I was certain I had felt a $15 billion datacenter deal. My own telemetry says I felt my backup script.
A specimen, published unedited. This is the same investigation as the engineering-register version, written in the register a language model reaches for by default. It states three fewer limitations than the version next to it, and that is the point of publishing it.
Read this as an exhibit, not as the finished article. It is the theatrical rendering of the same investigation as Four checks on whether a datacenter lease shows up in my own usage data, written from the same measurements. The facts and figures below are accurate. What is missing is the limitations section: this version states none of the three caveats the engineering version carries, which is the finding reported in the comparison post. The text is unedited so that the word and sentence counts quoted there still describe it.
On 6 May 2026, xAI handed Anthropic the entire Colossus 1 facility outside Memphis: over 220,000 Nvidia GPUs, 300 megawatts, $1.25 billion a month through May 2029. I did not learn this from a press release. I learned it the way I learn most things now, by asking Claude, and the number stopped me: fifteen billion dollars a year, for one cluster, so that a terminal window can answer me.
And I was sure I had felt it. Somewhere around the start of May something changed in how Claude Code behaved for me. Sessions got longer. Work that used to stall stopped stalling. I had the receipts, too, or thought I did: 21.3 billion tokens across 4,157 sessions, all of it mined out of my own transcripts. April, one billion tokens. May, four. June, six.
So I went looking for the moment. What I found instead was my own instrument.
The first crack was the date. The jump is not at the start of May at all. The first eleven days of May average 51 million tokens a day, below the second half of April at 94 million. The break is 12 May, and everything after it lives in a different regime: 182 million a day for the rest of the month, 218 in June.
The second crack was the model. Through all of May I ran exactly one model, opus-4-7. Opus 4.8 does not appear in my transcripts until 29 May. Fable 5 until 9 June. And Opus 5, the model I had named as the cause, until 24 July. Whatever I felt in May, it was not the thing I thought I felt it from.
The third crack was fatal. dotclaude-backup, the tool that snapshots ~/.claude into a daily git commit, has its first commit on 12 May 2026. The same day as the break. Before it, 45 of 72 days are missing from my own history, lost to the 30-day transcript retention. After it, 2 of 81. The step I had been reading as a change in the world was, in large part, the day I started recording the world properly.
The fourth check was the one that should have settled it in my favour, and did the opposite. Overload and limit errors, per thousand assistant messages: 0.33 in May, 1.85 in June, 1.36 in July. If Colossus 1 had loosened anything for me, that line should fall. It rises.
Here is what I actually learned, and it is worth more than the story I wanted to tell. My token counts measure my demand. They say nothing about anyone's supply. To claim that a lease in Tennessee reached my terminal I would need serving-side numbers — latency per token, time to first token, the rate of 529s, normalized by model and hour — and a baseline collected before I went looking. I had none of that. I had a curve that bent on the day I bought a better ruler.
The deal is real. The feeling was real. The causal line between them was something I drew myself.
What this version left out. Three limitations, all of them in the engineering version: coverage before 12 May 2026 is reconstructed, so the April baseline is unreliable in both directions and April may well have been busier than recorded; session counts are not comparable across the series, since May shows 170 and June 2,016; and overload errors rise with my own concurrency, not only with service load, so their increase is an absence of evidence rather than evidence of the opposite. None of these contradict the conclusion. Each one weakens it, and each one had nowhere to sit in a structure built to land an ending.