There is a number that should keep every AI investor awake at night, and it is not the one they expect.
The consensus centers on the spending-to-revenue gap: hyperscalers will deploy roughly $660–690 billion in capital expenditure in 2026, while the direct AI revenue generated by those investments is around $51 billion. That is a 10:1 ratio. When cloud computing was at the same stage of its adoption curve in 2011, the ratio was 2.4:1. The gap is real, the concern is valid, and the conclusion most people draw from it — that we are in a bubble — is both correct and catastrophically incomplete.
The deeper truth is structural, and it hides in a line item that almost no one outside accounting departments thinks about: depreciation.
The Hidden Variable
Here is the math that matters. The five largest hyperscalers — Amazon, Alphabet, Microsoft, Meta, and Oracle — collectively spend approximately 75% of their capex on AI-specific infrastructure: GPUs, networking gear, cooling systems, and the data centers that house them. That is roughly $450 billion flowing into AI hardware in a single year. These companies currently depreciate their servers over five to six years, assuming the hardware will generate economic value throughout that period.
Michael Burry, the investor who predicted the 2008 housing crisis, argues the real useful life of these chips is two to three years. If he is right, the industry is understating its depreciation by an estimated $176 billion between 2026 and 2028. That is not a rounding error. It is the difference between profitability and a writedown cycle that would reverberate through the entire technology sector.
But Burry's analysis, while directionally provocative, misses a crucial nuance. Bernstein's analysts have shown that the cash operating cost of running an older GPU — electricity, colocation, maintenance — is an order of magnitude lower than the rental price the market will pay for that compute. An Nvidia A100 chip, now five years old, still commands hourly revenue of approximately $0.93 against cash costs of $0.28. The contribution margin remains above 70%. The chip is not economically dead. It is merely less powerful than its successors.
This creates a paradox. The hardware simultaneously depreciates in capability and retains value in a supply-constrained market. As long as demand for compute exceeds supply — and every hyperscaler reports that it does — even last-generation silicon earns its keep. CoreWeave's CEO has noted that when a batch of H100 chips came off contract, they were immediately rebooked at 95% of their original price.
There is a fair counterargument here. GPUs do not become useless overnight. They cascade down through tiers of compute intensity: frontier chips train the next foundation model, last-generation silicon handles enterprise fine-tuning, and older hardware runs inference for chatbots, document processing, and recommendation engines. Most AI workloads do not require frontier compute. The secondary market for older chips is real and currently robust. But — and this is the part the optimists elide — that cascade only functions in a supply-constrained world. The secondary market for A100s is healthy because there are not enough GPUs to go around. If $450 billion in new silicon enters the market every year, supply eventually catches up with demand. And when it does, the cascade does not gracefully unwind. It collapses simultaneously at every tier.
The real risk, then, is not that the chips wear out too fast. It is that the market flips from supply-constrained to demand-constrained. The moment there are more GPUs than workloads to run on them, the economic assumptions underpinning $600 billion in annual spending collapse overnight. Five-year-old chips that were generating 70% margins become stranded assets that nobody wants at any price. Jensen Huang himself, with characteristic bluntness, has hinted at this dynamic. When Nvidia launched Blackwell, he joked about the predecessor architecture: "You couldn't give Hoppers away."
The Circular Financing Problem
The depreciation question cannot be separated from a second structural issue that few mainstream analyses connect to it: the circular financing architecture that now undergirds the entire AI economy.
Consider the web. Nvidia invests $100 billion in OpenAI. OpenAI uses that capital to build data centers filled with Nvidia chips. OpenAI commits $300 billion in cloud infrastructure deals with Oracle. Oracle uses that commitment to justify tens of billions in Nvidia chip purchases. Microsoft, which owns 27% of OpenAI and accounts for nearly 20% of Nvidia's revenue, provides Azure infrastructure to OpenAI and is CoreWeave's largest customer. CoreWeave, in which Nvidia holds a significant equity stake, issues billions in debt to expand capacity, with Nvidia guaranteeing to absorb any unsold compute through 2032.
Anthropic's CEO Dario Amodei has defended these arrangements as rational, given the capital requirements of building frontier AI. And he is not wrong. The technology is extraordinarily expensive, supply is genuinely constrained, and the arrangements do solve a real coordination problem. But the distinction between a virtuous cycle and a house of cards is whether real end-user demand exists outside the loop. And that is where the data gets uncomfortable.
According to Deloitte's 2026 State of AI survey, only 20% of enterprises report that AI is currently driving revenue growth. 74% say it remains an aspiration. Two-thirds of organizations remain stuck in the pilot phase, having not begun scaling AI across their operations. An MIT analysis found that 95% of companies see zero return on their generative AI investments. These are not numbers that justify $660 billion in annual infrastructure spending.
The Jevons Counterargument
The bull case has a name, and it is 161 years old. The Jevons Paradox, first articulated by William Stanley Jevons in 1865, holds that when the efficiency of resource use improves, total consumption of that resource increases rather than decreases. Cheaper coal did not reduce coal demand; it made coal economically viable for applications that previously could not afford it, and total consumption exploded.
