Find this research useful? Add aibubblequestion.com to your Google Preferred Sources, one click, affects only your own results.

The AI Bubble Question
All pages

The Evidence

The Demand Question

Everything on this site reduces to one question: is enough real, independent demand showing up, fast enough, to pay for the build-out? Here is the evidence, with facts separated from estimates.

Updated

The revenue, company by company

Microsoft (disclosed): its AI business passed a $37 billion [annualized run rate](/glossary#arr), up 123% year over year, in the quarter ended March 2026, nearly triple the $13 billion rate of January 2025 (GeekWire). A management metric, not an audited segment, but the closest thing to hard evidence that AI sells.

OpenAI (leaked/reported): $13.07 billion of revenue in 2025, against about $34.0 billion of costs, a $20.92 billion operating loss, and a $38.53 billion loss attributable to OpenAI, per leaked audited financials corroborated by the Financial Times (Runtime Wire); run rate has since surpassed $40 billion annualized as of mid-August 2026, roughly doubling its late-2025 pace, reported via an internal staff communication, not an audited or public figure, superseding the ~$25 billion February 2026 snapshot this page previously cited (Bloomberg, 13 Aug 2026). Explosive growth and enormous losses, simultaneously.

Anthropic (estimates): from $316 million ARR in March 2024 to $1.4 billion a year later to $9 billion at the end of 2025 to $47 billion in May 2026 to $65 billion at the end of July 2026 (TechCrunch), each figure from third-party reporting or estimation, never an Anthropic statement, filing, or audited document; Anthropic declined to comment when last asked contested. Treat every figure in this chain with real caution, especially the newest: the direction (hyper-growth) is well supported, the magnitude is not.

Meta (disclosed, indirect): Q1 2026 revenue of $56.3 billion (+33% YoY); Q2 2026 revenue $60.8 billion (+28% YoY), advertising revenue $59.4 billion (+27% YoY) (Meta Q2 2026 8-K, Ex. 99.1). Management's figures for advertiser adoption (8 million in Q1, 9 million in Q2) and a reported >$75 billion Advantage+ run rate do not appear in either filed exhibit; they are earnings-call disclosures, not filed figures. Q2 operating margin fell to 31% from 43% a year earlier (operating income down 8% to $18.8 billion), partly on $2.40 billion of legal charges and $1.18 billion of severance, a complication for a pure-upside "AI ad tools" framing. Real monetization, permanently entangled with the core ads business; no standalone AI line. Google similarly cites AI as a Cloud growth driver without breaking out a number.

Add it up: no named, dated report with disclosed methodology was located behind a $140 billion global generative-AI-revenue figure for 2026; estimates in that space vary by more than double depending on scope and horizon, so the number is dropped here rather than restated without a source (see building-ahead-of-demand for the matching scope problem on the capex side). What the companies above disclose directly is in the tens of billions, not hundreds, against hyperscaler capex heading past $600-750 billion a year. The revenue is real and compounding at rates almost nothing in business history matches. It is also several times smaller than the spending, and part of it originates inside the ecosystem.

Enterprise adoption: production or pilots?

Four figures, four surveys, four scopes, not one coherent picture. Enterprise gen-AI spending tripling from $11.5 billion (2024) to $37 billion (2025) is Menlo Ventures' own venture-capital estimate, not company-reported revenue. The 24%-scaled-production/48%-in-pilot split is from Mochikabu's survey of DACH-region regulated enterprises, not a global sample. 37% of organizations, not 39%, report at least some EBIT impact from AI use, per McKinsey's global 2026 survey (1,719 respondents, 97 nations, fielded 4 May to 8 June 2026, published 25 Aug 2026); McKinsey itself calls this share "about the same" as the year before, so the shift from 39% looks like survey noise rather than a real move, but 37% is the current published figure. The roughly 44% year-one-ROI figure has no locatable primary report behind its usual "McKinsey and Forrester, 2025" attribution and is dropped. Verdict: acceleration, not pullback, with the profits still concentrated and mostly prospective.

