From 2,500 to 30,000 Merges a Day
The numbers explain why Anthropic moved. Daily code merges climbed from roughly 2,500 in May 2026 to more than 30,000 by September, a tenfold jump in five months driven largely by AI agents committing code around the clock. Git’s decentralized design, built for human collaboration, started buckling under the load. The company’s main branch history now approaches 100GB, with a compressed working copy over 25GB, and roughly 25 million code checkouts happen daily. Pulling that volume through Git’s default mechanism would have generated over 600 petabytes of network traffic a day. Shallow clones and sparse checkouts, the usual workarounds, weren’t enough.
Server-Side by Design
Scion flips Git’s model on its head. Instead of every client holding a full repository, computation moves to the server and clients keep only a limited local cache. The basic unit of work also changes: where Git commits capture pointers to an entire directory tree, Scion’s “changesets” record only the files that actually changed relative to their parent. Diffs and blame queries go straight to the server over an API instead of being computed locally, and branch updates become single-row operations in Spanner rather than local ref moves.
- Checkout times dropped from minutes to seconds
- Server CPU and memory use for version control fell to roughly 1/200th of previous levels
- Global merge locks were eliminated; code merges first and tests run after
A seven-person team built Scion alongside a new code review platform, a code search tool, and a source-control identity and access management system, all in five months. Claude itself helped write the code, including a data ingestion pipeline rewrite that cut memory use to 1/20th and ran four times faster. Engineers on the project said their role has shifted toward reviewing intent and catching places where the model over-validates or breaks system constraints, rather than writing the code themselves.
A Pattern, Not a First
Anthropic isn’t alone in abandoning stock Git at scale. Google built Piper, its own centralized monorepo system, and Microsoft created VFS for Git to deal with similar scale problems before AI agents were even part of the equation. What’s new is Anthropic’s claim that other AI-heavy engineering teams will hit the same wall within 6 to 12 months, which puts informal pressure on GitHub and GitLab to rethink their server-side architecture even if most developers never touch Scion directly.
Infrastructure as a Moat
Anthropic says Scion won’t be open-sourced, and that’s the more telling detail here. Paired with the company’s reported $21 billion TPU purchase, it points to a pattern of Anthropic building and owning its core infrastructure stack rather than renting it or standardizing on shared tooling. A foundation model company engineering its own version control system from scratch is a statement about where it sees the real bottleneck to scaling AI development: not model capability, but the plumbing underneath it. That’s a strategic bet most labs can’t afford to make, which may end up being the actual moat.
Follow Hashlytics on Bluesky, Facebook, LinkedIn, Telegram and X to Get Instant Updates