Applied to AI, the argument runs as follows: as inference costs decline, and they are declining precipitously, with DeepSeek R1 offering performance comparable to frontier models at a fraction of the cost, the addressable market for AI compute expands dramatically. Services that were uneconomical at $15 per million tokens become transformative at $0.55. 80% of Americans who need a lawyer cannot afford one. Most small businesses cannot access financial planning. Most students cannot afford tutoring. As inference gets cheaper, these are not aspirational markets. They are inevitable markets.
The data support this. Enterprise spending on generative AI surged from $1.7 billion in 2023 to $37 billion in 2025, capturing 6% of the global SaaS market and growing faster than any software category in history. AI-native startups now capture nearly $2 for every $1 earned by incumbents at the application layer — 63% of the market, up from 36% the prior year. Anthropic's revenue run rate surpassed $9 billion in January 2026, up from roughly $1 billion at the end of 2024. Claude Code alone reached $400 million in ARR by mid-2025, growing from $17.5 million in April.
The Jevons case is compelling, but it rests on a critical assumption: that demand expansion occurs fast enough to justify the infrastructure build-ahead. And this is where the analysis must acknowledge a wildcard that most infrastructure-focused commentary underweights: consumer demand. Historically, it was not enterprise adoption that justified massive infrastructure buildouts — it was consumer behavior. The internet's fiber was justified by streaming. Broadband was justified by social media. The smartphone's chipsets were justified by apps. If AI follows this pattern — if personal assistants, AI-native search, AI agents, and AI entertainment create the kind of habitual, daily-use consumer behavior that smartphones did — then the current infrastructure may look not overbuilt but undersized within five years. But the current data on consumer willingness to pay is sobering: only a small fraction of ChatGPT's user base pays for the service. The consumer flywheel has not ignited yet. It might. But "might" is carrying a great deal of weight in a $660 billion capex plan.
The Variable Nobody Is Modeling
There is a second bull case, distinct from Jevons, that the oversupply narrative must contend with — and it may be the more powerful one. It is not about cheaper compute creating more demand for the same tasks. It is about AI fundamentally expanding the supply of cognitive labor, which in turn creates demand for more compute in a self-reinforcing loop.
The mechanism works like this. Studies consistently show performance gains of 10% to 25% in knowledge tasks — writing, research, programming, and financial analysis. Erik Brynjolfsson, writing in the Financial Times in early 2026, argues that the productivity takeoff is now visible in U.S. economic data: labor productivity grew roughly 2.7% in 2025, nearly double the 1.4% annual average of the prior decade. The St. Louis Fed estimates that workers using generative AI save about 5.4% of their work hours. EY-Parthenon projects that AI could lift economy-wide labor productivity by 1.5% to 3% over the next decade.
If those numbers hold — and they are contested — an NBER study of 6,000 executives across four countries found that the vast majority report little measurable impact on operations — the implications for compute demand are profound. A 15% productivity gain in software engineering does not just save costs. It enables a team of ten to produce what previously required twelve. If the market for software is elastic — and it has been for forty years — the response is not to fire two engineers. It is to keep all 10 and ship 15% more product. Which requires more compute to build, test, deploy, and run.
This is the cognitive labor supply shock. AI does not merely substitute for human workers. It augments them, making each unit of human cognitive output more valuable. Organizations that capture those gains invest more in the tools that created them. The result is not reduced compute demand. It is a compounding cycle where productivity gains fund further AI adoption, which drives further productivity gains, which funds further compute investment.
MIT economist David Autor has argued that mid-level professions — nursing, design, production management — can now access expertise previously available only to specialists, boosting the applicable value of their labor. If this pattern scales across knowledge work, compute demand could expand by an order of magnitude that current infrastructure forecasts do not capture.
There is a deeper reframe here that most bubble analyses miss entirely. Nearly every AI demand model — bullish or bearish — benchmarks against the software market: enterprise SaaS spend, cloud revenue, application-layer adoption. That is a market measured in hundreds of billions. But if AI functions not as software but as synthetic cognitive labor — performing tasks, not merely assisting with them — then the addressable market is not the global software TAM. It is the global knowledge-work TAM: roughly $30 to $50 trillion in annual spend on human cognitive labor across professional services, software development, finance, legal, education, and healthcare. If AI captures even 10% of that market over the next decade, the compute required would dwarf anything currently being built. And the usage model is shifting in a direction that compounds this effect. Today, most AI interaction is conversational — a user asks a question, receives an answer. But the emerging paradigm is agentic: autonomous AI systems running continuously in the background, monitoring codebases, executing research workflows, managing customer interactions, performing financial analysis around the clock. The shift from interactive to agentic compute multiplies per-user consumption by orders of magnitude. If millions of organizations deploy dozens of always-on agents, the compute demand curve bends upward in ways that make today's $660 billion buildout look not reckless but prescient.
