Meta's AI-Powered Capacity Efficiency: Automating Hyperscale Performance Optimization

Meta's Capacity Efficiency Program is a groundbreaking initiative that leverages unified AI agents to automate the identification and resolution of performance issues across its massive infrastructure. By encoding the expertise of senior efficiency engineers into reusable skills, these agents operate through a standardized tool interface, drastically reducing the time spent on manual investigations. This system not only recovers hundreds of megawatts of power—enough to power hundreds of thousands of homes annually—but also frees engineers from tedious tasks, allowing them to focus on innovation. The program combines offensive (proactive optimization) and defensive (regression detection) strategies, with AI accelerating both. Below, we answer key questions about how this unified platform works and its impact on hyperscale efficiency.

What is the Capacity Efficiency Program at Meta?

The Capacity Efficiency Program is Meta’s systematic effort to optimize power usage across its global data centers and servers. Since Meta serves over 3 billion people, even a tiny performance regression of 0.1% can cause massive energy waste. The program has historically relied on two pillars: offense (proactively finding opportunities to improve energy efficiency through code changes) and defense (monitoring production systems to catch regressions and pinpoint their root cause to a specific pull request). However, manual resolution became a bottleneck as the workload grew. To overcome this, Meta built a unified AI agent platform that encodes domain expertise into composable skills. These agents now automate both finding and fixing issues, enabling the program to scale without proportionally increasing headcount. The result is a self-sustaining efficiency engine that handles the long tail of performance problems.

How do unified AI agents help in finding and fixing performance issues?

Meta’s unified AI agents are designed to encode the specialized knowledge of senior efficiency engineers into repeatable, modular skills. These skills are combined through a standardized tool interface, allowing agents to automatically investigate performance issues on both sides of efficiency: offense and defense. For instance, when a regression is detected, an agent can dive into logs, trace metrics, and identify the root cause in minutes—a task that previously took engineers about 10 hours. The agents don’t just diagnose; they can also generate ready-to-review pull requests that fix the issue. This automation compresses hours of manual work into roughly 30 minutes, recovering hundreds of megawatts by stopping wasteful power usage quickly. Moreover, because the agents are built on a unified platform, they can be easily updated with new expertise without retraining from scratch, making the system highly scalable across Meta’s expanding product areas.

What is the difference between offense and defense in efficiency?

Efficiency at hyperscale is a two-front battle. Offense refers to proactive code changes that find optimization opportunities before they become problems. Engineers (now aided by AI) search for ways to reduce resource consumption—like improving algorithm efficiency or eliminating redundant computations—and deploy changes across the fleet. Defense, on the other hand, involves monitoring production systems for regressions (unintended performance drops) that sneak through testing. Meta’s in-house tool FBDetect catches thousands of such regressions weekly. The goal of defense is to quickly detect, root-cause, and mitigate these regressions before they compound across the fleet. AI agents accelerate both fronts: offense by autonomously finding and implementing optimizations, and defense by automating investigation and fix generation. Together, they form a continuous loop that keeps Meta’s infrastructure running efficiently while reducing human workload.

What is FBDetect and how does it work?

FBDetect is Meta’s internal regression detection system for production performance. It continuously monitors resource usage metrics—like CPU, memory, and network bandwidth—across the fleet. When it detects a statistically significant deviation from expected behavior, it flags it as a potential regression. FBDetect can correlate the change with recent code commits, often pinpointing the exact pull request that caused the regression. The tool catches thousands of regressions weekly. Traditionally, engineers would then manually investigate the root cause, which could take many hours. However, with the integration of Meta’s unified AI agent platform, FBDetect can now trigger automated analysis. AI agents take the regression signal, examine relevant logs and traces, and quickly identify the faulty code change. They can even generate a mitigation pull request if the fix is straightforward. This reduces the time from detection to resolution from hours to minutes, preventing wasted megawatts from compounding across the infrastructure.

How much power has the program saved and what is the impact?

Meta’s Capacity Efficiency Program, powered by AI agents, has recovered hundreds of megawatts (MW) of power—enough to supply electricity to hundreds of thousands of American homes for a full year. This is a direct result of both offensive optimizations (e.g., reducing unnecessary computation) and faster defensive regression resolution. Before AI automation, each regression fix required hours of manual work; now, AI compresses that to about 30 minutes, meaning fewer wasted resources and quicker recovery. Moreover, the program has scaled MW delivery across a growing number of product areas without needing to proportionally expand the human team. For example, AI-assisted opportunity resolution now handles a volume of wins that engineers would never have time to address manually. The cumulative effect is a significant reduction in Meta’s overall energy footprint, contributing to its sustainability goals while maintaining performance for billions of users.

What is the long-term goal of this AI efficiency engine?

The ultimate vision for Meta’s Capacity Efficiency Program is to create a fully self-sustaining efficiency engine. In this future, AI agents will manage the entire lifecycle of performance optimization: from detecting regressions and opportunities, to diagnosing root causes, to deploying fixes—all without human intervention in routine cases. Engineers would only step in for novel or complex issues that the agents cannot handle. The AI platform is designed to continuously learn from new data and encoded expertise, evolving to tackle the long tail of performance issues that would otherwise be economically impractical to address manually. This approach allows Meta to maintain efficiency at hyperscale without linearly growing the efficiency team, freeing up engineers to innovate on new products and features. Ultimately, the program aims to deliver ever-increasing power savings while ensuring that user experience remains unaffected by performance regressions.