AI Capacity Planning at Scale: Meta's Strategy | WEKA

Описание: How does Meta balance rapid AI innovation with long hardware procurement cycles?
Elisa Chen, Data Scientist on Meta's AI Infrastructure team, sits down with WEKA's Chief AI Officer Val Bercovici to discuss the critical challenges of AI capacity planning at hyperscale.
In this fireside chat from AI Infra Summit 2025, Elisa reveals the gap between AI model iteration cycles—which can happen monthly or even weekly—and hardware procurement timelines that can take months to complete. Meta's approach focuses on three key levers: elastic resource allocation through GPU-as-a-service, efficiency optimization by matching specific hardware to workloads (H100s for foundation model training vs. A100s for fine-tuning), and dynamic quota allocation across teams to maximize utilization.

Chapters:
00:00 - What is the gap between AI innovation and hardware procurement?
01:17 - How do you predict future AI capacity needs?
03:02 - What are the levers you can use for immediate capacity needs?
05:07 - How does GPU elasticity differ from CPU cloud elasticity?
06:43 - How does user metadata inform capacity planning, including in specific regions?
08:43 - What alternative energy sources are emerging for data centers?

👉 Learn how the global memory shortage exacerbates already lengthy hardware procurement timelines: https://www.weka.io/blog/ai-ml/the-me...

What is the biggest challenge in AI capacity planning?
Organizations struggle without proper data foundations, telemetry, and instrumentation to measure capacity ROI. Elisa and Val explore hardware-workload matching, disaggregated prefill and decode architectures, and why 95% GPU utilization remains an aspirational target for most companies.

👉 Learn why storage architecture is the new bottleneck impacting scale and capacity: https://www.weka.io/blog/ai-ml/why-st...

Why does regional capacity planning matter?
Elisa explains how data privacy regulations like GDPR, regional usage patterns (for example, peak WhatsApp usage in India during business hours), and jurisdiction-specific policies make capacity planning highly localized rather than globally homogeneous. The discussion also covers energy as the ultimate bottleneck, with token-per-watt efficiency becoming the critical benchmark for AI infrastructure at gigawatt scale.

Key topics covered:
• GPU capacity planning strategies
• Elastic resource allocation
• Hardware procurement cycles
• Workload-specific GPU selection
• Dynamic quota management
• Regional compliance requirements
• Energy efficiency optimization
• Disaggregated inference architectures
• Observability for AI workloads
• Capacity ROI measurement frameworks.

Whether you're managing AI infrastructure at enterprise scale or planning your organization's GPU strategy, this conversation provides actionable insights from one of the world's largest AI deployments.

👉 Hear more from Elisa during a panel discussion with other AI leaders hosted by WEKA CMO Lauren Vaccarello: https://www.weka.io/resources/video/w...

About the Speaker: Elisa Chen is a data scientist on Meta’s Ads AI infrastructure team, focusing on the intersection of AI resource management and large-scale recommendation systems to optimize infrastructure to serve high-performing ads models efficiently.

👉 Connect with WEKA:
Website: https://www.weka.io?utm_source=youtub...
LinkedIn: https://www.linkedin.com/company/weka...
X: https://x.com/weka?utm_source=youtube...