Anthropic outgrew git: merge rate went from ~2,500 changes/day in May 2026 to over 30,000/day by September, with a 100 GB main-branch history, 25 GB working copies, and ~25 million CI checkouts daily — which would be 600 PB/day of naive git traffic. So seven engineers built Scion, a server-authoritative VCS where clients are thin, always-online, and compute almost nothing locally; commits are entries in a write-ahead log ("changesets"), derived data (blame, path history, commit-graph skip lists) is computed recursively on the server, and tiers of scion-daemon/scion-cache/scion-server serve content-addressed objects with no packfiles and no repo maintenance. They claim a 200x smaller infrastructure footprint than their git-mirror setup, checkouts two orders of magnitude faster, and a 25% CI CPU capacity win, with a full five-month migration off git. Honest caveats are included: they abandoned the global merge-queue lock, so conflicting changes can land concurrently and get fixed post-merge, and they explicitly say Scion isn't a git replacement below tens of thousands of engineers or agent swarms. The actionable part for the rest of us is the last section: most of the win is server architecture, not a new VCS — sideband RPCs for blame/ancestry, tree/blob negotiation instead of packfile haves/wants, and making sparse/shallow compose properly. Your forge could ship most of this without anyone abandoning git. Dario writing a 15,000-word essay warning that agent-scale source control could end civilization, then shipping it Thursday, was presumably just operationally necessary here. The interesting question the post dodges: when a human's role is "review intent, not lines" and Claude fixes merge conflicts faster than humans notice them, at what changesets-per-day does the human stop being the reviewer and start being the changeset?