Is AI a Bubble? The Case for Undercounting the Upside

December 2025 Update: Hyperscaler AI CapEx projections revised upward to $405B for 2025, up from initial $250B estimates. McKinsey's $2.6-4.4T annual value creation forecast gains validation as GitHub Copilot reaches 1.8M paying subscribers with 46% of code now AI-generated. Enterprise AI adoption hits 87% among large organizations with 130% YoY spending growth. The infrastructure investment case strengthens as productivity gains materialize faster than dot-com era comparisons suggest.

The question answers itself wrong. When Ray Dalio calls AI investment "very similar" to the dot-com bubble, when the IMF warns of an impending bust, when Apollo's chief economist notes that today's top tech companies trade at higher valuations than their 1990s counterparts—the framing assumes the central risk is overinvestment.¹ ² ³

The central risk runs the other direction. The speculation undercounts the upside.

Many AI companies will fail. The majority of startups raising at inflated valuations will return nothing to investors. Enterprise pilot projects will continue their 95% failure rate.⁴ Correction will come, probably painfully. None of that contradicts the thesis. The winners emerging from the current investment wave will generate value so disproportionate to the capital deployed that today's $380 billion in annual hyperscaler CapEx will look like a rounding error.⁵

The dot-com parallel actually makes the case. Amazon and Google emerged from that wreckage to create trillions in value. The infrastructure those bubble-era companies built—the fiber, the data centers, the protocols—powered decades of economic transformation. The losers lost everything. The winners changed civilization. AI follows the same pattern at larger scale and faster speed.

The value creation math that bubble discourse ignores

Bubble analysis fixates on the wrong ratio. Analysts compare AI investment to AI revenue and find a gap. The gap exists. The comparison misses the point.

The relevant comparison measures investment against value creation potential. When Anthropic deploys Claude across enterprises, the metric that matters isn't Anthropic's revenue. The metric that matters encompasses the productivity gains, cost reductions, and capability expansions across every organization using the technology. When a pharmaceutical company uses AI to identify drug candidates in weeks instead of years, the value created vastly exceeds the compute costs and licensing fees.

McKinsey estimates generative AI could add $2.6 to $4.4 trillion annually to the global economy across 63 use cases analyzed.⁶ That range represents a single year's impact, not cumulative value. Against $380 billion in annual infrastructure investment, the ratio favors the bulls by an order of magnitude—even if McKinsey's estimates prove optimistic by half.

The productivity evidence already accumulates. GitHub Copilot users complete tasks 55% faster than non-users.⁷ BCG consultants using GPT-4 produced 40% higher quality work in 25% less time.⁸ Customer service operations report 14% productivity improvements for agents using AI assistants, with the largest gains among novice workers.⁹ These numbers emerge from early, primitive deployments. The models improve monthly.

Every major platform shift in computing history generated value multiples exceeding initial investment. Personal computers created roughly $10 in economic value for every $1 invested in the industry during its formative decades.¹⁰ The internet generated similar or higher ratios. Mobile computing amplified the pattern. AI represents a general-purpose technology more fundamental than any of these—a technology that augments cognitive work the way previous revolutions augmented physical and information work.

The bubble framing treats AI like Pets.com: a company with no path to value creation selling products nobody needed. AI demonstrably creates value in every deployment. The question concerns magnitude and distribution, not existence.

Why adoption speed implies value creation speed

The St. Louis Fed publishes data that should terrify AI skeptics. Three years after the IBM PC launched in 1981, personal computer adoption reached 19.7%.¹¹ Three years after the internet opened commercially, adoption reached 30.1%.¹² Three years after ChatGPT launched in November 2022, generative AI adoption reached 54.6% by August 2025.¹³

The disparity reflects a structural advantage with profound value implications. Previous platform shifts required users to purchase new hardware, learn new interfaces, change established workflows, and rebuild processes around unfamiliar capabilities. AI adoption requires none of this. The technology reaches users through existing devices, existing browsers, existing applications. A knowledge worker in 2024 gained access to capabilities that didn't exist in 2022 without purchasing anything new or learning any new interface paradigm.

The adoption speed matters for value creation because value compounds with deployment. A technology that reaches 50% adoption in three years generates value during those three years. A technology that takes fifteen years to reach the same penetration delays value creation proportionally. AI's adoption velocity means value creation velocity.

