Everywhere you look right now some commentator is talking about an AI bubble, saying “ChatGPT isn’t profitable! AGI is far away! It’s a circular economy! We’re in a bubble. We’re in a bubble. We’re in a bubble.” Either the people saying these things don’t understand bubbles, startups, adoption cycles, or circular economies because … this is not a bubble. It is a civilization-changing technology of nearly unlimited potential. This does not mean there is no risk, far from it; but the risk is likely not economic. Let’s take a look at some historical bubbles and separate the hype from the hysteria from the data.
In financial markets, few phenomena generate as much debate and retrospective hand-wringing as bubbles. Yet despite their recurring nature throughout economic history, the precise definition of a bubble remains surprisingly contentious among economists and market analysts. Understanding what constitutes a genuine bubble, examining historical precedents, and distinguishing speculative excess from real value creation has never been more critical, as artificial intelligence reshapes the global economy amid concerns that we’re witnessing another irrational exuberance.
Defining the Bubble
At its core, a market bubble represents a significant and sustained deviation between an asset’s price and its intrinsic value, driven primarily by speculation rather than fundamentals. Economists generally identify several key characteristics:
- rapid price appreciation that far exceeds historical norms
- widespread public participation driven by fear of missing out
- easy credit or abundant capital fueling purchases, a compelling narrative that justifies “this time is different,”
- and ultimately, a dramatic collapse that destroys substantial wealth.
What distinguishes a bubble from mere overvaluation is the self-reinforcing nature of price increases. As prices rise, they attract more buyers, which pushes prices higher still, creating a feedback loop disconnected from underlying economic value. This speculative frenzy typically involves leverage, where borrowed money amplifies both gains and eventual losses. Crucially, bubbles are characterized by assets trading at prices that cannot be justified by any reasonable projection of future cash flows or utility.
Three Historic Bubbles
The Dutch Tulip Mania (1636-1637)
Perhaps the most famous bubble in history occurred in the Dutch Republic during the 1630s, when tulip bulb prices reached extraordinary levels before collapsing spectacularly. At the peak, rare tulip varieties sold for more than ten times the annual income of a skilled craftsman. Single bulbs of the Semper Augustus variety reportedly traded for amounts equivalent to the price of luxurious Amsterdam canal houses.
The mania was fueled by a futures market for tulip bulbs that allowed speculation with minimal capital. Buyers purchased contracts for bulbs still in the ground, creating a derivatives market detached from the physical commodity. The psychological dynamics were classic bubble behavior: newcomers entered the market after witnessing neighbors’ wealth accumulation, convinced that prices would continue rising indefinitely. When the bubble burst in February 1637, prices collapsed by over 95%, leaving countless speculators holding worthless contracts and destroying fortunes virtually overnight.
The Dot-Com Bubble (1995-2000)
The late 1990s witnessed extraordinary speculative frenzy centered on internet-related companies. The NASDAQ Composite index surged from under 1,000 points in 1995 to over 5,000 by March 2000, driven by investor conviction that the internet represented a paradigm shift requiring new valuation metrics. Traditional measures like price-to-earnings ratios were dismissed as outdated, with investors focusing instead on metrics like “eyeballs” and “page views.”
Companies with minimal revenue and no path to profitability achieved billion-dollar valuations through initial public offerings. Pets.com, the poster child for excess, spent lavishly on Super Bowl advertising while losing money on every transaction, ultimately collapsing just nine months after its IPO. Webvan, a grocery delivery startup, was spending millions on logistics infrastructure with a handful of customers. The bubble was characterized by a complete disconnect between stock prices and business fundamentals. When reality reasserted itself beginning in March 2000, the NASDAQ would eventually lose 78% of its value, wiping out approximately $5 trillion in market capitalization.
The Housing Bubble (2004-2007)
The mid-2000s housing bubble represented a more dangerous phenomenon because it involved the leveraging of household balance sheets and the global financial system. Home prices in the United States rose 124% between 1997 and 2006, fueled by historically low interest rates, lax lending standards, financial innovations like subprime mortgages and collateralized debt obligations, and widespread belief that housing prices never decline nationally.
Speculation became rampant, with investors purchasing multiple properties as investments, often with no money down and no income verification. The assumption that prices would continue rising indefinitely encouraged excessive leverage throughout the financial system. When prices began falling in 2006, the collapse triggered a global financial crisis, as the complex securities built on these mortgages proved nearly worthless, threatening the entire banking system and precipitating the Great Recession.
