GitHub 新增 400 万 CPU 核心，并发布 Copilot Fusion 多模型路由

After provisioning more than 3 million CPU cores and 120 petabytes of high-speed storage in August following a series of reliability incidents, GitHub is now doubling down on expanding its infrastructure by adding more than 4 million CPU cores and 160 PB of storage.

This comes as GitHub experienced a massive surge in AI-driven commit activity recently, with 7.38 billion commits on GitHub in September alone—a 5x increase over the previous month.

"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," Jay Parikh, EVP of CoreAI, Microsoft, wrote in a blog post ahead of GitHub Universe, the hosting service’s flagship annual developer conference.

The company is also rebuilding its Git architecture to decouple compute and storage to enable a possible 35x more throughput and to prevent developer tools from buckling under synthetic load.

At the conference, GitHub also announced the general availability of Copilot Fusion, an intelligent routing engine that decides which frontier AI models to dispatch based on task difficulty and to balance cost, latency, and reasoning capability.

Fusion introduces three effort tiers—Standard, High, and Max. According to GitHub, the flagship Fusion Max tier improves task-completion rates over top-tier frontier baselines while capping estimated inference cost at 1.1 times the baseline model.

“Different tasks need different models. Developers shouldn’t have to manage every model choice to get the right result at the right price. Copilot Fusion evaluates what a task needs across reasoning, code generation, debugging, and tool use, then chooses how to execute across one or more models,” Parikh wrote in the blog.

The San Francisco-headquartered company is also rolling out critical safeguards to protect software supply chains and prevent hallucinations in libraries.

It has introduced AI-powered code scanning to check every pull request and repository for vulnerabilities. It is also rolling out reachability analysis to help teams prioritise alerts by showing whether their code calls a vulnerable function, with a call path they can inspect in GitHub. Moreover, the package firewall will inspect npm packages at download time to block known malicious packages, typosquatted names, and zero-day dependency attacks.

“Moving faster only helps if we can maintain the quality of what ships. Copilot Code Review surfaces meaningful, actionable bugs that developers need to catch," the Microsoft executive noted.

GitHub has also added more features to the Copilot app to evolve it into a project manager.

Copilot Projects lets developers track goals, context, active terminal sessions, and work-in-progress code diffs across parallel development tracks. Copilot Connectors natively integrates the Model Context Protocol, letting Copilot pull real-time production telemetry and deployment contexts from services like Cloudflare, Vercel, and Sentry across the CLI, web app, and VS Code. Meanwhile, Copilot Automations lets teams configure reusable, background agent tasks.

GitHub also announced that Copilot access is expanding to all GitHub Enterprise users without additional seat purchases.

Earlier, in September, GitHub rewrote the runtime powering its Copilot CLI, app, and SDK in Rust. GitHub said this added roughly 100 MB of working-set memory per client, increased process-management complexity, and limited startup speed, throughput, and server density.