Enterprise adoption numbers make the pattern concrete. By 2025, 87% of large enterprises had implemented AI solutions.¹⁴ Weekly usage of generative AI tools jumped from 37% to 72% year over year.¹⁵ Enterprise AI spending grew 130% in a single year.¹⁶ These organizations aren't spending out of enthusiasm. They're spending because the technology produces returns.

The pilot failure rate of 95% that critics cite actually supports the bullish case when examined carefully.¹⁷ Enterprises run pilots to find what works. A 95% failure rate among pilots means 5% success rate—and successful pilots scale into production deployments that generate sustained value. The enterprises continuing to invest after experiencing high pilot failure rates have found the 5% that works. They're optimizing, not abandoning.

The coding productivity revolution

Software development offers the clearest window into AI's value creation because productivity gains are measurable, immediate, and already operating at scale. Unlike abstract forecasts about future applications, coding productivity improvements generate hard data from millions of developers using AI tools daily.

GitHub Copilot provides the most robust evidence. GitHub's internal research found developers using Copilot completed tasks 55% faster than developers working without assistance.⁷ The study controlled for task complexity and developer experience levels. By late 2024, Copilot had accumulated over 1.8 million paying subscribers generating $400 million in annual recurring revenue.²⁷ More remarkably, 46% of code in Copilot-enabled repositories now comes from AI suggestions.²⁸

The technology has evolved beyond autocomplete. Early versions suggested the next few lines of code. Current versions propose architectural patterns, generate comprehensive test suites, write documentation, and debug complex problems. The shift from "suggest the next line" to "implement this feature" happened in roughly 18 months.

The ecosystem extends well beyond Copilot. Cursor, an AI-native IDE, grew from zero to over 40,000 paying users by late 2024, with monthly growth rates exceeding 50%.²⁹ Claude Code, Aider, and Cline enable autonomous coding agents that execute multi-file changes with minimal human intervention. Developers increasingly describe what they want rather than writing every character. The workflow transformation resembles the shift from assembly language to high-level programming—a categorical change in abstraction level.

The economic math makes the bubble framing difficult to sustain. The US Bureau of Labor Statistics reports median software developer compensation at approximately $127,000 annually.³⁰ A 55% productivity improvement translates to roughly $70,000 in equivalent additional output per developer. The US employs approximately 4.4 million software developers.³¹ Even conservative assumptions—20% adoption at 30% productivity gain—yield $26 billion in annual value creation from US software development alone. Global developer populations exceed 28 million.³²

The skeptic's counterargument deserves serious engagement: AI-generated code contains bugs, introduces security vulnerabilities, and creates technical debt. Developers spend as much time debugging AI output as they save generating it. The productivity studies were funded by AI companies with obvious incentive to produce favorable results. Real-world usage shows junior developers becoming dependent on tools they don't understand, degrading their fundamental skills.

The critique conflates naive usage with skilled usage. Studies showing quality degradation typically examine novices using AI as a crutch rather than professionals using AI as leverage. The BCG/Harvard study that found 40% quality improvement specifically examined experienced consultants—not junior developers fumbling with unfamiliar tools.⁸

Code review processes catch AI errors the same way they catch human errors. The security concern has merit but misses the trajectory: static analysis tools now scan AI-generated code automatically, and GitHub Advanced Security integration means AI suggestions receive security review before merge. The tooling catches up to the capability.

Most importantly, developers vote with their fingers. Copilot retention rates exceed 80%.³³ Cursor grew from zero to 40,000 paying users in months. If the tools delivered net negative value, developers would abandon them. They're not abandoning them; they're expanding usage.

Software represents a $650 billion global industry.³⁴ A 20% productivity improvement across the industry creates $130 billion in annual value—from one sector among dozens where AI creates measurable gains. The infrastructure investment enabling AI coding tools represents a fraction of that value creation. The bubble framing asks whether AI investment matches AI revenue. The correct framing asks whether AI investment matches AI value creation. In software alone, the ratio isn't close.

The dark fiber lesson: losers fund winners

Between 1995 and 2000, telecommunications companies invested roughly $2 trillion building 80 to 90 million miles of fiber optic networks.¹⁸ When the bubble burst, 95% of that fiber sat dark.¹⁹ The companies that built it—Global Crossing, Level 3, Qwest, WorldCom—went bankrupt or nearly so. Investors in those companies lost everything.