Why AI Is Not a Bubble
Despite superficial similarities to previous manias (rapid capital inflows, soaring valuations, and breathless media coverage) the current AI surge displays fundamental differences from historical bubbles that suggest genuine value creation rather than speculative excess.
First, AI is demonstrating clear, measurable productivity improvements across industries today, not promises of future profitability. ChatGPT has gone from zero in revenue to 20 billion a year in revenue in three years. This is one of the fastest growth cycles we have ever seen, yes; and it is based on real revenue and real value. This is not a matter of flowers suddenly becoming more expensive than they were because they are in demand. This is a pattern of actual subscriber growth and real business utility. Companies are achieving documentable cost reductions, efficiency gains, and revenue enhancements. McKinsey research indicates that generative AI alone could add $2.6 to $4.4 trillion annually to the global economy. Unlike dot-com companies burning cash without viable business models, AI is generating immediate returns on investment for enterprises implementing these technologies.
Second, the revenue growth supporting current valuations is real and accelerating. Leading AI companies are posting substantial, profitable revenue increases, not the imaginary metrics that characterized earlier bubbles. Cloud computing revenue directly attributable to AI workloads is growing at 40-50% annually, representing genuine customer spending on productive capabilities, not speculative positioning.
Third, the capital being invested is flowing toward infrastructure with tangible utility beyond the hype cycle. The hundreds of billions being spent on data centers, chips, and computing capacity create durable assets with multiple use cases. This contrasts sharply with tulip futures or speculative home flipping, where value evaporated completely upon price collapse. Even if AI growth moderates, this infrastructure retains substantial value.
Fourth, adoption is being driven by rational business decisions rather than retail speculation. Corporate adoption of AI follows rigorous ROI analysis and pilot programs demonstrating efficacy. This bottom-up, use-case-driven implementation differs markedly from the top-down speculation that characterized previous bubbles.
We are witnessing the emergence of a genuine general-purpose technology comparable to electricity or computing itself. AI is transforming workflows across virtually every industry, from drug discovery to software development to customer service. The breadth and depth of application distinguish it from narrower innovations that captured imaginations but delivered limited practical value.
Certainly, some individual AI companies may be overvalued, and a correction in specific stocks seems inevitable. However, the fundamental technological transformation underway represents authentic value creation. The infrastructure being built, productivity being unlocked, and capabilities being developed will generate returns for decades, distinguishing this moment from history’s speculative manias. What we’re experiencing isn’t a bubble destined to pop: it’s the foundational phase of technology’s next era.
Q&A: Common AI Bubble Concerns
Q: Aren’t AI companies just propping each other up with circular deals? We keep hearing about tech giants investing in AI startups while simultaneously becoming their biggest customers. Isn’t this artificial demand?
A: This concern confuses standard enterprise technology adoption patterns with circular financing schemes. When Microsoft invests in OpenAI and then purchases OpenAI services to integrate into its products, that’s not a fake deal, it’s vertical integration, a practice as old as modern business itself. When companies buy NVIDIA chips they do so because they need the chips to accommodate demand, not to prop up NVIDIA. Microsoft is buying a capability it then resells to millions of actual end-users through products like Copilot, which generated over $1 billion in revenue in its first year. Copilot is now struggling. Is this a sign of a bubble? No. It’s a sign of a product failure, partnership failure, of product market fit failure, which is not in and of itself a sign of a bubble.
The “customer-investor” relationship is typical in B2B technology markets. Amazon Web Services didn’t become profitable by selling only to unrelated third parties; major AWS customers include Amazon’s own retail operations. Google Cloud serves Google’s internal needs while selling to external enterprises like Apple. This strategic alignment allows companies to develop technologies at scale while proving real-world utility.
What would indicate artificial propping-up would be if these companies were buying services they don’t use or can’t monetize. Instead, we’re seeing genuine integration into products with measurable user adoption and revenue generation beyond the initial investor-customer relationship.
Q: ChatGPT and similar AI services require massive investment but aren’t profitable yet. Isn’t this exactly like WeWork or other startups that burned billions before collapsing?
A: This question reveals a fundamental misunderstanding of how transformative technology startups operate and why the comparison to failed ventures like WeWork is inapt.
First, let’s address the startup economics. High capital requirements before profitability aren’t indicators of a bubble, they’re the defining characteristic of infrastructure-scale technology platforms. Amazon lost money for its first nine years, investing heavily in warehouses, technology, and logistics. Those weren’t wasted billions; they were strategic investments building durable competitive advantages. Amazon’s cumulative losses through 2002 exceeded $3 billion, yet no serious analyst would call that period a bubble. The company was deliberately prioritizing growth and market position over near-term profitability.
