No Dark Tokens

The dark fiber contrast

Responding to the worry that AI infrastructure spending will eventually turn into the kind of glut that followed the dot-com telecom buildout, Brad Gerstner argues the two situations are structurally different: "In 1999, 2000 we called it dark fiber for a reason. There was nobody using it, and we knew it when we put it in the ground. There's not a dark GPU in the world today. There's not a dark token in the world today."¹ The telecom buildout of that era laid fiber ahead of demand, and capacity sat unused for years, the canonical overbuild and glut. Gerstner's claim is that the AI buildout is consuming capacity as fast as it is built, and points to the tell on recent earnings calls, where every major operator reported being constrained by tokens rather than sitting on unused capacity, including a claim from one company that it would generate more revenue if it simply had more tokens to sell.

Supply as the binding constraint

The ceiling, in his account, is physical rather than a matter of demand: only so many memory wafers, logic wafers, and units of powered data center capacity exist in the world, which caps how many tokens can be produced at all. He contrasts this with the year 2000, when only roughly thirty five million people were on broadband, so demand itself was the limiting factor. Today, with several billion people connected, demand vastly exceeds the supply of tokens available to serve it, which is why he expects the buildout to feel close to exponential, citing a combined power footprint for two leading labs moving from roughly three gigawatts at the start of a year to around ten gigawatts by year end and a projected twenty gigawatts the year after.

Why it matters

The claim functions as the supply-side complement to the more general observation that demand for intelligence expands as it gets cheaper: together they describe not an overbuild but an under-build relative to demand. If every unit of current capacity is already spoken for, then constraining new data center construction does not trim a glut, it directly throttles growth. Gerstner's own fund is heavily concentrated in AI and compute, which is a reason to weigh the claim as directional rather than neutral, and the claim itself is a point-in-time observation rather than a guarantee that the next wave of new capacity stays fully absorbed once it comes online. A separate line of critique complicates the picture further: even tokens that are fully consumed are not necessarily useful ones, since a meaningful share of generated tokens may go toward unproductive or low-value output, meaning a world with no literally dark tokens could still contain a great many idle ones in effect, rationed by a hard supply wall while still being wasted on the demand side.