--- title: "I Gave Away Ten Billion Tokens" date: 2026-10-17 summary: "Building four open-source systems took about 9.9 billion tokens of AI compute, most of it on a working signals agent for AdCP, an open standard for agentic advertising. The code is MIT-licensed, so the next team can fork it instead of spending that compute again." standfirst: "Four open-source systems on the Building page took about 9.9 billion tokens of AI compute to build, and about 96% of it was the model re-reading context rather than writing code. In agentic building the search is the expensive part, so a fork passes the compute on. The essay shows how the count was made and corrected, and why a repo without a license isn't a donation." canonical: https://nofluffadvisory.com/writing/i-gave-away-ten-billion-tokens/ --- *Building four open-source systems took about 9.9 billion tokens of AI compute. Call it ten billion. Most of it built a working signals agent for AdCP, an open standard for agentic advertising. The code is MIT-licensed. Fork it, and you skip most of that search.* --- ## The number The four systems are on my [Building](/building/) page. They live in three public GitHub repos, because the AdCP Ecosystem prototype sits inside the adaptor's. Most of the 9.9 billion went into one repo. The [AdCP Signals Adaptor](/building/adcp-signals-adaptor/) took about 8.85 billion tokens over 509 commits. AdCP is the Ad Context Protocol, and the adaptor is a working signals agent for it: buyer agents can find its audiences and activate them. The rest is [Agentic Audience Intelligence](/building/audience-intelligence/) (321 million), the [AI Growth Operator](/building/ai-growth-operator/) (530 million) and the [AdCP Ecosystem](/building/adcp-ecosystem/) prototype (181 million). A token is a small chunk of text, a few characters long. I counted every token the model processed, read or written. About 96% of the total is context the model re-read: the repo, the tests, the spec, its own earlier attempts. Only about 366 million tokens were new input and output. ## Code is what's left after the search When you build with agents, the expensive part is the search: the failed attempts and the reviews that send work back. A commit is the part of that search that survived. The adaptor has 509 of them. Most are decisions that held up, and behind each are attempts that didn't. When you fork a repo, you get the decisions without paying for the search. That's the donation. ## Why ad tech should care Ad tech is about to build the same plumbing many times over. Every seller that adopts AdCP needs an agent. Every data owner on AdCP needs a signals endpoint. Every brand team that builds audiences with agents has to pick a method, like a look-alike, and know when the data is too thin to build one. I co-lead AdCP's Signals and Measurement working group. A standard becomes real when a team can run working code against it on day one. Reference implementations lower the cost of saying yes. A team that rebuilds these systems from scratch could spend a number like mine to get there. The [ledger on my Building page](/building/#donated) has a slider for that math. If 105 teams reused all four systems instead of rebuilding them, that's about 1.04 trillion tokens the industry wouldn't have to spend again. That figure is a hypothetical. GitHub doesn't report who reuses code, so the reuse number on my site is a slider you set yourself. ## The count, corrected Here's the method. About 4 billion of the tokens are logged: every model response in my Claude Code session transcripts from June to October, attributed by file path. Work in a repo counts for that repo. Work on my site about a repo, like its case-file page, counts as site work. The other 5.8 billion is estimated. Commits those logs don't hold, the adaptor's early build from March to early June and all of Agentic Audience Intelligence, are priced at 14.6 million tokens each. That rate is a floor. It divides 33 billion logged tokens by all 2,260 non-merge commits on this site, and the 371 made before the logs began add commits but no tokens. The adaptor's own logged commits since June ran at more than twice that rate, and at least 225 of its 378 earlier commits carry a Claude co-author line. My first count said 10.8 billion. Before publishing it, I ran an adversarial review: agents whose only job was to break the number. They found three problems. The estimate priced merge commits, which add no work of their own. One day was counted twice, once in the logs and once in the estimate. And site work about the repos had been booked as work on the repos. Fixing all three took about 900 million tokens off the total. ## A repo without a license is a display When I went to publish the number, two of the three repos had no license file. The adaptor, which my own site called MIT-licensed, had none on GitHub. Neither did the AI Growth Operator. Without a license, public code is still all rights reserved. Anyone can read it. Nobody can safely build on it. Those two repos held about 97% of the tokens. Both now carry the MIT license, which lets anyone use, change and sell the code as long as they keep the copyright and license notice. If you publish code you want other people to use, add the license on day one. It's the cheapest step in the whole project. ## What to do with this If you're scoping an AdCP integration, fork the adaptor before you budget a build. If you're building with agents, count your tokens and publish the receipt. And if you sell into a standard you want adopted, put a working reference implementation in public. The four systems, the ledger and the method are on my [Building page](/building/#donated).