The critical distinction lies in unit economics and the path to profitability. WeWork’s fundamental problem was that its core business model (leasing long-term, subleasing short-term) had negative unit economics that worsened at scale. Each new location increased losses. The business couldn’t ever be profitable without completely reinventing itself.
ChatGPT and similar AI services face the opposite dynamic. Their current unprofitability stems from massive upfront infrastructure investment and the deliberate choice to underprice services during market development. The unit economics are improving, not deteriorating. As model efficiency improves, inference costs decline dramatically. GPT’s inference costs have dropped over 95% since launch through optimization. Simultaneously, usage is growing exponentially, allowing fixed infrastructure costs to be amortized across more users.
OpenAI’s revenue trajectory illustrates this: the company crossed $2 billion in annual recurring revenue within two years of ChatGPT’s launch, among the fastest revenue ramps in technology history. More importantly, the marginal cost of serving an additional user is declining while willingness to pay is increasing, as evidenced by successful enterprise tier pricing and API usage growth.
The investment required isn’t disappearing into operational expenses that must be repeated indefinitely. It’s building models, infrastructure, and capabilities that become more valuable and less expensive to operate over time. Each generation of AI models requires substantial training investment, but once trained, can serve billions of queries with decreasing marginal costs.
Furthermore, the monetization paths are multiplying. Beyond direct subscriptions, AI companies are licensing models to enterprises, selling API access at improving margins, and creating entirely new business categories. The revenue diversification and growth rates demonstrate genuine market demand, not artificial hype.
Compare this to genuine bubble companies: Pets.com was losing money on every transaction and had no path to profitability because shipping heavy bags of dog food to individual homes couldn’t ever compete economically with retail distribution. The business model was fundamentally broken. AI services are fundamentally sound but capital-intensive during the buildout phase, a crucial distinction.
Is investment in these companies building durable value? Does a clear path to profitability exist? For leading AI companies, both answers are demonstrably yes. The infrastructure being built has decades of utility, the technology is improving rapidly, costs are declining, and revenue is growing at rates that validate the investment thesis. This is how transformative technology platforms are built. This is evidence of a genuine technological revolution in its capital-intensive development phase.
Q: But hasn’t AI failed to deliver on its promises before? We’ve had multiple “AI winters” where the technology couldn’t live up to the hype. Why should we believe this time is actually different?
A: The AI winter comparison is historically important but fundamentally misunderstands what’s changed. Previous AI disappointments resulted from a consistent pattern: researchers made genuine theoretical breakthroughs, media and investors extrapolated those breakthroughs into near-term capabilities that didn’t exist, funding flooded in based on unrealistic expectations, and when the technology couldn’t deliver on overpromised timelines, investment collapsed.
The 1960s AI spring promised human-level machine intelligence within a generation based on early successes in game-playing and theorem-proving. The 1980s expert systems boom suggested AI would soon replace human expertise across domains. Both collapsed because the underlying technology hit hard theoretical and computational limits. Expert systems were brittle, requiring manual encoding of knowledge that didn’t scale. Early neural networks couldn’t handle complex real-world problems due to computational constraints and algorithmic limitations.
What makes the current era fundamentally different isn’t hype, it’s demonstrated capability at commercial scale. We’re not talking about laboratory demonstrations or narrow proof-of-concept projects. AI systems today are performing economically valuable work across millions of businesses. GitHub Copilot is writing roughly 40% of code for developers who use it. AI systems are reading medical imaging with accuracy matching or exceeding specialists. Customer service operations are handling millions of interactions autonomously. Legal document review, financial analysis, content generation; these aren’t just future promises, current deployments are generating measurable value.
The technical breakthroughs enabling this are profound and durable. The transformer architecture solved fundamental problems that plagued earlier approaches. Massive computing infrastructure enables training at scales previously impossible. We’ve crossed a threshold where models demonstrate emergent capabilities; they can perform tasks they weren’t explicitly trained for, showing genuine generalization and not just pattern recognition.
Previous AI winters occurred because the technology fundamentally couldn’t do what was promised. Today’s AI is doing things skeptics claimed were decades away, and doing them at commercial scale. The concern isn’t whether AI works: millions of deployed systems prove it does. Yes there are failures and limitations. There always will be. The question is how quickly capabilities will advance and how broadly they’ll be adopted, which are matters of degree, not existential questions about viability.