The fiber itself retained value. The infrastructure that destroyed its builders enabled its successors. Netflix streams video over cables that WorldCom's shareholders paid for. Cloud computing runs through dark fiber that bankrupted its original owners. The economic value created by that infrastructure across two decades dwarfs the $2 trillion invested by at least an order of magnitude.

The losers funded the winners. Equity investors in bubble-era telecom took the losses. The broader economy captured the gains. Amazon didn't build the internet infrastructure it depends upon; it built upon infrastructure that other companies' investors financed and lost.

AI follows the same pattern with one crucial difference: the ratio between infrastructure cost and value created is even more favorable. Fiber required physical installation across continents. AI infrastructure concentrates in data centers that serve global user bases. The capital efficiency of AI infrastructure exceeds the capital efficiency of physical network infrastructure by a substantial margin.

Today's AI infrastructure investment will produce losers. Startups valued at billions will return zero. Some infrastructure built speculatively will sit underutilized for years. The companies that survive and scale will generate value so enormous that the aggregate investment across winners and losers will seem modest in retrospect.

The S&P 500 has more than tripled since the dot-com peak.²⁰ The technology sector has grown faster. The bubble didn't destroy value permanently; it destroyed it for specific investors while creating conditions for long-term value creation. The same dynamic applies to AI, compressed into a shorter timeframe.

Why efficiency gains accelerate rather than reduce infrastructure demand

DeepSeek's January 2025 announcement demonstrated that frontier AI models could be trained for a fraction of previous costs—roughly $6 million compared to $600 million for GPT-4.²¹ The market interpreted this as bearish for AI infrastructure: if models could be built cheaply, demand for expensive hardware would collapse.

The interpretation revealed a fundamental misunderstanding of how general-purpose technologies create value.

In 1865, English economist William Stanley Jevons observed that the Watt steam engine's improved efficiency didn't reduce coal consumption—it increased it.²² Efficiency made steam power economically viable for applications that couldn't justify earlier, less efficient engines. More applications meant more total coal consumption despite lower consumption per application.

AI efficiency operates identically. When training costs drop from $600 million to $6 million, AI becomes viable for thousands of organizations that previously couldn't afford it. Universities, research institutions, small companies, and individual developers enter the market. Each participant consumes less compute than a frontier lab but collectively they consume more than frontier labs consumed when only frontier labs could participate.

The market figured this out within 48 hours. NVIDIA's stock recovered from its $593 billion single-day loss as institutional investors recognized the overreaction.²³ Days after DeepSeek's announcement, Meta raised its 2025 AI spending guidance to $60-65 billion.²⁴ The company saw cheaper training as a reason to train more models for more use cases, not a reason to reduce investment.

The most compelling evidence came from outside the technology sector. Chevron and GE Vernova announced plans to build power plants co-located with data centers, generating enough electricity for 3 million homes without connecting to the grid.²⁵ Industrial companies don't build dedicated power generation for a bubble. They build it for sustained, growing demand.

Efficiency improvements accelerate value creation by expanding the addressable market. Every efficiency gain makes AI viable for more use cases serving more users generating more value. The infrastructure required to serve that expanded market exceeds the infrastructure required to serve the previous, smaller market.

What the winners will look like

The bubble narrative assumes AI investment resembles dot-com investment: speculative capital chasing companies with no viable business model. The comparison ignores the structural differences.

Dot-com companies sold products that competed with physical alternatives while offering worse unit economics. Pets.com competed with grocery stores that didn't need to ship heavy bags of dog food individually. Webvan competed with supermarkets that could amortize delivery infrastructure across massive order volumes. The bubble companies offered inferior economics dressed up in internet aesthetics.

AI companies offer genuinely superior economics. The marginal cost of serving an additional AI query approaches zero once the model exists. The productivity gains from AI deployment reduce costs for customers rather than adding to them. The technology creates value on both sides of the transaction.

The winners emerging from the current cycle will likely concentrate in a few categories. Infrastructure providers controlling scarce resources—compute, data, and specialized talent—capture durable value. Platform operators achieving network effects in data and distribution compound advantages over time. Application companies solving specific high-value problems extract meaningful portions of the value they create.

Some winners already exist. NVIDIA's 80% market share and 78% gross margins reflect durable positioning.²⁶ Microsoft's Azure integration with OpenAI creates distribution advantages that compound. Anthropic and OpenAI themselves may capture substantial portions of the value their models generate, though the competitive dynamics remain unsettled.

