The Next Bubble

# The Case for an AI Software Bubble ## Overview The current AI software market has several characteristics associated with the early stages of financial bubbles: extraordinary revenue growth, rapidly rising valuations, intense competition, aggressive assumptions about future adoption, and a powerful narrative that the technology will transform the economy. The important distinction is that this is not necessarily a bubble because AI software is overvalued today. It is a potential bubble because investors are increasingly pricing companies according to what AI software *could become*, rather than what the industry can sustainably earn. There is already substantial evidence that the technology has genuine commercial value. OpenAI has surpassed $25 billion in annual revenue, while Cursor has grown from more than $1 billion to more than $2 billion in annual revenue. Gartner expects global AI software spending to reach approximately $453 billion in 2026, up from $283 billion in 2025. That makes the current environment more interesting than a traditional speculative mania. There is a real product, real demand and real revenue. The potential bubble lies in how much future value investors are already capitalizing into today's companies. ## The Narrative The central narrative behind AI software is that AI agents will become a new computing interface capable of performing work previously handled by humans using conventional software. Instead of buying software that helps employees perform a task, companies may increasingly buy AI systems that perform the task itself. Programmers may use agents to write code, lawyers may use them to analyze documents, and financial professionals may use them to perform research and analysis. If this continues, AI could address an enormous market because it is effectively competing for a share of the economic value of knowledge work. This creates a powerful investment narrative: AI will automate knowledge work, knowledge work represents enormous economic activity, and AI software companies will capture a meaningful portion of that value. The argument is not irrational. The problem begins when investors move from believing that AI will create enormous economic value to assuming that today's leading AI companies will necessarily capture most of it. ## Revenue Growth and Valuation One of the clearest signs of potential speculative excess is the speed at which valuations are increasing alongside revenue. Cursor provides a useful example. In November 2025, the company raised $2.3 billion at a valuation of approximately $29.3 billion after reporting more than $1 billion in annual revenue. Only months later, it had reportedly surpassed $2 billion in annual revenue, and reports subsequently suggested that it was discussing another financing round at a valuation of around $50 billion. The concern is not that Cursor's growth is illegitimate. It demonstrates genuine demand. The issue is what investors are extrapolating from that growth. A rapidly rising valuation assumes that growth can continue, pricing will remain attractive, margins will eventually become substantial, and competitors will not take away the company's market. The question therefore changes from whether AI software can generate billions in revenue to whether those companies can retain enough of that revenue as durable economic profit. ## AI Software Has Different Economics From Traditional SaaS Traditional SaaS (Software as a Service) has an attractive economic structure because serving another customer usually adds relatively little cost once the software has been built. AI software is different because many products incur variable inference costs as customers use them. This means that increasing usage can increase both revenue and expenses. AI companies must therefore balance subscription prices against the cost of generating responses, while competition pushes them toward lower prices and greater efficiency. That creates uncertainty around the quality of AI revenue. If customers can change their usage, switch models or move between providers, annual revenue may not have the same durability as traditional software subscriptions. This matters for valuations because investors are not only assuming that AI usage will grow. They are assuming that customers will continue paying enough for that usage to produce attractive margins. ## AI Could Also Destroy Existing Software Economics AI is not simply creating a new software category. It may change the economics of software itself. Gartner estimates that approximately $234 billion of enterprise application software spending could be exposed to "agentic arbitrage" by 2030, representing roughly 20% of enterprise SaaS spending. AI agents could increasingly perform tasks across CRM, accounting, project management, analytics and other applications, potentially reducing the amount of traditional software needed. This creates a paradox. AI software companies are being valued on the assumption that AI will capture enormous amounts of software spending, while the same technology could commoditize existing software and weaken pricing power across the industry. It is therefore unclear where the economic value will ultimately accrue. It may go to model developers, application companies, businesses controlling proprietary data, or platforms with strong distribution. Some existing software companies may instead lose value as AI makes their products less necessary. AI can therefore create enormous economic value without every AI company becoming enormously profitable. ## Adoption Is Ahead of Proven Economic Returns AI adoption is already widespread, but adoption does not necessarily equal economic value. McKinsey's 2025 global survey found that almost two-thirds of organizations remained in the experimentation or piloting phase for AI. 62% were experimenting with AI agents, but only 39% reported any enterprise-level EBIT impact from AI, and most of those companies said the impact was less than 5% of EBIT. This does not mean AI is failing. It means the transition from experimentation to economically significant deployment is still underway. That creates a potential gap between **AI adoption, AI spending and AI revenue** on one side and **sustainable productivity gains and corporate profits** on the other. Financial markets can price future economic benefits years before they appear in earnings, creating room for valuations to move ahead of proven returns. ## Valuations Depend on the Future The valuation problem becomes particularly visible at the frontier-model companies. Reuters reported in August 2026 that Anthropic's potential IPO valuation was being justified partly by projected revenue of approximately $190–200 billion in 2028, compared with a current annual revenue run rate of around $47 billion. Anthropic may eventually achieve those numbers, but the valuation requires investors to make enormous assumptions about future demand, pricing, costs, competition and market share. If any of those assumptions prove weaker than expected, the company could experience a significant valuation adjustment