Uploaded July 2026 | Updated November 2026, 2 days ago
Meta has a massive amount of internal infrastructure, but the demand for GPU compute led us to look to the cloud to grow our footprint. This led to our initial forays into cloud, and what it took to enable and integrate into Meta’s overall infrastructure footprint.
In order to create capacity hyperliquidity, we’ve leveraged multi-cloud as a form of optionality. Come learn how we’ve fully integrated cloud into our platform, OS, and fleet management stack to make the experience ubiquitous for our internal customers.
Learn more about the @Scale conferences here: atscaleconference.com
Meta has a massive amount of internal infrastructure, but the demand for GPU compute led us to look to the cloud to grow our footprint. This led to our initial forays into cloud, and what it took to enable and integrate into Meta’s overall infrastructure footprint.
In order to create capacity hyperliquidity, we’ve leveraged multi-cloud as a form of optionality. Come learn how we’ve fully integrated cloud into our platform, OS, and fleet management stack to make the experience ubiquitous for our internal customers.
Learn more about the @Scale conferences here: atscaleconference.com


![Unmixing the World: Making Sound as Programmable as Code | Jessica Powell, Audioshake
Sound is one of the world’s richest but least accessible data sources. Most audio exists as a single, tangled mix—impossible to edit for humans and noisy for machines to understand. This talk explores how advances in audio separation are unmixing the world: splitting sound into its component parts so it can be edited, searched, and programmed like code.
For humans, this means new creative control, immersive experiences, and copyright-safe workflows. For machines, it provides the fine-grained data needed to power multimodal training, boost speech recognition, and enable entirely new AI applications. By transforming raw sound into structured, usable data, separation is becoming the intelligence layer for audio—laying the foundation for the next generation of media, AI, and human-computer interaction.
Jessica Powell, the CEO and co-founder of Audio Shake, delivered a presentation on her companys audio source separation technology. [00:52]
She began by outlining the shared problem both humans and machines face in understanding and filtering complex, real-world sound. [01:33]
Powell explained that Audio Shakes technology addressed this by splitting audio into its component parts, such as separating a voice from noise or one instrument from another. She detailed the technical challenges of balancing processing speed with quality and asserted that Audio Shake outperformed open-source models on both fronts. [04:34]
The presentation then covered a wide range of applications, first for human-centric creative workflows in music, film, and broadcasting, and then for machine-centric uses like improving voice AI accuracy, enabling the editing of generative media, and creating higher-quality training data for multi-modal AI models. Unmixing the World: Making Sound as Programmable as Code | Jessica Powell, Audioshake](https://i.ytimg.com/vi/qfJwq-rznFg/mqdefault.jpg)


![Scaling Privacy Infrastructure for GenAI Product Innovation | Ming Qiao and Ram Rathnam, Meta
Generative AI is reshaping product experiences while introducing new privacy challenges. This talk explores how Meta’s Privacy-Aware Infrastructure (PAI) helps navigate this evolving landscape, protecting user data while enabling safe GenAI product innovation at scale, using Meta AI glasses as an example. We’ll highlight core PAI technologies such as data lineage and safeguards that embed privacy directly into the product development lifecycle, empowering product teams to move fast while delivering trusted experiences to Meta’s ~4bn users.
The speakers, Ram Ratnam and Ming, delivered a presentation on Metas approach to privacy protection for its Generative AI products, with a specific focus on the Meta AI glasses. [00:31]
They explained the unique privacy challenges posed by wearable AI technology and introduced Metas Privacy Aware Infra (PAI), a foundational system designed to embed privacy controls directly into the engineering stack. [08:17]
The talk detailed the PAIs four-phase playbook—Understand, Discover, Enforce, and Demonstrate—and highlighted the critical role of comprehensive, cross-stack data lineage in providing the visibility needed to enforce purpose limitations and ensure compliance at scale.
Ram Ratnam is a Technical Program Manager with the Meta Reality Labs Trust Team.
Ming Qiao is a Software Engineer on Metas privacy infrastructure team. Scaling Privacy Infrastructure for GenAI Product Innovation | Ming Qiao and Ram Rathnam, Meta](https://i.ytimg.com/vi/sXXZNRfHq2E/mqdefault.jpg)
![Ultra Low Latency Connect | Ishan Khot and Hani Atassi, Meta
Software engineers from Meta, Ishan and Hani, delivered a presentation on a new system they developed called Ultra Low Latency Connection (ULLC) to improve conversational AI. [01:37]
They began by identifying the awkward pause caused by connection latency in AI interactions, which degraded the user experience. [00:44]
After initial optimizations proved insufficient to break the one-second connection barrier, they redesigned the systems architecture. The core of ULLC involved parallelizing the signaling and media connection processes, which traditionally ran sequentially, thereby eliminating a full round-trip time. [08:56]
They detailed the technical challenges, such as ensuring data streams were routed to the same data center, and the solutions they implemented, including a latency mapping system and a reliable fallback path. The presentation concluded by showcasing the results of ULLC in production—faster connection times, increased user engagement, and higher call success rates—and exploring future applications for the technology in edge inferencing, media processing, and wearables like the Ray-Ban Meta AI glasses. [14:56] Ultra Low Latency Connect | Ishan Khot and Hani Atassi, Meta](https://i.ytimg.com/vi/sYtAJD8NVU0/mqdefault.jpg)



