Exponential View Puts AI's Demand Side at $110 Billion

Exponential View estimates generative AI demand-side revenue at $110 billion over the past 12 months, with a recent monthly run rate above $175 billion, based on a deduplicated model that counts end-customer spending. The report aims to assess whether AI revenue is catching up with infrastructure spending, finding that hyperscaler AI revenue just covers depreciation on AI capex.

Azeem Azhar https://www.azeemazhar.com/about and the Exponential View research team put a hard number on the demand side of the generative AI boom: $110 billion in deduplicated sales over the past 12 months, with the most recent month implying a revenue run rate above $175 billion. The estimate, published June 25 in Exponential View's State of the AI Economy https://www.exponentialview.co/p/the-state-of-the-ai-economy essay and accompanying June 2026 report https://intelligence.exponentialview.co/ , is an attempt to answer the question the market has mostly inferred from chip orders, cloud capex and private-company leaks: how much are customers actually paying for AI? Azhar is not a neutral bystander to the AI cycle. He founded Exponential View https://www.exponentialview.co/p/the-state-of-the-ai-economy in 2015 after earlier work as a technology correspondent, operator and founder, and his public biography describes him as an investor in more than 30 technology startups. That background matters because the report is aimed less at whether AI is useful than whether the current buildout has a revenue base large enough to support it. The report's central move is methodological. Exponential View says it counts the dollar spent by an end customer and avoids counting the same dollar again when it flows through the supply chain. Its example: if a customer spends $1 on Claude and Anthropic spends 50 cents with Amazon to serve that customer, the reported demand-side number is $1, not $1.50. That distinction is the whole fight. The supply side of AI is visible because much of it runs through public companies selling chips, memory, networking gear, power equipment, cooling systems and data center capacity. The demand side is harder because many of the companies collecting AI revenue, including OpenAI, Anthropic, Cursor and ElevenLabs, are private and do not publish audited segment revenue. Even the public hyperscalers, including Amazon, Google and Microsoft /article/we-put-ideogram-4-head-to-head-against-openai-google-and-microsoft-in-four-image , do not consistently break out AI revenue as a separate line item. Exponential View says its model uses public statements from hyperscalers and neoclouds, supplier and customer disclosures, reported leaks and company self-reports, each weighted for confidence. The result, according to the authors, is an item-by-item financial model for major companies and business units rather than a top-down market-size estimate. The caveats are as important as the headline number. The $110 billion estimate does not include internal AI-driven uplift at Meta or Google, such as recommendation systems improving ad revenue. It does not include efficiency gains from internal AI tools at large technology companies. It also excludes professional services and systems integration, meaning a Fortune 500 company's broader AI program may be larger than the portion counted as revenue to an AI vendor. China is outside the first version of the model, even though Exponential View says it has modeled Chinese revenue separately. Those exclusions make the number narrower than the total economic impact of AI, but cleaner as a measure of customer spending. They also make the report useful as a check on one of the market's most persistent questions: whether AI revenue is catching up with AI infrastructure spending. On that point, Exponential View says AI-attributable hyperscaler revenue "just about" clears the depreciation expense tied to AI infrastructure. The report separates AI-oriented capital expenditure from ordinary hyperscaler capex, noting that Google, Microsoft and Amazon were already spending around $120 billion annually on capex before ChatGPT, with Amazon's number including logistics investments and the comparison excluding Meta. The depreciation assumptions drive the conclusion. Exponential View says it depreciates compute assets over six years and other infrastructure over 14 years. A six-year useful life for AI compute is not a settled accounting truth; it is a bet that demand will remain high enough, and fleet management will improve enough, to keep GPUs economically productive for longer than skeptics assume. If useful lives shorten because hardware is displaced faster or utilization falls, the payback picture deteriorates. The report also argues that token price declines do not necessarily shrink the market. Exponential View estimates that a 10% cut in token prices leads to 12% to 18% more tokens used across providers, enough for total spend to rise. That is the case investors want to believe: cheaper inference expands usage faster than it compresses revenue. But the report is careful about what a token can and cannot measure. The authors argue that a token remains a billing unit rather than a true unit of economic value. Their proposed alternative is quality-adjusted output tokens, which would account for visible output and model capability, not just raw token volume. External survey data supports the report's picture of broad adoption without full enterprise maturity. McKinsey's 2025 State of AI survey found that most organizations remained in experimentation or pilot mode and that nearly two-thirds had not begun scaling AI across the enterprise. It also found that 88% of respondents reported regular AI use in at least one business function, up from 78% a year earlier, but only a minority reported enterprise-level EBIT impact. That gap between adoption and bottom-line transformation is where the report's $110 billion number lands. It is large enough to make generative AI a real software and cloud market, not just a capex story. It is still small relative to the amount of capital being committed to data centers, chips and power, especially if demand growth slows or infrastructure costs prove less flexible than modeled. The report's strongest contribution is not the precision of the estimate, which depends on private-company self-reports and leaks that outsiders cannot fully audit. It is the framing: AI's economic test is moving from supply scarcity to demand durability. The chip shortage created the first phase of the boom. The next phase depends on whether enterprises and consumers keep increasing paid usage as model prices fall, products converge and CFOs start asking which AI bills are actually tied to revenue, margin or labor substitution. Azhar's team is effectively arguing that the answer is already measurable, even if imperfectly. The AI economy, in their model, is no longer a promise waiting for future monetization. It is a $110 billion trailing revenue pool running above $175 billion annualized, with a customer-spending base big enough to begin testing whether the infrastructure buildout is rational. The unanswered question is whether that base compounds fast enough. If token prices keep falling and usage grows faster, AI revenue can keep expanding while products become cheaper to use. If usage elasticity weakens, the market will be left with the harder math: expensive compute, fast-moving models and revenue that looks real but not yet large enough to justify every data center being planned around it.