Nscale is reportedly hunting $3.5 billion in pre-IPO financing, which is the kind of number that stops registering as money and starts registering as weather. GPUs, power contracts, cooling, land, substations. Somebody pays for all of it eventually. And I keep circling back to the smallest, least glamorous slice of that buildout: the headless Chrome instances that cloud AI agents spin up so they can click a button for you.

That’s it. A browser. Sitting in a rack somewhere, drawing power, holding a few hundred megabytes of RAM, so it can load Gmail and locate the reply box.

Meanwhile there’s a Chrome on this laptop with 31 tabs in it. Already running. Already signed into everything. It cost nothing to start this morning and it will cost nothing when I close it tonight.

Where the rented browser shows up on your invoice

You will never see a line item labeled “browser compute.” What you see instead is a pricing page with three columns and a lot of asterisks.

Cloud agent products have to recover the cost of every session they boot, so the economics get laundered into tiers: a free plan capped at some small number of tasks, a $20 tier that quietly meters “agent runs,” and a per-task price for anything heavier. OpenClaw started charging extra for exactly this reason, and I wrote about the agent tax that came with it. Nobody’s being greedy. Chrome in a datacenter is a real cost and the invoice has to land somewhere.

But the shape of that invoice tells you what you’re actually buying. You’re not paying for intelligence. You’re paying for a machine to be awake.

The session starts as a stranger, which is the expensive part

Cost is the boring half of this. The interesting half is that every one of those rented browsers boots up with no idea who you are, and the workarounds for that are worse than the problem.

A fresh cloud Chrome has no cookies, no session tokens, no saved passwords, no 2FA device trust, no browser fingerprint history with any site you use. So the product has to bridge the gap somehow, and the options are all a bit grim: OAuth flows that hand a third party standing access to your mailbox, cookie exports that you paste into somebody’s dashboard, credential vaults, or a “log in for me” prompt that asks you to type your bank password into an agent’s viewport. Then Cloudflare looks at a datacenter IP running an automated browser with a blank profile, decides it’s a bot, and serves a challenge that the agent fails. Which it will, repeatedly, because that’s what bot detection is designed to do.

I’ve watched this fail on internal tools that don’t even have a login page, just an SSO redirect that trusts the device. The agent gets bounced at the door and reports back that the page was empty.

Dassi takes the other road. It runs in the side panel of the Chrome you already have open, on the tab you’re already looking at, using the session you already authenticated. No token handoff, no profile to provision, no cold start. It’s on the Chrome Web Store and free. The reason it can be free is that the browser was never the expensive part for us. It’s your browser. We don’t rent it.

Zero marginal compute

Your Chrome is already open. Adding an agent to it costs the provider nothing per session, which means there’s nothing to meter, which means no tier exists to gate you out of your tenth task on a Tuesday.

BYOK, or: the only bill is the one you signed up for

There is still inference to pay for, and I’m not going to pretend otherwise. Language models cost money per token and always will.

The difference is who holds the contract. With bring-your-own-key, you point the agent at Anthropic or OpenAI or Google or DeepSeek, and the tokens bill to an account you opened, at whatever rate that provider publishes, with no reseller margin stacked on top and no second party deciding your monthly cap. Or you sign in with a ChatGPT subscription you’re already paying for and use the thing you already bought.

What you don’t pay for is somebody else’s datacenter Chrome. That’s the whole trick, and it’s a damn simple one. The model is identical either way. Same weights, same API endpoint, same latency. Only the plumbing changed, and the plumbing is where the $3.5 billion goes.

I’ve argued before that cloud browser agents can’t see your tabs, and cost is really just the same fact viewed from the accounting side. A browser that doesn’t know you is expensive twice: once to run, once to authenticate. A browser that already knows you is free on both counts.

Nscale will probably get its money. The buildout continues regardless of what any of us think about it, and some enormous fraction of that capacity will do real work that genuinely needs a GPU farm.

Loading your inbox in a rack in Texas isn’t that work.