Extending Meta to the Public Cloud
summary
Meta moves GPU demand to public clouds in under three months, enabling first capacity in ~60 days using the offnet services boot-and-attest flow and transparent encryption via TTLS transparent TLS so on‑prem services require no code changes. Reliability comes from shipping fleet health and a PCM public cloud manager that folds cloud maintenance into on‑prem maintenance trains. Result: web runs from cloud (about 1% of employee traffic), new vendors onboard in <1 month, regions in ~1 week, and co-designed SKUs in ~2 weeks.
AI-written summary · about
chapters (10)
- 0:00Meta’s cloud journey
- 1:49Public cloud for GPUs
- 3:36Security challenges
- 6:24Reliability hurdles
- 8:10Generic cloud capacity
- 12:06Transparent TLS
- 14:34GPU maintenance in cloud
- 16:44Public Cloud Manager
- 18:31Co-designing cloud hardware
- 19:56Cloud strategy state
related talks
19mBuilding Meta's Internal Cloud: the opportunities and challenges
21mCapacity Planning for AI Training Hardware
37mHybrid OnPrem + Cloud Architectures
31mOn-Prem is the new Black
28mCloud The Final Destination
26mOpenStack in 0 to 45 seconds. Engineering at the speed of light
27mSupporting Security Sensitive Tenants in a Bare-Metal Cloud
23mSupporting AI holistically Across the Infra stack