GitHub strengthens its infrastructure to absorb the wave of AI agents

GitHub adds millions of CPU cores and hundreds of petabytes of storage after reliability incidents and a sharp increase in AI-related commits. The platform also deploys Copilot Fusion and new safeguards to secure code.

GitHub significantly strengthens its infrastructure after several reliability incidents and an explosion of AI-generated activity: in August, the platform had already added more than 3 million CPU cores and 120 petabytes of fast storage, and now announces more than 4 million additional CPU cores and 160 PB of storage. Meanwhile, GitHub reported 7.38 billion commits in September, five times more than the previous month, in a context where agents are changing how software is developed. The challenge is not only to handle the load; it is also to prevent tooling from collapsing under artificial volumes.

Quand la capacité ne suffit plus, l’architecture doit suivre

Jay Parikh, Executive Vice President of CoreAI at Microsoft, sums up the situation bluntly: “Agents are changing the volume and shape of software development. Adding capacity matters, but capacity alone won’t get us there. We’ve also had to change the architecture,” he writes in a post published before GitHub Universe, the platform’s flagship annual conference. In other words, GitHub is not just adding servers in a chain.

The company is also rebuilding its Git architecture to separate computation from storage, aiming to achieve up to 35 times more throughput and limit saturation risks under synthetic load. This choice matters for development teams, as code tools are now used not only by humans but also by agents who open, modify and test code at high speed. The issue becomes almost trivial for machines; much less so for engineers.

Copilot Fusion, un routage intelligent entre plusieurs modèles

GitHub has also made Copilot Fusion generally available, a routing engine that chooses which AI models to use depending on task difficulty, balancing cost, latency and reasoning capacity. The service introduces three effort levels — Standard, High and Max — and GitHub claims Fusion Max improves task execution rates compared to high-end benchmarks while capping inference cost at 1.1 times that of the base model.

The principle is simple: not all requests need the same model or the same price. “Different tasks need different models. Developers shouldn’t have to manage every model choice to get the right result at the right price,” writes Jay Parikh. Fusion analyzes what a task requires — reasoning, code generation, debugging, tool usage — then distributes execution over one or more models. For customers, it is mainly a promise of smoothness; for GitHub, it is a way to keep costs under control.

Sécuriser la chaîne logicielle pendant que l’IA accélère

GitHub is also deploying several protections aimed at the software supply chain and libraries. The platform launches AI-assisted code scanning to check every pull request and repository for vulnerabilities, as well as scope analysis to prioritize alerts depending on whether the code calls a vulnerable function. The latter adds a call path directly viewable in GitHub, helping teams sort real risks from false positives faster.

Another component: a package firewall that inspects npm packages at download time to block known malicious packages, typosquatted names and zero-day dependency attacks. GitHub thus aims to reduce hallucinations in libraries and limit unpleasant surprises in build chains. Developers save time, but attackers also seek to gain time; the balance of power remains very concrete.

Copilot veut aussi devenir un chef de projet

The Copilot app is finally enriched with broader functions, making it look less and less like a simple code assistant. Copilot Projects allows tracking goals, context, active terminal sessions and ongoing code diffs across multiple development tracks. Copilot Connectors natively integrates the Model Context Protocol to retrieve production telemetry and deployment contexts from Cloudflare, Vercel and Sentry, in the CLI, web app and VS Code.

Copilot Automations adds reusable background agent tasks. And to further expand usage, GitHub announces that access to Copilot is extended to all GitHub Enterprise users without additional license purchases. Already in September, the company rewrote in Rust the runtime powering Copilot CLI, the app and the SDK, after noticing it added about 100 MB of working memory per client and complicated process management, impacting startup speed, throughput and server density.

By the numbers

  • 3 million CPU cores — added in August after reliability incidents.
  • 120 PB — fast storage added in August, according to GitHub.
  • 4 million CPU cores — new capacity announced before GitHub Universe.
  • 160 PB — additional storage announced with the new scale-up.
  • 7.38 billion commits — recorded in September, 5 times the previous month.

To read