Ian King’s recent warning about hyperscaler spending landed with the thud of someone who has watched this movie before. His argument, roughly, is that the roughly $750 billion flowing into data centers from the likes of Microsoft, Amazon, Google, and Meta will end the way every chip boom has ended: badly, with excess capacity, write-downs, and a lot of expensive steel sitting idle. It is a sobering thesis, and it deserves more than a reflexive dismissal.
The pattern he is pointing to is real. Memory, fiber, networking gear, and compute capacity have all followed the same arc over the past four decades. Demand surges, capital floods in, supply overshoots, prices collapse, and the weakest players get washed out. The telecom bust of the early 2000s is the canonical example. Companies laid fiber at a frantic pace, convinced that internet traffic would double every few months forever. Traffic did grow, but not fast enough to fill the pipes that had already been dug, and a generation of equipment vendors and carriers paid the price.
So why might this cycle play out differently, or at least not identically? Three reasons keep coming up among analysts who track the space.
First, the demand signal is not speculative in the way fiber was. Hyperscalers are not building against a forecast of future users. They are building against contracts, internal workloads, and revenue that is already showing up in earnings. Cloud divisions at the major providers have been growing at rates that would have seemed implausible a decade ago, and AI training and inference workloads are consuming capacity as fast as it comes online. That does not mean the demand is infinite, but it is more grounded than the “if we build it, they will come” logic that powered earlier booms.
Second, the composition of the spending is different. A meaningful share of the $750 billion is going into custom silicon, advanced packaging, liquid cooling, and power infrastructure, not just generic server racks. That spending is harder to walk away from and slower to depreciate, which cuts both ways. It raises the exit cost if demand softens, but it also means the capacity is more specialized and less likely to become a commodity glut overnight.
Third, the buyers are among the most cash-rich companies in corporate history. Microsoft, Amazon, and Google can absorb a bad year or two of returns on invested capital in a way that a leveraged telecom startup could not. That does not make a poor investment wise. It does mean the shakeout, if it comes, may look more like a slowdown in spending growth than a wave of bankruptcies.
None of this is an argument for complacency. The bear case has teeth. If AI adoption stalls, if enterprise buyers pull back, or if a handful of large customers decide they have overbuilt, the correction could be sharp. Nvidia’s customers are concentrated, and any sign that they are digesting rather than expanding would ripple through the entire supply chain within a quarter or two. Memory makers, foundries, and equipment vendors would feel it first.
The more interesting question is not whether the boom ends, but how. King’s framing suggests a repeat of history. The specifics of this cycle suggest something messier: a plateau rather than a cliff, with winners and losers separated less by timing than by whether their capacity is genuinely useful when the music slows.
For investors and operators, the practical takeaway is to watch utilization, not headlines. If hyperscalers start extending depreciation schedules, delaying announced builds, or shifting capital toward power and cooling rather than raw compute, that is the signal worth paying attention to. The $750 billion figure is eye-catching, but the number that matters is how much of it converts into revenue-generating workloads before the next downturn arrives.
Every chip boom has ended. That part is not in dispute. What is up for grabs is whether this one ends with a bang, a shrug, or something in between.