re paying for twice, and with BYOK you can at least see the one bill that remains.

I’ve written before about how cloud browser agents can’t see your tabs, which I framed as a capability problem. The energy framing is the same fact wearing different clothes. A remote Chrome is blind and redundant, and those are two symptoms of one design choice: putting the browser somewhere other than where the human is.

Or you could look at it from the other end. Desktop AI went local over the past year, quietly, mostly for latency and privacy rather than carbon. Browser agents are the holdout, still shipping tasks to a datacenter so a machine with no memory of you can log in as you.

The rebound effect says efficiency gains get eaten by increased use. Fine, probably true. But some of the compute isn’t serving increased use at all. It’s serving a copy of a machine already running on your desk, and you’re the one who left it on.