How the AI Boom Might Really End: Five Scenarios

Summary

Martin Alderson outlines five ways the current AI investment boom could unwind, while arguing that AI is already creating substantial economic value. He divides the risks into two groups: AI may fail to deliver further value or may be constrained by external events, or AI may work well while the investors funding it fail to capture the returns. The first group includes a scaling wall in which larger training budgets produce only marginal gains, severe alignment failures that trigger political backlash, and opposition to data-centre construction. Alderson considers the first possibility unlikely for now because model capabilities continue to advance, and he argues that local data-centre bans would have limited global effect, although financing could still be disrupted by a chain of project failures. The second group is presented as more consequential. Open-weight models that become good enough for common tasks could intensify the inference price war, compress frontier-lab margins, and push value toward users and infrastructure providers. The article cites large price differences for short-context and long-context workloads, while noting that the pricing of newer models and the capabilities of frontier labs remain uncertain. A separate risk is a major hardware or software efficiency breakthrough that lets supply expand much faster than token demand, potentially lowering GPU prices and weakening the investment case. Alderson says the key indicator would be broad declines in spot-market GPU prices rather than token-cost indexes that ignore demand growth. His conclusion is that the boom could end with the same number of tokens, or more, moving through a more fragmented market at lower margins. In that scenario, frontier-lab valuations could fall while demand for compute, memory, and power remains strong. Debt financing is the main accelerant: even a modest change in expectations could unwind infrastructure investments quickly, despite continued demand for GPUs. The article also argues that headline chip performance gains can overstate underlying efficiency because lower numerical precision and higher power consumption account for much of the improvement.