There is a not-so-subtle danger of fast-paced and large-volume expansion of AI infrastructure. It creates a rubberband effect. Acceleration should not mean faster, it might mean higher volume, but not faster on an existing platform. Picture a cartoon character running forward with an enormous elastic band around his waist. His legs keep moving, and for a while he makes progress. But the farther he runs, the more tension accumulates behind him. Eventually, the force pulling backward can exceed his ability to move forward. The running continues right up to the moment he is thrown in the other direction.
Companies keep buying, deployment takes time, the competitive standard changes, and customers’ willingness to pay develops on its own schedule. The snapback happens when the remaining cash those chips can generate no longer supports the investment made in them (or it appears this way). This argument concerns silicon: the GPUs and TPUs used to train and run AI models.
My skepticism is focused on the scale and pace of the more recent commitments, particularly the roughly six quarters leading into late 2026. I question whether their economics can justify the expansion. Even where earlier silicon earned attractive returns, those results cannot establish what a much larger subsequent purchase will earn under different prices, workloads, and competitive conditions. The response I favor is calculated pacing. Every substantial new commitment should still be tested against credible paid demand, the performance of recently deployed capacity, realistic operating costs, and the time available to earn a return.
For leadership teams, the relevant questions are concrete. How much recent capacity is doing paid work? What prices are customers actually paying after discounts? How much cash remains after power and operating costs? How dependable are the customer commitments? What happens to the expected return if deployment is delayed, prices fall, macroeconomics change, or newer chips shorten the period of competitive advantage? These all create rubber-band effects.
Right now mt view is built on an increasing number of accelerated infrastructure build-outs, while the specific consumer and enterprise demand for hardware is not crystal clear. The next generation of silicon needs an economic case of its own, grounded in what customers will pay and what the equipment can earn during its useful competitive life.
The rubberband stretches when commitments keep compounding faster than that case develops. Running harder can conceal the tension for a while. It cannot make the tension disappear. China will snap first. That’s a warning for the United States.
