On a quiet Sunday, Serenity drops a one-liner with a big read-through: there is nothing quite like the scent of more AI capex, and there are probably some ways to go if leading frontier models are out of capacity from too much demand. The frame is simple. If the best AI models are turning away demand because they have run out of compute, the bottleneck is not appetite, it is capacity. And the only cure for a capacity shortage is more spending on chips, data centers and power. The timing lines up with the news flow: hyperscalers are steering toward roughly $700B of AI capex in 2026, and reports this week say Meta is in talks to lease about $10B of compute to Anthropic, a rival literally renting a rival's infrastructure because nobody has enough. Serenity's implicit bet is that the capex cycle has further to run, which is bullish for the whole AI-infrastructure stack: GPUs, optical interconnects, memory, and the neoclouds that rent capacity out. The counter-case is that demand-driven spending eventually meets a digestion phase, but as long as frontier labs stay capacity-constrained, the buildout keeps feeding itself.
Why it matters · When frontier models turn away demand for lack of compute, the bottleneck is capacity, and the only fix is more capex, which flows straight to the AI-infrastructure chain.
Worth asking · Capacity-constrained AI demand, a durable capex supercycle, or a bubble waiting to digest?