Same GPU, 50% More Output: Why the Operator Matters — Oliver Kennon, SemiAnalysis

Buying GPUs gets you capacity. How much of it turns into useful work depends on who runs them. That is the core of this talk by Oliver Kennon of SemiAnalysis, as summarized in its YouTube description. No transcript was available, so the claims below are his as the description summarizes them. He treats compute as a commodity market in which GPU hours are the input and intelligence is the output. His headline example is that on the same Nvidia B200 chip, the serving stack alone can deliver about 50% more output. The serving stack is the software that schedules requests and runs the model on the hardware.

Kennon separates the price a provider charges for rented GPUs from what it costs to produce output on them, and says your position on the cost curve depends on your competitors, not just on you. He counts cluster security as part of whether the output is useful at all, and points to SemiAnalysis's ClusterMAX, which rates GPU cloud operators. Version 3.0 of those ratings is covered in an earlier Intel item on ClusterMAX 3.0.

The talk closes on what it takes to build credible compute grades and futures. In general, a future is a contract to buy at a set price later, and futures markets depend on a standardized grade of the thing traded, as with oil or wheat. According to the description, Kennon covers how to grade comparable compute and what compute futures could hedge. The description does not say which serving stacks or workloads are behind the 50% figure.