3418 - 在云端构建和管理 GenAI.pdf

《3418 - 在云端构建和管理 GenAI.pdf》由会员分享，可在线阅读，更多相关《3418 - 在云端构建和管理 GenAI.pdf（22页珍藏版）》请在三个皮匠报告上搜索。

1、Orlando,FLOctober 69IBM TechXchange 20253418Phil AlfanoField CTOApptio,an IBM CompanyPhilip.ABuilding and Managing GenAI in the CloudAgenda1234Complexities of GenAI in the CloudHow IBM Cloudability HelpsGenAI Cost OptimizationGenAI Allocation&Unit Economics IBM TechXchange|2025 IBM CorporationSignif

2、icance of GenAI Running on Public Cloud“Generative AI could constitute$200 billion to$300 billion of cloud spending by 2030,as investment moves beyond mega technology companies and foundation model providers”Goldman SachsState of FinOps 2025:Top priorities next 12 monthsThe Workhorse of GenAI:GPU-ba

3、cked VMsPopular ML ModelsPopular ML ModelsJambaAmazon NovaAmazon TitanGPT-4.1o3o4-miniGemini 2.5Gemma 3Popular VM FamiliesPopular VM FamiliesAWS:P,G Azure:NC,ND,NG,NVGCP:A4,A3,A2,G2Generative AI Deployment ModesChatbot ServicesChatbot ServicesMonthly subscription($20-250/month)Examples:ChatGPT Plus,

4、Claude ProPredictable costs,no infrastructureLimited customization,usage capsSaaS APIsSaaS APIsPay per token($0.0005-$0.02/1K)Examples:OpenAI API,Anthropic APIPay for actual usage,flexibleUnpredictable costs,monitoringManaged ServicesManaged ServicesVariable pricing based on usage including per toke

5、nExamples:AWS Bedrock,Azure OpenAIEnterprise security,scalingPlatform lock-in,complex pricingDIY DeploymentDIY DeploymentHigh upfront costs($10,000+)Self-hosted open-source modelsComplete control,data privacyTechnical expertise requiredHow are you charged?Operation:Operation:You are charged for“infe

6、rence”generating outputs based on input data Charged by token(e.g.,$0.003 per 1k tokens)Important:includes input and output tokensTraining/Building models:Training/Building models:charged for VM and storage usage,or hourly charge as set by managed serviceIBM TechXchange|2025 IBM Corporation6What is