Han_开源AI工厂软件：NVIDIA云合作伙伴如何构建与规模化运营GPU基础设施.pdf

1、Han VanholderOpportunities for providers offering differentiated infrastructure and servicesBare metal remains essential for frontier training,large fine-tuning,and sovereign AI workloadsPerformance:Virtualized vs.Bare MetalX X1.3X1.3XVirtualizedBare Metal+30%performance advantageNeocloud providers

2、are scaling rapidly,offering GPU-as-a-Service at unprecedented growth ratesNeocloud Platform SignalX X6.2X6.2X2024 Baseline2025622%YoY growth Netris ARR,2025)Kubernetes is the preferred orchestration layer for production AI workloadsContainer Users Running K8s in Production66%66%82%82%20232025+16pp

3、in two years infrastructure convergenceSources:ML Trends(epoch.ai),McKinsey,Business Wire(Netris)Bare metal is not going awaybut providers are evolving toAI native servicesBy 2030,inference will leadin enterprise spending anddata center demandToken generation andutilization are explodingThe Economic

4、s of AI are Tilting Towards Inference EfficienciesWIPOpportunities for providers offering differentiated infrastructure and servicesBare metal is not going awaybut providers are evolving toAI native servicesBy 2030,inference will leadin enterprise spending anddata center demandToken generation andut

5、ilization are explodingInference will comprise most of the spending by 2030GPUaaS Market:Training vs.Inference($B)010203040202520262027202820292030TrainingInferenceDatacenter demand:Inference will be the dominant workload by 20300501001502002502025 2026 2027 2028 2029 2030CAGR 25-3022%Sources:Analys

6、ys Mason,McKinseyBare metal is not going awaybut providers are evolving toAI native servicesToken generation andutilization are explodingGlobal data center demand by workload,2025-30(gigawatts)Training 22%Inference35%Non-AI11%The Economics of AI are Tilting Towards Inference EfficienciesWIPOpportuni