When sellers explain why you must buy now, check your wallet. The more the AI industry justifies its own spend, the clearer the fragilities look. A skeptical analyst asked insiders to stress his doubts and left even more convinced: the economics do not add up, the power constraints are real, and the feedback loop is one way. In game theory, when everyone repeats the same message, it often signals coordination, not truth. Markets rarely peak on silence. They peak on chorus.
Start with incentives. Hardware vendors, cloud providers, and consultants sell a story of inevitable scale. The buyers are often the same firms that also supply the parts, finance the leases, and control the software distribution. That is a principal agent problem, wrapped in a vertical stack. Insiders are rational. They talk their book. But when the distribution of claims is narrow and the counterpoints come from the purchase orders, not independent P and L, you have a signaling equilibrium. The message optimizes for budget approval, not base rates. History rhymes. Fiber buildouts in 1999, 3G spectrum in the early 2000s, and shale drilling in the last decade all shared one feature: the pitch sounded most compelling from the companies that needed you to believe it.
Every transformational capex cycle eventually collides with constraints that do not negotiate. GPUs need power, cooling, land, high voltage interconnects, and water. Lead times for large transformers run years. Interconnection queues delay projects more than delivery trucks. At scale, inference is an energy business wearing a software badge. The marginal cost per token is not free. Training spend can be capitalized. Inference spend shows up in operating costs each month. That matters. Electricity does not care about your valuation multiple. The purest test of a technology wave is whether unit economics improve as you grow. For many AI workloads today, growth pushes you into higher power prices, scarcer sites, and stricter regulation. That is fragility, not scale.
Optimists frame AI as a productivity multiplier on the order of electricity or the internet. If that is right, it must show up in cash costs per task and gross margins for customers, not just in demo reels. Some observers note that most AI spend is coming from tech giants redeploying profits from their core monopolies. That is not a sin, but it is a signal. Cross subsidized booms are durable until the core business stumbles or regulators cap the subsidy. A true productivity revolution escapes the lab and pays for itself outside the original profit pool. Until CFOs see lower error rates, faster cycle times, and fewer heads for the same output, the payback remains a promise. The right inversion is simple: if AI is as inevitable as advertised, why do so many features ship with negative gross margins and usage caps designed to limit the bill?
Training is a one time mountain. Inference is the toll road you drive every day. Enterprises will not absorb an unbounded inference tax. Common sense and basic probability say adoption plateaus below the booster slides: a minority use intense generation every hour; many will default to simpler tools; most workflows do not need a stochastic co pilot at each step. That is not heresy. It is how software adoption always looks. The average ROI for new enterprise systems skews lower than the case studies, and usage decays after novelty. When the novelty fades, the budget questions start. What happens to the margin profile when customers push back, when they set hard cost per seat caps, or when they move models on prem to avoid cloud markups? The answer decides who eats the inference bill.
AI does not sit outside the financial system. It is wiring into it. Regulators have warned that autonomous agents interacting in markets can drift toward coordinated outcomes, even without explicit collusion. Shared training data, similar reward functions, and identical cost signals make behavior correlate. That is a recipe for crowded trades and synchronized errors. Add supply chain concentration at the chip foundry, HBM memory choke points, and a single vendor for the hottest accelerators, and your tail risks get fatter. One plant fire or export ban ripples across models, clouds, and customers. Enthusiasts counter that more AI makes the system smarter. Maybe. But complexity plus tight coupling breeds brittleness. The 2010 flash crash was not fixed by wishful thinking. It was contained by circuit breakers and hard limits.
Investors love the narrative that a few platforms will capture all the value. Sometimes they do. More often, the winner’s profits get bid away in the race, or flow to less glamorous toll collectors. In railroads, it was landowners and freight customers. In the mobile boom, carriers fought churn while landlords and tower REITs raised rents. In AI, the eventual value may accrue to regulated utilities selling electrons, to memory suppliers with pricing power, to landlords controlling permits, or to software incumbents that bundle basic models into existing seats and squeeze point solutions. Shareholders can be directionally right and financially wrong if estimates bake in permanent scarcity and monopoly margins. Competition, commoditization, and regulation are not bugs. They are the base case.
Run simple scenarios. Power prices up 20 percent. Transformer deliveries slip 12 months. Export controls tighten one notch. Model performance improvements slow from 10x to 2x year on year. Customers cap AI opex at 5 percent of IT spend. Which business models break? Which stocks need flawless execution to justify today’s assumptions? This is not cynicism. It is risk management. The Fed urges institutions to embrace both the power and responsibility of AI, which includes contingency planning for unintended outcomes. Scenario work exposes where fragility hides: in levered data center financing, in take or pay contracts for power, in SLAs that assume uptime the grid cannot provide, and in teams that built their budget on the straight line of last year’s adoption curve.
The right contrast to euphoria is not doom. It is discipline. Look for systems that benefit from volatility, not narratives that require endless straight lines. In practical terms that means business models with variable costs that flex with demand, balance sheets that do not depend on short term refinancing, and pricing power that does not rely on hype. It means focusing on unit economics over total addressable markets, on throughput rather than headline FLOPs, on customer retention without subsidies. As one strategist put it, AI can transform work. That does not absolve markets from asking who pays, when, and at what margin. When insiders offer more reasons to accelerate spend, treat it like any other cycle. Test the incentives, check the physics, and keep your distance from the edge.