without the underlying technology actually failing. This is a classic feature of speculative markets: prices increasingly reflect a particular future rather than a broad range of plausible futures. ## Competition Could Destroy the Scarcity Premium Another risk is that investors may be treating AI capabilities as more scarce and defensible than they ultimately prove to be. The industry is already experiencing intense competition between OpenAI, Anthropic, Google, Meta, Chinese developers and numerous startups. As models become more capable and interchangeable, customers gain more negotiating power. Pricing pressure is already visible. Reuters reported in August 2026 that OpenAI had cut the price of its GPT-5.6 Sol developer model by more than 20% because of competition from Anthropic and Chinese AI developers. If this continues, the economic value of AI could increasingly accrue to customers rather than software providers. Businesses may receive enormous productivity gains while paying progressively less for the underlying intelligence. AI could therefore become extremely useful without producing the extraordinary margins that current valuations may imply. ## Reflexivity The AI software market also contains a strong reflexive mechanism. Investors believe AI will transform software, so they invest heavily in AI companies. Those companies can use the capital to hire talent, develop products, acquire customers and subsidize growth. Better products strengthen the narrative, which attracts more capital and supports higher valuations. At first, this can be productive because capital accelerates genuine technological development. But eventually the process can become self-reinforcing speculation, where high valuations themselves become evidence that AI's future is inevitable. If sentiment reverses, the mechanism can operate in the opposite direction. Financing becomes harder, companies prioritize profitability over growth, subsidies disappear and weaker competitors fail. The market then discovers which businesses have genuine competitive advantages and which depended heavily on abundant capital. ## The Core Bubble Mechanism The strongest case for an AI software bubble is therefore not that AI is overhyped. It is that AI could be **dramatically underpriced as a technology while simultaneously being overpriced as an investment**. AI could genuinely transform software development, customer service, law, finance and other industries. But if thousands of companies compete for the same economic value, prices can fall. If models become commodities, margins can fall. If AI agents reduce the need for traditional software, some existing markets can shrink. The critical question is therefore not whether AI creates value. It almost certainly will. The question is **who captures that value and how much investors are paying today for their share of it**. This distinction is important because a technology can create enormous economic value while producing disappointing investment returns. The internet did this during the dot-com bubble: the technology transformed the economy, but many companies purchased at extreme valuations failed to capture enough of that value to justify their prices. ## Why This Could Become a Bubble The current environment combines several conditions that make a bubble possible. AI is a genuine technology, which means the narrative has a strong foundation rather than being based on fiction. Revenue growth is also real and unusually visible, giving investors compelling numbers from which to extrapolate. At the same time, valuations increasingly depend on future dominance, while competition is already producing pricing pressure and industry-wide economic returns remain relatively early. This combination can create a dangerous feedback loop: strong technology produces strong expectations, strong expectations attract capital, capital accelerates growth, growth reinforces expectations, and valuations rise further ahead of proven profitability. The bubble would not necessarily require AI revenue to collapse. It could emerge simply because **future growth and profitability have been priced too aggressively**. ## The Counterargument There are strong reasons the AI software market could continue expanding without producing a major bubble. AI adoption is accelerating, spending is growing rapidly, and companies such as OpenAI and Cursor are already generating billions of dollars in annualized revenue. Gartner expects AI software spending alone to reach approximately $453 billion in 2026. The market could therefore be correctly recognizing the beginning of an enormous technological transition. If AI becomes the dominant interface for knowledge work and today's leading companies capture a substantial portion of that economic activity, today's valuations could eventually prove reasonable. The more sophisticated argument is not that AI software is guaranteed to be a bubble. It is that **the success of AI as a technology and the success of AI investors are two different questions**. ## What Would Confirm the Bubble? The strongest warning signs would be slowing AI software revenue growth while valuations continue expanding, accelerating price competition, declining gross margins, increasing customer churn and continued dependence on enormous future revenue forecasts. Another important signal would be continued growth in enterprise AI adoption without a corresponding improvement in measurable enterprise-level returns. If businesses keep spending heavily on AI but struggle to generate meaningful productivity and profit improvements, the gap between the investment narrative and economic reality would become increasingly difficult to justify. If this happens alongside continued aggressive investment, the industry could enter a classic late-bubble cycle: competition increases, prices fall, margins deteriorate, growth disappoints and investors reassess the valuations they assigned to future growth. ## Bottom Line The strongest case for an AI software bubble is not that AI is useless or that its revenues are fake. The evidence points in the opposite direction. AI software is already generating enormous and rapidly growing revenue, while Gartner expects global AI software spending to exceed $450 billion in 2026. The potential bubble comes from the capitalization of future growth. Investors are increasingly pricing companies on the assumption that AI will capture a massive portion of the value of knowledge work, that today's leaders will remain dominant, that pricing will remain attractive and that extraordinary revenue growth will eventually become extraordinary profits. Yet enterprise-wide ROI remains early, competition is pushing prices downward, and AI could simultaneously create enormous new software markets while undermining the economics of existing ones. That creates the central paradox: **AI software can become really important technology and still experience a financial bubble.** The technology determines whether AI succeeds. Competition determines who captures the value. Valuation determines whether investors make money from that success.

