Using Hybrid Strategy to Achieve AI Capacity Agility - Top Lessons Learned.pdf

1、Using Hybrid Strategy to Achieve AI Capacity AgilityTop Lessons LearnedAI CLUSTERSJin Zhang,Technical Program Manager,MetaPolina Vasileva,Technical Program Manager,MetaAI Inflection PointThe advent of AI has changed all of our assumptions on how to scale our infrastructure BEFORE LLMS PREDICTABLE,LI

2、NEAR GROWTHAI scaled with users recommendation models had stable,well-understood compute needsLargest training jobs ran on 128 GPUs;clusters topped out at 4,000 GPUsAFTER LLMS EXPONENTIAL,UNBOUNDED DEMANDGPU demand grew 30 x in two years 4,000 to 129,000 GPUs in a single cluster Every prior assumpti

3、on about data centers,power,cooling,and network had to be rebuilt from scratchWere in the AI RaceMeta AI Demand Capacity FASTWere building tens of gigawatts this decade,and hundreds of gigawatts or more over time.Mark Zuckerberg,Meta Compute AnnouncementPrometheus1GW+in 2026Hyperion5GW over several

4、yearsHowever,Building New Data Centers Takes TimeMetas Hybrid StrategyNo single infrastructure model can keep pace with AI capacity demand1FOUNDATION LAYERSelf-BuildOwned gigawatt-scale campuses designed from the ground up for AI density.Lowest long-term cost per GPU-hour.2BRIDGE LAYERLeased&ColoBui

5、ld-to-suit facilities that bridge the gap between AI demand spikes and the multi-year lead times for construction.Faster to deploy,higher unitcost.3AGILITY LAYERCloud PartnershipsStrategic hyperscaler and NeoCloud agreements that provide immediate burst capacityHighest unit cost,fastest time-to-capa

6、city.How we engineer,invest,and partnerto build this infrastructure will become a strategic advantage.Top Lessons LearnedMETASHYBRIDMODEL5 Key Lessons1HARDWARE STRATEGYInfrastructure HeterogeneityManage diverse hybrid infrastructure as one unified fleet2VENDOR STRATEGYHyperscalers vs NeoCloudsPortfo