A Billion Personal Agents, Starting from Scratch

The scale I mean is not one person vetting a law firm. It is a billion people, each with a personal agent. Grok, Claude, OpenAI's Dots, Meta's Muse, and whatever ships next. Ten times a week that agent goes looking for a service or a product: an accountant, a courier, a sofa, a clinic. Ten a week is about 500 a year. A billion people is 500 billion searches. And if each agent starts from the website, every one of those searches rereads pages the business already knows by heart.

That is the case for an AI website profile, a short file at /llms.txt. One search is a rounding error. Five hundred billion of them, each reconstructing the same facts from menus and cookie banners, is a bill for compute and a bill for electricity. The profile is how the bill comes down. The arithmetic below is our estimate. The inputs are in the open.

One search, then the year

Start with a single search that compares eight firms. In testing we have published, text extracted from an 80KB homepage is about 18 percent of the file: roughly 14,400 characters, or about 3,600 tokens. Inner pages weigh about 40 percent of that, about 1,440 tokens. A homepage plus four inner pages is about 10,000 tokens for one firm. Eight firms, read once, are about 80,000 tokens. Eight profiles of 1,500 tokens are 12,000. The cut is 68,000 tokens, 85 percent.

That homepage is a heavy example. The HTTP Archive's July 2025 crawl put the median desktop HTML response at 35KB and the 75th percentile at 78KB. A median site is a smaller job. The percentage cut holds, because the profile stays 1,500 tokens when the template grows.

Now multiply. 500 billion searches times 80,000 tokens is 40 quadrillion tokens of page text in a year. The same searches over 12,000-token profiles are 6 quadrillion tokens. The difference is 34 quadrillion tokens that never had to be read. If the agent keeps each page in the thread, one eight-firm search is about 260,000 tokens rather than 80,000, and the year scales by the same three times.

Anthropic lists Claude Sonnet 5.5 at $2 per million input tokens. Thirty-four quadrillion tokens at that price is about $68 billion. Claude Haiku 5.5 is $0.10 per million on prompts under 100,000 tokens, which brings the same tokens to about $3.4 billion. The cheaper model changes the invoice. It does not change the number of pages fetched, or the wait while an agent opens forty of them.

What that year costs in electricity

Google measured the median Gemini Apps text prompt in May 2025 at 0.24 watt-hours, counting the chips, the machines around them, idle capacity, and the data centre overhead. Five hundred billion prompts at that weight are 120 billion watt-hours, which is 0.12 terawatt-hours. The International Energy Agency puts data-centre electricity at about 415 terawatt-hours in 2024. Priced as a single short reply, the year is real and small beside that total.

A search that opens eight websites is not a short reply. The IEA's later briefing says reasoning and agentic tasks can use hundreds or thousands of times the energy of simple text generation. At 100 times Google's prompt, one search is 24 watt-hours and the year is 12 terawatt-hours, about 3 percent of data-centre electricity in 2024. At 1,000 times, one search is 240 watt-hours and the year is 120 terawatt-hours, about 30 percent of that same total. Those two ends are the IEA's words, hundreds and thousands, applied to a measured prompt. They are not a meter clamped to one law-firm site.

Microsoft Research, on frontier models, found that 15 times the tokens raised the energy of a query about 13 times, from a median 0.34 watt-hours to 4.32. In that test, energy moved with the reading. An 85 percent cut in the tokens of the crawl is the same order of cut in its electricity. On the agentic year, that is on the order of 10 terawatt-hours avoided at 100 times a short prompt, and about 100 terawatt-hours at 1,000 times. The profile does not delete the search. It is how 500 billion searches stay near a short read, instead of landing in that 12 to 120 terawatt-hour band.

This is a scenario. Meta introduced Muse in September 2026 and described it as an agent built to work for billions of people. OpenAI's Dots are personal agents too. Neither number above is a headcount of users this month. It is the year those products are pointed at, if each person really does hand an agent ten lookups a week and the agent really does begin at the raw site.