New winners will emerge. The applications that AI enables—from autonomous vehicles to drug discovery to personal assistants that genuinely assist—will create trillion-dollar companies. The current investment funds the infrastructure those companies will need. The capital deployed today purchases optionality on transformations we can identify but not yet measure.

The actual risks

Intellectual honesty requires acknowledging what could go wrong, and the risks differ from the bubble narrative.

The concentration risk is real. Five companies—Amazon, Microsoft, Google, Meta, and NVIDIA—dominate AI infrastructure investment. If AI value creation concentrates as heavily as AI investment, the benefits flow to shareholders of a handful of companies rather than broadly across the economy. The winners win enormously; everyone else gains less than the aggregate numbers suggest.

Regulatory intervention could disrupt value capture. Governments increasingly view AI as strategically important and potentially dangerous. Heavy-handed regulation could slow deployment, fragment markets, or redistribute value from companies to compliance costs. The EU AI Act, export controls, and emerging national AI strategies all create uncertainty.

Technical limitations could constrain applications. Current models hallucinate, require careful prompt engineering, and fail unpredictably at edge cases. If these limitations prove fundamental rather than solvable, the addressable market for AI shrinks. The productivity gains measured in controlled studies might not replicate in messy real-world deployments at scale.

Competitive dynamics could destroy margins. If AI capabilities become commoditized before winners establish durable advantages, the value created might flow entirely to customers rather than to investors. This outcome would be excellent for economic productivity and terrible for AI equity investors.

These risks merit serious consideration. They differ fundamentally from bubble risks because they concern value distribution rather than value existence. The value creation happens regardless. The question concerns who captures it.

The frame that clarifies

The bubble framing asks: "Is AI investment excessive relative to AI revenue?" The answer is yes, obviously. Revenue lags investment in every infrastructure buildout. The question misunderstands infrastructure economics.

The correct framing asks: "Will the value created by AI justify the investment made to build AI infrastructure?" The evidence strongly suggests yes.

The winners will be worth far more than the entire investment category is worth today. The infrastructure built by losers will enable winners just as dot-com infrastructure enabled Google and Amazon. The transformation AI enables across the global economy—in productivity, in capability, in the nature of cognitive work—will make current investment levels look conservative in hindsight.

Many companies will fail. Investors in those companies will lose money. The aggregate value created will vastly exceed the aggregate capital deployed. The speculation undercounts the upside because the upside involves transformation of cognitive work across the global economy, and no financial model accurately captures that.

Organizations positioning for AI's value creation wave face decisions about infrastructure timing, architecture, and capability development. Teams at Introl deploy AI infrastructure across 257 global locations, helping enterprises build the foundation for capturing value from technology that's moving faster than any platform shift in history.

Key takeaways

For infrastructure strategists: - AI value creation math favors bulls: McKinsey estimates $2.6-4.4T annual economic impact vs $380B infrastructure investment—10:1 ratio even at pessimistic assumptions - Dot-com fiber lesson: $2T invested, 95% sat dark, original investors lost everything, but infrastructure enabled Google/Amazon to create trillions in value - DeepSeek efficiency breakthrough triggered Jevons Paradox: cheaper training enabled more organizations to train, increasing total compute demand

For executive leadership: - Adoption velocity unprecedented: 54.6% generative AI adoption in 3 years vs 30.1% internet and 19.7% PC at same point post-launch - 95% pilot failure rate supports bullish case: successful 5% scale into production deployments that generate sustained value - Winners will concentrate: NVIDIA (80% share, 78% margins), platform operators with network effects, and specific high-value application companies

For financial planning: - Software productivity alone creates $130B annual value: 20% improvement across $650B global industry from one sector among dozens - GitHub Copilot: 55% task completion improvement, 46% of code now AI-suggested in enabled repositories, 1.8M paying subscribers - Investment vs revenue framing misses point; correct framing: investment vs value creation across entire economy

For risk management: - Real risks differ from bubble risks: concentration (5 companies dominate), regulatory intervention, technical limitations, margin destruction from commoditization - Risks concern value distribution, not value existence: who captures created value matters more than whether creation occurs - Competitive dynamics could destroy margins: commoditization may flow value to customers rather than investors

For strategic planning: - Many AI companies will fail; losers fund winners as in every infrastructure buildout - Dark fiber precedent: Netflix streams on cables WorldCom shareholders paid for; infrastructure value exceeds builder equity - The speculation undercounts the upside because transformation of cognitive work across global economy exceeds any financial model

References

1. Fortune. "Everyone's Wondering If, and When, the AI Bubble Will Pop." Fortune, September 28, 2025.

2. Al Jazeera. "IMF Says AI Investment Bubble Could Burst, Comparable to Dot-Com Bubble." Al Jazeera Economy, October 14, 2025.