But this is a J-curve argument. General-purpose technologies suppress measured productivity during an initial investment phase before entering a harvest phase. The personal computer followed this pattern — a decade of investment before the productivity surge of the late 1990s. Cloud computing followed it too. The question is whether investors and lenders will hold their nerve during the dip. And this is where historical analogy becomes treacherous.
The Telecom Parallel Nobody Wants to Hear
In 1999, telecommunications companies laid 28 million miles of fiber optic cable on the premise that internet bandwidth demand would grow indefinitely. They were right about the demand. They were catastrophically wrong about the timeline. Less than 5% of installed fiber capacity was operational when the bubble burst. The companies that built the infrastructure went bankrupt. The companies that eventually used the infrastructure — Netflix, YouTube, the entire streaming economy — did not exist yet.
The AI infrastructure buildout rhymes with this history in ways that should concern even the most committed bulls. The hyperscalers are spending on the assumption that a wave of AI-powered applications will eventually consume every unit of compute they deploy. Many of those applications do not exist yet. The ones that do exist are, in many cases, being financed by the same capital that is building the infrastructure to serve them.
But the parallel breaks down in one crucial respect: the companies building AI infrastructure today are not single-purpose telecoms. They are the most profitable businesses in human history. Amazon, Microsoft, Alphabet, and Meta generate enormous cash flows from advertising, e-commerce, cloud services, and enterprise software. Their AI spending is aggressive — Bank of America estimates they will push capex to 94% of operating cash flow in 2026, up from 76% in 2024 — but they are not leveraged telecoms betting everything on a single thesis. They can absorb writedowns that would destroy smaller companies.
The more vulnerable nodes are elsewhere: the neoclouds like CoreWeave, whose entire business model depends on GPU demand remaining insatiable; the AI labs like OpenAI, which projects $14 billion in losses by end of 2026 despite $20 billion in revenue; and Oracle, which has staked its corporate transformation on a $300 billion deal with a company that has never been profitable.
The Real Framework
So here is the synthesis that neither the bulls nor the bears are offering:
Yes, this is a bubble. The circular financing, the 10:1 capex-to-revenue ratio, the fact that AI-related stocks now represent 28.7% of the U.S. market index (up from 9.7% a decade ago), the price-to-sales ratios exceeding dot-com peaks — these are textbook indicators. Even OpenAI's chairman Bret Taylor has said plainly that AI will transform the economy and that we are in a bubble, and that a lot of people will lose a lot of money.
No, this does not mean AI is fake. Enterprise spending on generative AI has grown 22x in two years. Productivity gains are real and measurable — PwC documents 3x higher revenue-per-employee growth in AI-exposed industries. The technology is demonstrably transformative. The question was never whether AI works. The question is whether the financial structure built around it can survive the gap between investment and return.
The correction will be selective, not systemic. The hyperscalers will survive because they have diversified revenue and balance sheets that can absorb multi-year investment horizons. The casualties will be the neoclouds, the over-leveraged infrastructure plays, and the AI startups whose burn rates assumed capital would flow indefinitely. This is the 2000–2002 pattern: the technology was real, the internet was transformative, and the companies that overborrowed against tomorrow's revenue went bankrupt while the platforms that survived — Amazon, Google — went on to define the next two decades.
The depreciation schedule is the timing mechanism. A correction, if it comes, will not be caused by an accounting line item. It will be triggered by slowing AI revenue growth, falling GPU prices, or macroeconomic tightening — the usual suspects. But the depreciation schedule determines when the underlying demand-supply mismatch becomes visible on income statements. Depreciation is a non-cash expense; shortening a GPU's useful life from six years to three does not change the cash already spent. Sophisticated investors know this. But markets are not populated exclusively by sophisticated investors — algorithmic trading, passive index flows, and earnings-driven sentiment all react to reported numbers. And the cash flow picture is already strained: hyperscalers are pushing capex to 94% of operating cash flow, leaving little margin for error before depreciation revisions take effect. If GPU useful life compresses — which Nvidia's own annual chip release cadence suggests is plausible — the resulting acceleration of depreciation expense will not kill the hyperscalers. But it will amplify any demand disappointment, spook the market into repricing, and expose every entity in the ecosystem running on borrowed time and money. There is a milder scenario, and it is probably the most likely one: AI demand grows steadily but more slowly than projected, valuations compress by 20–30%, and the infrastructure proves useful over a longer timeline than investors price a crash, more painful than the bulls expect. But even the mild version is governed by the same underlying variable. The question is not whether AI works. It is whether demand can grow fast enough to justify the financial commitments already made.
The smartest position is not to bet for or against AI. It is to bet on the inevitability of the technology while respecting the fragility of its current financing. The infrastructure being built today will, like fiber optic cable in 1999, eventually prove essential. But the companies that built it may not be the ones that profit from it. And the reckoning between investment and return — between the trillion dollars going into the ground and the applications that will justify it — is not a question of if. It is a question of when.
The clock is the depreciation schedule. And it is already ticking.