The objection worth keeping

Labs will cache. A billion agents will not each download the same homepage from scratch if a shared index already holds it. The waste that remains is every distinct business, every page that changed, and every agent that does not share a cache with the others. A 1,500-token profile is a small thing to store and a small thing to refresh. Forty pages of HTML, re-extracted into a fresh thread, are not.

Cut the scenario by ten. One hundred million people, or searches that are a tenth as heavy. The agentic year falls to about 1 to 12 terawatt-hours. The top of that range is still a few percent of all the electricity data centres used in 2024. The direction does not depend on the billion being exact.

A profile is the business describing itself. It does not check a licence, read the reviews, or replace the page that closes the sale. It covers the first pass: who this is, who they serve, where they work, how to start. And a thin file fails it. In a scan of 69 small-business sites we published in June 2026, 24 served a real llms.txt and 4 were properly structured. StackScan, updated on 31 October 2026, got a file back from 18.0 million of 151.6 million live sites, and 19.9 percent of those files were empty or described a site that was not open. A template with the brand name swapped in adds a fetch. It does not move a search from the crawl to the short read.

What has to be true for the saving to show up

On 3 September 2024 Jeremy Howard proposed /llms.txt as a short summary plus links to the pages an agent should actually read, ideally as clean markdown. The 10 August 2026 revision keeps the file small enough to fit in context. The detail lives in the links. A dump of the whole site into one file spends the tokens the profile was meant to avoid.

If you publish a site, write that short file. Say what you do, who it is for, where you operate, and how a client starts. Link the few pages that prove it. If you run an agent, fetch the file first and stop when it answers. Open the site when you are checking a claim the profile should not be trusted to settle.

We build these profiles at Platinum. The year above does not depend on that. It depends on a billion agents having something shorter than the website to read, and on the file being specific enough that they use it.

Sources

1. Google Cloud, Amin Vahdat and Jeff Dean, "Measuring the environmental impact of AI inference," 21 August 2025. Median Gemini Apps text prompt in May 2025: 0.24 Wh, including chips, host machines, idle capacity, and data-centre overhead. https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference
2. International Energy Agency, "Executive summary," Key Questions on Energy and AI. Agentic and reasoning tasks can use hundreds or thousands of times the energy of simple text generation. Replacing every conventional search with a simple AI text query would use under 4 TWh a year. https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary
3. International Energy Agency, Energy and AI (2025). Data centres about 415 TWh in 2024, around 1.5 percent of global electricity; base case about 945 TWh in 2030. https://www.iea.org/reports/energy-and-ai
4. Microsoft Research, "Energy Use of AI Inference: Efficiency Pathways and Test-Time Compute," September 2025. Median 0.34 Wh per frontier query; 4.32 Wh when the query carried 15 times the tokens. https://www.microsoft.com/en-us/research/publication/energy-use-of-ai-inference-efficiency-pathways-and-test-time-compute/
5. Meta, "Introducing Muse," September 2026. Describes Muse as a personal agent built to work for billions of people. https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/
6. Jeremy Howard, "The /llms.txt file," llmstxt.org, 3 September 2024, revised 10 August 2026. https://llmstxt.org/
7. HTTP Archive, "Page Weight," Web Almanac 2025, July 2025 crawl. Median desktop HTML response 35KB; 75th percentile 78KB. https://almanac.httparchive.org/en/2025/page-weight
8. Anthropic, Claude API pricing, read 13 November 2026. Claude Sonnet 5.5 at $2 per million input tokens. Claude Haiku 5.5 at $0.10 per million input tokens under 100,000 tokens. https://platform.claude.com/docs/en/about-claude/pricing
9. StackScan, "llms.txt Statistics 2026," published 10 September 2026, updated 31 October 2026. 17,956,268 of 151,649,745 live sites returned a file; 19.9 percent of those files were empty or described a closed site. https://www.stackscan.com/blog/llms-txt-statistics
10. Our own figures for the 80KB homepage, the 18 percent extraction, and the June 2026 scan of 69 sites are in "What Happens When an AI Agent Tries to Read Your Website" and "Your llms.txt Is Probably Garbage," on this blog.