3. Yahoo Finance. "Big Tech's AI Investments Set to Spike to $364 Billion in 2025 as Bubble Fears Ease." Yahoo Finance, 2025.

4. Fortune. "Everyone's Wondering If, and When, the AI Bubble Will Pop."

5. CNBC. "How Much Google, Meta, Amazon and Microsoft Are Spending on AI." CNBC, October 31, 2025.

6. McKinsey & Company. "The Economic Potential of Generative AI." McKinsey Global Institute, June 2023.

7. GitHub. "Research: Quantifying GitHub Copilot's Impact on Developer Productivity and Happiness." GitHub Blog, September 2022.

8. Harvard Business School. "Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality." Working Paper 24-013, 2023.

9. Brynjolfsson, Erik, Danielle Li, and Lindsey R. Raymond. "Generative AI at Work." NBER Working Paper 31161, 2023.

10. Nordhaus, William D. "Two Centuries of Productivity Growth in Computing." Journal of Economic History 67, no. 1 (2007): 128-159.

11. St. Louis Fed. "The State of Generative AI Adoption in 2025." Federal Reserve Bank of St. Louis, November 2025.

12. ———. "The State of Generative AI Adoption in 2025."

13. ———. "The State of Generative AI Adoption in 2025."

14. Second Talent. "AI Adoption in Enterprise Statistics & Trends 2025." Second Talent, 2025.

15. Knowledge at Wharton. "2025 AI Adoption Report: Gen AI Fast-Tracks Into the Enterprise." Wharton School, 2025.

16. ———. "2025 AI Adoption Report."

17. Fortune. "Everyone's Wondering If, and When, the AI Bubble Will Pop."

18. Technostatecraft. "Dark Fiber—an Archaeology of the Dot-Com Bubble." Technostatecraft, 2025.

19. ISE Magazine. "The Perils of Irrational Exuberance—The 25th Anniversary of the Dot-Com Boom." ISE Magazine, 2025.

20. S&P Dow Jones Indices. "S&P 500 Historical Data." S&P Global, 2025.

21. Fortune. "What Is Jevons Paradox? The Reason Satya Nadella Says DeepSeek's New AI Is Good News for Tech." Fortune, January 27, 2025.

22. Jevons, William Stanley. The Coal Question: An Inquiry Concerning the Progress of the Nation, and the Probable Exhaustion of Our Coal Mines. London: Macmillan, 1865.

23. The Motley Fool. "Does DeepSeek's Breakthrough Help or Hurt Nvidia Stock? 3 Views You Should Know." The Motley Fool, February 3, 2025.

24. Substack. "The Jevons Paradox in AI Infrastructure: DeepSeek Efficiency Breakthroughs to Drive Energy Demand." AI Proem, 2025.

25. Axios. "DeepSeek Illustrates AI's Jevons Paradox." Axios, January 29, 2025.

26. Seeking Alpha. "Google Vs. Nvidia: Trillions of Dollars at Play." Seeking Alpha, 2025.

27. Microsoft. "GitHub Copilot Revenue and Subscriber Growth." Microsoft Fiscal Year 2024 Earnings, 2024.

28. GitHub. "GitHub Copilot Usage Statistics." GitHub Next Research, 2024.

29. The Information. "Cursor AI Coding Tool Growth." The Information, December 2024.

30. U.S. Bureau of Labor Statistics. "Occupational Employment and Wages: Software Developers." BLS, May 2024.

31. CompTIA. "State of the Tech Workforce 2024." CompTIA Research, 2024.

32. Statista. "Number of Software Developers Worldwide." Statista, 2024.

33. GitHub. "GitHub Copilot Retention and Engagement Metrics." GitHub Enterprise Report, 2024.

34. Gartner. "Worldwide Software Market Forecast." Gartner Research, 2024.