Inference costs: collapsing, which cuts both ways

The cost of running AI modelsRunning a trained AI model to answer questions or generate text. Training builds the model once; inference is the ongoing cost of using it, every single time.Full definition in the glossary, for fixed capability, fell roughly 10x a year: GPT-3.5-class output dropped from $20 per million tokens (late 2022) to $0.07 (late 2024), about 280x in two years (Stanford HAI, AI Index Report 2025). The additional claim of a further 30-60% decline from mid-2025 to spring 2026 has no locatable named-model-basket or methodology behind it and is dropped. For the bulls, collapsing costs are how AI becomes ubiquitous, the semiconductor story again. For the bears, they mean the chips bought at peak prices earn less per hour every quarter, and customers increasingly self-host on cheaper hardware. Both readings are true; they differ on who captures the falling cost.

Utilization: how busy are the chips?

Outside the hyperscalersOne of the giant cloud-computing companies that run global fleets of data centers: Microsoft, Amazon, Google, Meta, and (increasingly) Oracle.Full definition in the glossary, utilization is measured only narrowly: Cast AI's own sample of roughly 23,000 Kubernetes clusters it monitors found average GPU utilization near 5% (Cast AI), its own customer base, not a fleet-wide or hyperscaler-wide measurement. A widely repeated "86% of enterprises at or below 50% utilization" figure has no locatable source of record across two independent search passes and is dropped. Hyperscaler fleets almost certainly run hotter by pooling many customers, but no audited utilization numbers exist. The nearest thing to a "dark fiber" measurement for AI, and it points the same direction, with the same caveat: fiber was dark until it wasn't.

The GPU depreciation controversy

This is the demand question wearing an accounting costume. What the companies did is documented in filings, and it is not a uniform "3 years to 5-6" move: Microsoft raised server and network equipment life from 4 to 6 years, effective FY2023. Alphabet raised servers from 4 to 6 years and certain network equipment from 5 to 6, also FY2023. Meta set most servers and network assets to 5.5 years, effective January 2025. Amazon shortened a subset of servers and networking equipment from 6 to 5 years in 2025, citing the pace of AI-driven hardware change, the opposite direction from the others, and starting from 6 years, not 3. Oracle never changed from 5. Longer assumed life means less depreciationSpreading the cost of an asset over its useful life in the accounts. Stretch the assumed life and reported profits go up, with no change in the business.Full definition in the glossary charged each year, which means higher reported profit from identical hardware. A derived analyst model (TheCUBE Research, echoing figures associated with Michael Burry) puts the aggregate impact near $18 billion less depreciation in 2024 and a cumulative $176 billion by 2028, flattering earnings 21 to 27% versus unchanged assumptions, but the companies' own disclosed effects, in the years they actually reported them, do not sum to anything resembling that: Microsoft +$3.7 billion operating income (FY2023), Alphabet -$3.9 billion depreciation / +$3.0 billion net income (FY2023), Meta -$2.92 billion depreciation / +$2.59 billion net income (FY2025), and Amazon's FY2025 move running the other way (+$1.4 billion more depreciation, -$1.0 billion less net income). Read the aggregate as a model, not as a sum of the filings.

The critique, pressed most famously by Michael Burry: the extensions overstate the economic life of AI hardware. Chips can be physically running and still be economically obsolete against newer, more efficient parts, and if true end demand is as thin as he argues, the accounting life dramatically exceeds the earning life (Business Insider).

The defense: the life extensions mostly predate the AI boom (2020-2022), were disclosed in audited filings, and reflect real engineering improvements. Nvidia's response memo states that A100 chips shipped six years ago are "still running at full utilization," and that customers' 4-6 year schedules reflect observed longevity (CNBC). Analysts at TheCUBE and Tsai Capital add that no net extensions happened during the boom itself.

What the market data says: ServerBuyback's dealer-quoted pricing shows used H100s retaining roughly 75-85% of value after 24 months as of mid-2026, with prior-generation prices down 10-20% specifically against Blackwell/B200 availability, dealer-quoted estimates, not transaction-level or audited data, and not independently reconciled with the accounting-life figures above (ServerBuyback, June 2026). Older GPUs demonstrably retain value and see real use; they also demonstrably lose value on a clock measured in a few years, not decades. Both sides of the controversy can point at the same chart. Why this matters so much is laid out on Building Ahead of Demand: the shorter the true asset life, the smaller the window for demand to catch up. How this revenue and cost picture weighs into the overall scorecard is on Conclusions.