## The Fundamentals Silver’s 2025–2026 move was not a simple “industrial-demand bull market.” It began with a real fundamental backdrop: multi-year market deficits, strong industrial use, constrained mine elasticity, tight London liquidity, and rising bar-and-coin demand. It became a bubble-like overshoot once retail ETF inflows, momentum chasing, leveraged products, and exchange-margin dynamics took over. On the way up, silver moved from $35.82/oz on 5 June 2025 to $77.40 on 26 December 2025, then to a spot peak of $121.6 on 29 January 2026. On the way down, it fell to $56.41 on 24 June 2026, a 53.6% drawdown from the January spot peak, before stabilizing near $60.64 in early July. Reuters, BIS and CME evidence all point to the same mechanism: a fundamentally tight market that was then pushed into a self-reinforcing squeeze by speculative and leveraged capital. ## A Tight Physical Market The most important mechanical driver was the mismatch between available float and financial claims. World Silver Survey 2026 says the market remained in deficit in 2025 and was expected to remain in deficit in 2026; Reuters also reported that by end-September 2025 London vaults held 24,581 tonnes of silver, but 83% of that stock was already allocated to ETFs. That implies only a relatively small pool of readily available London metal against a market where SLV alone held 480.3 million ounces on 2 July 2026, and where the CFTC’s 23 June 2026 combined silver futures-and-options open interest represented 134,115 contracts, or about 670.6 million ounces notional. In other words, the paper market was large relative to both annual mine supply and the likely “free” physical float. ## Supply and Demand Conditions The supply side mattered, but mostly because it could not respond quickly. World Silver Survey 2026 reported 2025 mine production of 846.6 Moz, up 3%, and recycling of 197.6 Moz, up 2%, yet total demand still exceeded supply. Industrial demand actually fell 3% in 2025 to 657.4 Moz, largely because photovoltaic users accelerated thrifting and substitution as prices rose, but that weakness was more than offset by a 14% rise in coin and net-bar demand. Official-sector activity was negligible: the survey table shows net official sector sales of just 1.5 Moz in both 2024 and 2025. That means the late-2025 and January-2026 price explosion cannot be explained by sovereign accumulation; it was overwhelmingly a private-market phenomenon. ## The Rise of Speculation and Leverage On the demand side, the market changed character in late 2025. Earlier in 2025 the story was “silver is catching up to gold because of industrial demand and deficits.” By December 2025 and January 2026, Reuters and BIS describe an increasingly speculative market: large retail ETF inflows, persistent ETF premia to NAV, surging activity in both bullish and inverse leveraged silver ETFs, and record derivatives volumes. Vanda Research told Reuters that retail investors had bought $921.8 million of silver ETFs over 30 days by 15 January 2026, and then $171 million of SLV in a single day on 26 January; almost double the one-day peak seen during the 2021 silver squeeze. BIS concluded that retail-driven exuberance, leveraged ETF rebalancing, CTA/trend-following flows and margin-triggered liquidations amplified both the rise and the collapse. ## A Bubble Built on Fundamentals The 2026 Silver Episode does fit a bubble definition, but not in the sense of “zero fundamentals.” A better label is a fundamentals-based squeeze that turned into a speculative bubble. A price around $60/oz had support from deficits and tightness, Bank of America’s strategist told Reuters that about $60 was a fundamentally justified level, and Reuters quoted analysts on 2 February saying the market was searching for a more fundamentally supported floor around $60–$70. But the move to $121.6 required narrative contagion, one-sided ETF buying, leverage, and technical feedback loops. The subsequent crash after the Fed-chair shock, stronger dollar, and repeated CME margin hikes is exactly what one would expect when prices have moved materially beyond what physical fundamentals alone can sustain.