转向代币经济学：驾驭AI的新支出动态

A note from the authors: AI economics affect mostorganizations and the C-suite uniquely. This paper guides those familiarwith AI tokens in making strategicchoices. Traditional total-cost-of-ownership frameworksthe realityof AI Volatile workloads, newinfrastructure demands,and tokens as the algorithmic complexity, and infrastructure intensity. What exactlyare the thresholds to move across different consumption choices? It depends on the organization. Roughly a quarter of respondentsin a Deloitte 2025 survey4of data center and power executivessay they or their clients are ready to make the move off of cloudto alternatives as soon as costs reach just 26% to 50% of thosealternatives, showing high sensitivity to even modest price changes, while others plan to wait until cloud costs exceed 150% of thecost of alternatives. The decision point remains unclear given the Across industries, Generative AI (GenAI) has become the fastest-growing line item in most corporate technology budgets—alreadyconsuming up to half of IT spend in some firms.1Cloud bills are rising nearly 20% year over year, driven by AI workloads.2At the same time,geopolitical uncertainties are intensifying calls for data sovereignty and technology infrastructure independence, making manyenterprises think about AI sovereignty and gaining greater control Traditional total cost of ownership (TCO) approaches are no longerthe best way to manage AI economics. Leaders may be better servedby precision economics—the ability to track, predict, and optimize Unlike prior technology waves governed by licenses or virtualmachines, AI spend often scales in nonlinear and unpredictableways. AI capabilities run ontokens: small chunks of data—text,image or audio—that AI systems process in training, inference,and reasoning. Every AI interaction consumes tokens, and every The competitive divide will not likely hinge on who adopts AI first, buton who manages its cost structure with discipline. AI spend will likelyseparate value creators from value eroders. The former converttokens into measurable enterprise output; the latter accumulate The complexity of AI’s economics hides within these tokens.Costs rise not only with user adoption but with workload design, The elusive AI ROI Despite rising investment, many leaders appear to stillbe chasing measurable return on investment (ROI) from •Nearly half (45%) of 500 leaderssurveyed in Deloitte’s 2025 USTech Value surveyexpect it will takeup to three years to see return on •Six in 10 of those completing Deloitte’s2025 Tech Value survey believe moreadvanced AI automation will take even •Of the 1,326 global finance leaderssurveyed for Deloitte Global’s inauguralFinance Trends report,fielded May2025, 28% said AI investments are But the issue isn’t whether AI will deliver value—it’show to measure and manage that value in a way ROIframeworks cannot. For many organizations, adoptingAI is no longer optional; it’s a strategic response to That makes understandingtheeconomics of AI —how costs, workloadsand returns flow throughtokens—thenew Tokens:The newcurrency of AI Unlike traditional pricing based on compute time—which is relatively static—token-basedpricing ties cost directly to the actual work AI performs.Each token represents botha unit of computation and a unit of cost.In that sense, tokens are thetrue currencyof AI economics—as indispensable to machine intelligence as kilowatt hours are to •Nonlinear demand:Complex reasoning models improve performancebut can consume more tokens than simple inference tasks. •Fluctuating token use:Token use fluctuates with experimentationlevels, workload design, model choice and even prompt engineering. •Varying pricing:Token price keeps changing based on AI model While this volatility appears to stem from usage patterns, its roots are in the tech stack.The compute, storage, and networking decisions that power AI models determine how A token is not just a technical measure—it is an economic signal. Each token carriesthe compound effect of GPU design, storage, throughput, network latency, and facilityeconomics. The discipline lies in tracing lineage—from infrastructure to the AI model to How tokens are bought AI spending is not a single market; it fractures into different economic realities dependingon how organizations consume intelligence. Some leaders experience AI costs only as asoftware-as-a-service (SaaS) line item, others as metered application programminginterface (API) calls, and a growing group/cohort manage it directly through infrastructure Buying patterns •Generating through packaged softwareabstracts tokens almost entirely.Leaders see a predictable subscription or per-seat fee, but little transparency •Consuming through APIsmakes tokens explicit. Every query is metered,billed, and exposed. This brings transparency, but also volatility: Costs risebased on workload design, prompt length, and hidden choices of infrastructure •Running on owned infrastructure