The infrastructure investment this time is fundamentally different. Companies aren’t betting on theoretical future capabilities. They’re expanding proven systems already generating returns. The billions being invested in chips and data centers support AI workloads running today, not speculative technologies that might work eventually. Even if progress plateaus tomorrow, the current generation of AI would justify substantial ongoing investment based on demonstrated utility.
The AI winter analogy also ignores that even during those periods, foundational research continued and ultimately enabled today’s breakthroughs. The “winters” were investment cycles, not technological dead ends. Today’s spring builds on decades of accumulated knowledge, with the crucial addition that the technology now works reliably enough for production deployment. That’s the difference between speculation and reality, and why comparing today’s AI surge to historical disappointments misses how fundamentally the landscape has changed.
Q: But aren’t we still really far away from AGI? Doesn’t that prove this is all hype since artificial general intelligence was supposed to be the real breakthrough?
A: This question confuses two entirely separate value propositions: artificial general intelligence (AGI) as a theoretical end goal, and narrow AI as an immediately valuable commercial technology. This confusion actually reinforces why the current AI surge isn’t a bubble: the economic value being captured today doesn’t depend on achieving AGI at all.
To be clear, we’re likely still a ways away from AGI, depending on how you define it. AGI, a system with human-level reasoning across all cognitive domains, remains an ambitious research goal with massive technical hurdles. We don’t have clear consensus on how to achieve it, and many researchers debate whether current approaches can even get us there.
But here’s what matters for market fundamentals: that doesn’t matter for current valuations and investments.
The commercial AI revolution isn’t predicated on achieving AGI. It’s built on narrow AI systems that excel at specific, economically valuable tasks. A system doesn’t need human-level general intelligence to read radiology scans more accurately than most humans, to generate functional code that accelerates developer productivity by 40%, or to handle customer service inquiries at scale. These capabilities are valuable precisely because they’re reliable, deployable, and solve real business problems today.
There are historical parallels. The automotive industry generated trillions in economic value without ever achieving the original vision of fully autonomous vehicles that could navigate any condition. Commercial aviation became transformative despite never delivering the personal flying cars futurists promised. The internet created unprecedented wealth before we achieved the semantic web or AI agents that early visionaries imagined. Technologies create genuine value by solving practical problems, even when falling short of maximalist visions.
The GDP impact of current AI (McKinsey’s projection of $2.6 to $4.4 trillion annually) comes entirely from narrow applications we can deploy right now. Enterprise software automation, content generation, data analysis, personalized recommendations, supply chain optimization, drug discovery acceleration—none of these require AGI. They require AI that’s good enough at specific tasks to generate positive ROI, which is exactly what we have.
The AGI timeline is actually irrelevant to the bubble question. If anything, the fact that we’re capturing massive value without AGI strengthens the case against a bubble. Bubbles occur when prices rely on distant, speculative outcomes. The dot-com bubble required believing that companies with no revenue would somehow become profitable eventually. The current AI investment is generating returns now, with deployed systems creating measurable value today. The potential arrival of AGI would be upside, not the base case justifying current investment.
In fact, the obsession with AGI timelines often comes from AI skeptics seeking to move the goalposts. When AI systems achieve specific capabilities (playing Go, generating coherent text, creating images, writing code) skeptics respond, “But that’s not real intelligence.” The criticism shifts from “AI can’t do X” to “AI doing X doesn’t count because it’s not AGI.” This rhetorical move ignores that economic value comes from capability, not from matching some philosophical definition of general intelligence.
The businesses investing billions in AI infrastructure aren’t betting on AGI arriving by 2027. They’re deploying systems that improve operations today and will continue improving incrementally. Even if we hit a plateau well short of AGI, if current architectures can’t scale beyond a certain capability level, the technology would still justify enormous ongoing investment based on what it can already do.
Think of it this way: electricity transformed the economy without ever achieving wireless power transmission that early pioneers like Tesla envisioned. The internal combustion engine created the modern transportation system despite never becoming as efficient as thermodynamic theory suggested was possible. Technologies create value by being useful, not by achieving theoretical maximums.
The AI sector today is built on demonstrated, deployable capabilities generating real revenue. AGI remains an exciting research frontier, but confusing its timeline with current AI’s commercial viability is a category error. We’re not in a bubble waiting for AGI—we’re in a genuine technology buildout capturing value from narrow AI while researching more advanced systems. Those are fundamentally different propositions, and the latter is exactly what sustainable technology investment looks like.
- *AI use disclosure: Claude Sonnet 4.5 was used in the outline and proofreading of this article.
