Ninety-Six Billion and the Trap That Catches Winners

Nvidia will hit $100 billion in quarterly revenue in the next few months. The company just reported $96. 2 billion, with data center revenue alone at $89 billion — numbers so dominant they sound permanent, the way Intel's 95% grip on server CPU market share looked permanent in 2010.

But dominance in infrastructure is only as durable as the workload it was built to serve. When that workload disappears, the dominance evaporates so quickly it's hard to believe it was ever real.

Intel owned the data center. Their processors were the only sensible choice for enterprise servers, their margins were stupendous, and their architectural advantage seemed unassailable. Then cloud computing arrived and everything changed. Cloud datacenters needed scale-out systems that ran thousands of cheap, simple servers instead of a handful of expensive ones. Intel's chips were over-engineered for that problem. AMD's architecture fit better. Within five years, AMD gained 25 points of market share. Intel went from untouchable to vulnerable.

Nvidia now faces the identical setup. Its H100 chips are optimized for the training workload that built this entire revenue tower — massive memory bandwidth, enormous floating-point throughput, designed for the assumption that you're moving gigantic matrices through the system. That assumption is correct when you're training models. It becomes questionable when you're running inference at scale. Is what's actually starting to happen now that enterprises have stopped experimenting and started deploying. Inference has different constraints. Inference needs latency, efficiency, and throughput per watt. Those are not the same optimization target. Nvidia now faces the same risk: if large language model inference moves away from the H100's memory-bandwidth assumptions toward different optimization targets, that quarter becomes yesterday's story.

The $100 billion number is real. The vulnerability is also real. The question that will define the next five years isn't whether Nvidia can hold its current share. It's whether holding your current share matters when the workload you optimized for stops being the one customers actually need. Intel learned the hard way. Dominance in infrastructure looks like permanence until the precise moment it doesn't.