Engineering at Meta

What's your plan when the power goes out instantly, with zero warning?

For Meta's data centers, we needed an answer we could prove. That's why we built Instantaneous PowerLoss Storm, a new testing paradigm that simulates complete utility outages to validate our infrastructure can tolerate instant power loss.

In this post, our engineers walk through the defense-in-depth strategies, implementation tradeoffs, and validation approach that make this possible at scale.

Read more: https://engineering.fb.com/2026/06/03/data-center-engineering/lights-out-systems-on-validating-instant-power-loss-readiness/

Today we're announcing America's Workforce Academy (AWA), the largest private-sector commitment to the skilled trades in American history, beginning with a $115 million commitment in the first year alone.

This cost-free program supports all participants while they learn and then guarantees a job for all graduates in high-demand fields such as electrical work, mechanical systems and plumbing. Every graduate will be guaranteed a job on a Meta construction site.

Learn more: http://meta.com/AmericasWorkforceAcademy

For years, recommendation retrieval systems have relied on a complex mesh of microservices. But as scale and sophistication grew, so did the structural limits like network latency, version inconsistencies, and siloed development environments.

To break through these barriers, we’re introducing SilverTorch, a reimagined retrieval ecosystem built under a new paradigm: "Index as Model."

Instead of stitching neural networks into a microservice architecture, we’ve built the entire retrieval system as a single PyTorch neural network. Every component – the item index, eligibility filtering, scoring, and user tower – is now a tensor or operator inside one unified model.

A few highlights of the impact:
🔹 Up to 23.7x higher throughput than state-of-the-art approaches
🔹 20.9x more compute cost efficiency vs. CPU baselines
🔹 Unlocked capacity for complex neural reranking without breaking our strict sub-100ms latency budget
🔹 Greater engineering velocity by dissolving the boundaries between ML and infrastructure

Read the full breakdown of how we built SilverTorch and details from our SIGIR 2026 paper: https://engineering.fb.com/2026/05/26/ml-applications/silvertorch-index-as-model-new-retrieval-paradigm-recommendation-systems/

How do you migrate a data ingestion system that processes several petabytes daily without any downtime?

We just published a deep dive into how our team pulled this off. The system incrementally ingests up-to-date snapshots of Meta's social graph into the data warehouse for analytics, machine learning training, and product development across the company.

Key insight: The hardest problem wasn't just building the new architecture. It was also ensuring every single job could be migrated seamlessly while the system remained fully operational. That challenge was intensified by the sheer volume of jobs and the complexity of their logic. We addressed it with a rigorous 3-phase migration lifecycle, automated quality tooling, and clear migration planning.

Read the full technical deep dive on how our team pulled this off: engineering.fb.com/2026/05/12/data-infrastructure/migrating-data-ingestion-systems-at-meta-scale/

Good security should always be invisible. When Meta launched encrypted backups for Messenger in 2023, we set a new bar for end-to-end encrypted (E2EE) messaging at scale. With E2EE messaging backups your message history can travel with you across devices without ever being readable by any other party, including Meta.

Today, we’re evolving the work we began with Labyrinth, our protocol for end-to-end encrypting stored messages history between devices on your Messenger account. Labyrinth 1.1 improves backup reliability with a new sub-protocol that lets messages reach your encrypted backup as they’re sent, rather than waiting for your device to come back online, which is how Messenger encrypted backups currently work.

This means your messages are safe even if you lose your phone, switch devices, or have a long gap between sign-ins. Each message is wrapped with a message encryption key that the sender places directly into the recipient’s encrypted backup — like dropping a sealed envelope into a locked box only the recipient can open. No one but you and the people you’re talking to can read your messages – not even us.

We’re rolling Labyrinth 1.1 out broadly to Messenger and are already seeing meaningful gains with more messages successfully backed up and more people restoring their full message history when they change devices.

Read our updated white paper, “The Labyrinth Encrypted Message Storage Protocol” for more information: https://go.meta.me/4c30af

Today, we are announcing two landmark partnerships with Overview Energy and Noon Energy to deliver reliable, innovative power to our AI infrastructure and data centers.

🛰️ Harnessing Space Solar: We are partnering with Overview Energy to deploy up to 1 GW of orbit-to-grid solar energy to Earth. By utilizing satellites in geosynchronous orbit to beam near-infrared light to ground facilities, we can enable around-the-clock energy production from existing solar infrastructure.

🔋 Ultra-Long-Duration Storage: We have reserved up to 1 GW/100 GWh of energy storage capacity with Noon Energy. Their modular, reversible solid oxide fuel cell technology provides over 100 hours of storage, far surpassing the limitations of current lithium-ion systems.

These partnerships with Overview and Noon build on our approach to helping strengthen the grid through diverse and reliable solutions.

Learn more: https://go.meta.me/635755

We're breaking ground on a new AI-optimized data center in Tulsa, Oklahoma — our first in the state.

Once operational, this facility will be purpose-built for AI workloads. Here's what's going into it from an engineering standpoint:

💧 Water-efficient cooling: The facility uses a closed-loop, liquid-cooled system that recirculates the same water and requires zero water for server cooling during the majority of the year. We're also paying the full costs of water and wastewater service so those costs aren't passed on to consumers.

⚡ Grid investment + clean energy: We're matching 100% of the data center's energy use with clean energy and adding over 1,500 MW of clean energy to the Oklahoma grid. Through our utility bills, we'll pay hundreds of millions of dollars to fund grid infrastructure that benefit all consumers.

🎓 Technical workforce development: We're partnering with Tulsa Tech and Tulsa Community College on a cross-institutional program for digital infrastructure careers. The goal: 200+ graduates annually trained in cooling simulation, fiber optics, structured cabling, and AI/data analytics programming.

🌾 Water restoration at scale: A 10-year partnership with Phytech will deploy plant-sensor technology across approximately 1,500 acres of commodity crops — corn, soybeans, winter wheat — to save over 50 million gallons of water per year. Part of our goal to be water positive by 2030.

More on our Tulsa data center: https://go.meta.me/1bac0e

At Meta, real-time communication (RTC) powers everything from Messenger and Instagram video chats to Cloud Gaming and VR casting on Meta Quest. To meet the performance demands of billions of users, we spent years developing a specialized, high-performance variant of the open-source WebRTC library.

But permanently forking a large open-source project often leads to the "forking trap"—where internal features accumulate and merging external commits becomes prohibitive.

We recently concluded a massive multiyear migration to break this cycle, moving 50+ use cases from a divergent WebRTC fork to a modular architecture built on the latest upstream version.

To achieve this in a monorepo with strict static linking constraints, we engineered a dual-stack architecture. By building a shim layer—a proxy library with a unified, version-agnostic API—we ran both the legacy and latest WebRTC implementations simultaneously for safe A/B testing. We solved thousands of C++ One Definition Rule (ODR) symbol collisions using automated renamespacing and AST-based code generation, adding only ~5 MB to our binary size instead of the ~38 MB that duplicating higher-layer orchestration would have caused.

Instead of being years behind, we are now "living at head" with stable Chromium releases. This modernization dropped CPU usage by up to 10%, improved crash rates by up to 3%, and reduced compressed binary sizes by 100-200 KB across major apps.

Read the full technical breakdown on how we engineered our way out of the forking trap: https://engineering.fb.com/2026/04/09/developer-tools/escaping-the-fork-how-meta-modernized-webrtc-across-50-use-cases/