Multiple browser profiles on one machine: about 340 MB each

How many browser profiles one machine runs for AI agents: what a profile is, the memory a running browser takes, what refuses a start, and why plans count browsers.

Davide Valli · Founder, Scalebrowser

Published Oct 27, 2026 · 6 min read

Short answer

A running browser with a visible window took about 340 MB of physical memory with one ordinary page and 580 MB with three tabs in our measurement, while a stored profile costs nothing but disk, so one machine holds any number of profiles and runs as many at once as its free memory allows. Before every start the daemon checks three limits, a browser count, a memory budget and the machine's measured free memory, and refuses the start instead of letting the machine run out. Plans count browsers running at once and machines, not profiles: Fleet runs 50 at once on 10 machines for 249 USD a month.
340 MB, per running browser, beside a memory module with four chips

A hundred accounts need a hundred identities, and the first instinct is to count machines, or to count what a vendor charges per profile. Both counts start from the wrong unit. A profile that is not running is a folder on disk; only a running browser takes memory, and an agent rarely needs all hundred at the same moment.

What is a browser profile?

A browser profile is the directory in which one browser identity keeps everything that makes it the same browser tomorrow: cookies and logins, local storage, history, saved passwords, passkeys and settings. Chrome uses profiles to keep two people's browsing apart on one computer, inside one browser installation. For agents the word means more: each profile is its own browser with its own fingerprint, its own storage and its own network exit, so that two profiles look like two devices rather than two windows of one.

Two states follow from that definition, and they cost very different things:

StateWhat it holdsWhat it costs
Storeda directory with cookies, storage, passkeys and a stored identitydisk space only
Runninga browser process tree with its renderers and graphics processmemory, CPU and a concurrency slot

A fleet of 200 profiles with 5 browsers open at a time is the normal shape of agent work, because most accounts wait between tasks.

How much memory does each running browser need?

A running browser with a visible window took 334 MB of physical memory with one ordinary page and 583 MB with three tabs open when we measured it, and Google Chrome on the same machine stayed within 8 percent at every stage. Our plans documentation gives 250 to 400 MB for an ordinary page, which matches the one-tab case. Physical memory is what Task Manager shows for the browser; reserved memory is what Windows has to back with memory or the page file, and it runs about twice as high.

StagePhysical memoryReserved memoryGoogle Chrome, physical / reserved
Empty tab after 30 seconds138 MB282 MB131 / 292 MB
One tab after Wikipedia, GitHub and Hacker News334 MB664 MB343 / 694 MB
Three tabs with the same pages583 MB1,051 MB543 / 996 MB
Source: Scalebrowser measurement on one Windows 11 machine with 63 GB, engine 151.0.7922.99 with a visible window against Google Chrome 154.0.8037.57, 2 runs per stage, 27 September 2026

A heavy web application in a visible window takes more than these pages. A number per browser therefore sets a ceiling, not a promise, which is why the daemon asks the machine before every start. Three limits apply, and the strictest one wins:

  1. A browser count per machine, max_concurrent, 64 by default.
  2. A memory budget, ram_budget_mb, 24,576 MB by default: how much of the machine the browsers may take.
  3. The machine's measured free memory, read at the moment of the start, because a budget written for a 128 GB workstation says nothing about a 16 GB laptop.

A start that would break any of them is refused with error 4003, "this machine is full", before a browser is launched; the configuration reference lists the settings. The alternative is the one every operator has seen once: the browser that tips the machine over does not fail alone, it takes the running ones down with it, in the middle of their sessions.

A stored profile costs disk; a running browser took 334 MB with one page. Before every start the daemon checks three limits and refuses with error 4003 instead of letting the machine run out.

The memory arithmetic sets the upper end for common machine sizes. It divides the machine by the measured physical memory, and the operating system and your own programs take their share first:

Machine memoryBrowsers with one tab, 334 MB eachBrowsers with three tabs, 583 MB eachDefault limits that cap it
16 GBat most 49at most 28measured free memory
32 GBat most 98at most 56the 24,576 MB budget, which counts 400 MB per browser, so 61; with three tabs each, free memory runs out first
64 GBat most 196at most 112the budget and the count of 64, until you raise them

A bulk start stops at the first refusal of this kind and reports the profiles it never tried, because running out of memory says nothing about the profile that happened to be next.

Why pay for concurrency instead of profiles?

Paying for concurrency matches the bill to what costs something, because a profile at rest is a directory and a running browser is memory. Chromium's user data directory documentation describes a profile as "a subdirectory (often Default) within the user data directory" holding "history, bookmarks, and cookies". Charging per profile charges for that directory, and it pushes the buyer to reuse one identity for several accounts, which is exactly what makes accounts look linked. Hosted browsers bill a third unit, time: Browser Use charges 0.02 USD per browser hour, rounded up to whole minutes with a 1 minute minimum.

PlanPrice per monthBrowsers at onceMachinesProfiles
Start19 USD11unlimited
Solo39 USD51unlimited
Studio99 USD203unlimited
Fleet249 USD5010unlimited
Source: scalebrowser.net/pricing and the plans documentation, read 29 September 2026

Both numbers count across the whole account, so three machines with two browsers each are six browsers at once. Past Fleet, 5 more browsers cost 40 USD a month and one more machine 25 USD. Running 100 profiles on Solo means 5 of them work at a time and the other 95 wait on disk with their logins intact; an agent that leases several at once gets back whatever the plan and the machine allow, plus the reason it stopped. The guide to an agent browser on your machine covers what else a local browser changes, and the fleet documentation covers groups, presets and bulk starts. How Claude Code works with a real browser, and what each step costs in tokens, is covered in Claude Code with a real browser.

Scalebrowser

Scalebrowser gives each agent its own isolated browser with a persistent identity, on your own machine, so a run stays signed in, handles the captcha and finishes without anyone watching it.

See how Scalebrowser does it

Which exit should each profile use?

Each profile should leave through its own address, and one bought residential or mobile gateway is enough for many profiles, because a session tag in the username or password turns it into many stable exits. The daemon knows that spelling for fourteen providers and refuses to give one address to two profiles: a start whose address already belongs to another profile draws a new tag, up to five times, and is refused with error 4013 if every draw is taken. How long an address stays one profile's depends on its kind, 3 days for mobile, 30 for residential and 90 for datacenter, and the proxies documentation explains why. A datacenter exit stays a datacenter exit whatever the browser does, which is why sharp targets call for mobile proxies or residential ones.

Run it on your own machine

Seven days to try it with your own agents on your own sites. Starting the trial needs a card.

Written by

Davide Valli

Founder, Scalebrowser

Builds Scalebrowser, the browser layer for AI agents that runs on your own machine. Measures every change a web page could observe against a real browser before it ships, and writes up the ones that turned out wrong.

All articles by Davide Valli