Anthropic Builds Its Own Source Control System Scion: The Performance Dilemma Behind 30,000 PRs a Day

When an AI company's code submissions reach 30,000 pull requests a day, can Git still hold up? Anthropic's answer is no. According to tech media outlet The Pulse, Anthropic chose to build its own source control system called "Scion" after Git exposed severe performance bottlenecks in ultra-large-scale collaboration scenarios, and it is already in use in actual operations.

Why Git Fails at 30,000 PRs/Day

Anthropic's engineering team told The Pulse that when the collaboration scale of a codebase climbs to about 30,000 PRs per day, Git's response speed and merge efficiency begin to decline noticeably. For a team centered on model training and AI product iteration, this latency translates directly into a loss of R&D tempo—time spent waiting for code merges, branch switches, and history queries keeps getting amplified, ultimately becoming a systemic bottleneck for engineering efficiency.

The report did not disclose specific performance metrics, but the fact that Anthropic was willing to invest engineering resources into building an entirely new system suggests the problem clearly could no longer be solved through routine configuration tuning or process restructuring.

Git's performance bottlenecks are closely tied to its underlying data model and mode of operation. Git uses a content-addressable filesystem, and every commit, merge, and branch switch involves reading and writing a large number of objects (commit, tree, blob). In ultra-large-scale repositories, the longer the history and the more branches there are, the larger the object graph that operations like git log, git blame, and git merge must traverse, causing time consumption to grow non-linearly. In addition, Git's merge strategy is based on three-way merge, and when the number of concurrent branches surges, the computational overhead of conflict detection and merging rises significantly. At the hosting platform level, PR creation, review, and merging also involve additional metadata operations and permission checks, which likewise amplify latency under high concurrency. Therefore, when tens of thousands of PRs are reached per day, the decline in Git's response speed and merge efficiency is not a problem in any single link but the combined result of the data model, algorithms, and platform architecture.

What Scion Is: An Extension of Underlying Engineering Capabilities

Scion is a new-generation source control system developed internally by Anthropic, positioned to replace or supplement existing Git workflows to support its high-frequency, large-scale code collaboration needs. One detail: Anthropic did not choose to open-source it. When asked about this decision, the team responded (the report did not elaborate on the full reasoning), which also means Scion will not become public infrastructure like Git in the short term.

This choice itself is worth pondering: an AI company known for open research chose a closed-source path at the infrastructure level, suggesting that Scion may be deeply coupled with its internal training pipeline and model R&D processes, making it difficult to generalize if made public and potentially exposing internal engineering details.

The Industry Signal of Building a Custom Code Hosting System

Anthropic is not the first company to rebuild its development infrastructure due to scale issues, but the barrier to "building a custom source control system" is far higher than building custom CI/CD or code review tools. Git has existed for years as the industry standard for distributed version control, and replacing it means re-solving a series of foundational problems such as branch models, merge strategies, permission management, and audit trails.

This also indirectly reflects the code production intensity of leading AI companies: model iteration, experimental branches, and multi-team parallel development are causing cracks in the traditional "Git + hosting platform" combination under extreme loads. If Scion proves effective in practice, more high-intensity AI R&D teams may re-examine their version control foundations in the future.

The technical barriers to building a custom source control system manifest at multiple levels. First is the design of the branch model: it is necessary to decide whether to adopt a centralized or distributed branching strategy and how to efficiently handle mixed scenarios of long-lived and short-lived branches. Second is the merge strategy: beyond traditional three-way merge, real-time collaborative merge algorithms based on CRDTs (Conflict-free Replicated Data Types) or operational transformation (OT) must be considered to reduce the probability of conflicts under high concurrency. For permission management, fine-grained access control (such as role-based access control RBAC or attribute-based encryption ABE) must be implemented, and permission changes must take effect immediately. Audit trails require the system to record every operation in a tamper-proof manner, typically combining log-structured storage and hash chain technology. In addition, the storage layer must support high-throughput object reads and writes and version snapshots, potentially involving custom development of distributed filesystems or object storage. Re-solving these foundational problems means a custom system requires an engineering team with knowledge across distributed systems, storage, and compiler principles, with investment cycles often measured in years.

Why Not Open Source: An Unresolved Mystery

The biggest mystery surrounding Scion is why Anthropic chose to keep it closed source. From an engineering perspective, source control systems are inherently strongly coupled; if the internal implementation depends on a specific storage layer, permission system, or linkage mechanisms with training clusters, the maintenance costs and security risks after open-sourcing cannot be ignored. On the other hand, open-sourcing also means facing community expectations for generality, documentation, and long-term support, which is no small burden for a team whose main business is model R&D.

The Pulse's report leaves this question open without giving a complete answer. What is certain is that Scion's emergence shows that competition among AI companies has extended from the model layer to the engineering infrastructure layer—whoever can organize large-scale code collaboration more efficiently may gain an advantage in iteration speed.

Implications for Developers

For most teams, 30,000 PRs a day remains an unreachable scale, and Scion's details are difficult to reuse directly. But this event provides an observation window: when collaboration scale crosses a certain critical point, the performance ceiling of general-purpose tools becomes a real business constraint, and building custom infrastructure thus shifts from "optional" to "mandatory." For rapidly expanding AI teams, assessing the capacity of their version control systems in advance may be more cost-effective than rebuilding afterward.

One-line Summary

Anthropic built the closed-source source control system Scion due to Git's performance bottlenecks at a scale of 30,000 PRs per day.

Paragraph Summary

Anthropic's engineering team found that when code collaboration scale reaches about 30,000 pull requests per day, Git's response speed and merge efficiency decline noticeably, becoming a systemic bottleneck for R&D efficiency. To address this, Anthropic built a new-generation source control system called Scion to replace or supplement existing Git workflows. Scion is not open source, and the team did not fully disclose the reasons, presumably because it is deeply coupled with internal training pipelines and engineering details. Building a custom source control system has an extremely high barrier, requiring re-solving foundational problems such as branch models, merge strategies, permission management, and audit trails. This event reflects that competition among leading AI companies is extending from the model layer to the engineering infrastructure layer, and the performance ceiling of general-purpose tools may become a real business constraint for rapidly expanding teams.