驾驭人工智能热潮企业人工智能的采用和影响

Navigating the Boom:Confronting Generative AI’sMost Pressing Questions 2Jason Ader, CFA +1 617 235 7519Contents Key Conclusions.................................................................................................................3Executive Summary...........................................................................................................4Generative AI Frequently Asked QuestionsAre We in an AI Bubble?...............................................................................................7Is There Enough of a Return on Current Investments in AI Models and DataCenters to Justify the Spending?.........................................................................14Where Is the Value Accruing in the AI Space and Where Is the MostUntapped Opportunity for Investors?.................................................................15Are Scaling “Laws” Holding or Are LLMs Reaching DiminishingMarginal Returns?...............................................................................................17Is the LLM Becoming a Commodity?..........................................................................21Will Open-Source LLMs Diminish the Value of Closed-Source LLMs?.........................22Is There Room in the Market for Both LLMs and SLMs?.............................................24What’s the Killer App for GenAI or Is That the Wrong Question?..............................25How Is AI Being Monetized at the Application Layer?...............................................35What Is Agentic AI, How Does it Differ From AI Copilots, and WhatDoes it Mean for Enterprise AI Adoption?...........................................................38What Are the Main Barriers to Enterprise GenAI Adoption?.....................................41What Are the Main Physical Bottlenecks in the GenAI Buildout?...............................43Where Do LLMs Fit Into the Application Landscape?.................................................51Is AI a Threat to the Software Industry and/or Software Business Models?.............52How Necessary Are Nvidia GPUs Once the Heavy Lifting of TrainingModels Is Complete?............................................................................................53How Will AI Impact the IT Services Industry?...........................................................54Will Government Regulation Hold up the AI Market?................................................55When Is AGI Coming and Are We All Doomed?...........................................................57Our Best Ideas to Play the AI Theme...............................................................................58 Key Conclusions1.No Sign of AI Capex Letting Up. We expect sizable GenAI-related capex to persist through theend of the decade as model intelligence is still rapidly improving and well-capitalized hyper-scalers are engaged in an AI “arms race.”2.Test-Time Compute Is Next Frontier in AI Scaling. New vectors of AI model improvementbeyond pretraining should continue to support scaling “laws” and drive the need for ever-increasing computing power. In particular, test-time compute (used in reasoning models likeOpenAI’s o1 family) represents the next frontier in AI scaling/intelligence given its ability dur-ing the inference phase to generate multiple potential solutions, evaluate them, and select theoptimal one.3.Still Scratching the Surface on AI Use-Cases.Initial GenAI use-cases include customerservice, web search, software development, IT service management, content creation, andadvertising. Use-cases should expand rapidly with improving model performance, reliabil-ity, and increasing enterprise/consumer familiarity, but return on GenAI investment (i.e.,monetization of AI apps) will need to be proven out in the next few years to protect againstoverbuild risk.4.Physical Power and Infrastructure Are Main AI Bottlenecks.The greatest constraint on theAI buildout today is the physical infrastructure, such as power and data centers, not chips andtechnology. In a bull scenario, we estimate AI could cause U.S. data center electricity demandto inflect up to 15% growth annually (up from 1% over the past decade), consuming 500 TWhof electricity by 2030, accounting for 9% of total U.S. electricity consumption in that year (upfrom ~2% today). When combined with the “electrify everything” movement, reshoring, andautomation, AI is projected to more than double domestic electricity load growth annuallyover the next decade.5.Natural Gas Is Best Positioned for AI Data Center Energy Demands.We see natural gasas the greatest near-term winner from the AI boom given its flexibility to meet electricitydemand throughout the day, while nuclear and battery storage are likely to receive signifi-cant investment going forward to support the massive power demands of next-generationAI data centers.6.Enterprise Adoption Starting to Roll, but Still Early Days.We expect to see signs of a steadyincrease in enterprise