AI Is Not a Product Cycle The AI bears have a simple story: demand is not there, while the industry is building far too much capacity. Their latest exhibit is Big Tech’s roughly $ 3T of commitments for chips, data centres, power and leases. Because those items appear in SEC-filing footnotes rather than as current balance-sheet liabilities, they are presented as evidence of hidden excess and impending financial distress. This is sloppy. An obligation disclosed in an SEC filing is not hidden. It is disclosed. Nor is an item kept off the balance sheet under prevailing accounting rules automatically a concealed liability, an accounting abuse or a fraud. A future purchase commitment is not a bond coming due. An uncommenced lease is not unpaid debt. Calling everything “off balance sheet” is a useful way to make a large number sound sinister. It is not analysis. The question is straightforward: can these companies meet their obligations if AI demand weakens? That requires examining liquidity, payment schedules, termination rights, repricing provisions, asset redeployment, customer revenues and operating cash flow. A $ 3T gross number answers none of these questions. The danger, if there is one, is not in a footnote. It is in a funding cascade: contracts that cannot be deferred or reassigned, capacity that cannot be sold, and cash flows that no longer cover commitments. Show that mechanism. Show that the contracts are take-or-pay. Show utilisation is falling, prices are collapsing, enterprise adoption is stalling, or AI revenues are failing to cover the infrastructure bill. Until then, the bear case depends on one assertion: perhaps AI demand will not materialise. That is an unfalsifiable counterfactual, not an investment thesis. It cannot be disproved as stated because it provides no timeframe, measurable benchmark or identified constraint. It merely shifts the burden of proof: rather than demonstrate that demand is deteriorating, it demands that investors prove a demand shortfall can never occur. Any emerging technology can be made to look reckless under that standard. The burden should instead sit with those predicting a bust. They should identify the source of it: falling utilisation, collapsing compute prices, slowing adoption, weak AI revenue or a real economic limit on the work AI can perform. The implicit claim is more radical still: that the global economy is nearing saturation in its demand for compute. Governments have automated enough. Industry has extracted the available efficiencies. Businesses have delegated the tasks worth delegating. Consumers have discovered all the value they will find in cheap, programmable intelligence. This is not scepticism. It is an astonishingly confident forecast of technological exhaustion. AI is not a product cycle. It is the construction of an infrastructure layer for programmable intelligence. A chatbot is constrained by human attention. An agent is constrained by the scale of global economic activity. Compute is becoming a utility. The $ 3T is not proof that the world has bought too much intelligence. It is proof that it is preparing to use far more.