Everyone says hosting open-source models is a money printer—until someone ran the numbers and found H200 can't break even

Whether hosting open-source models is a good business has always had two camps. One says the weights are free and you charge by token, so gross margins are astonishingly high; the other says compute costs are what they are, and you only find out who is swimming naked once you run it.

Yingjun Wu recently did the math on DeepSeek V4.1 Flash, and the conclusion is not very optimistic. By his estimate, hosting this model on H200, even at 100% utilization, revenue still cannot cover GPU rental. Switching to B300, the break-even point is around 35% utilization.

Unpack that sentence: on H200 it is not a question of earning more or less—it is a structural loss. B300 is better, but the premise is that you can stably run above 30% utilization. Idle cards are burning money; anyone doing inference services knows this.

Some background is needed. DeepSeek V4.1 Flash is the lightweight version in DeepSeek's open-source model line, focused on low inference cost. H200 and B300 are both Nvidia data center GPUs; the former is the mainstay of the Hopper architecture, while the latter is from the Blackwell Ultra generation. Renting these cards is not cheap, billed by the hour, and running at full load versus sitting idle are two different worlds.

So to the claim that "hosting open-source models prints money," Yingjun Wu's response is: Profitable, yes. Simple, no.

Several variables are easy to overlook here. First is the model's actual throughput. Although V4.1 Flash is a Flash version, batch size, sequence length, and concurrency during inference all affect per-card output. Second is pricing; the price war over token-based charges is fierce, squeezing gross margin. Third is utilization; traffic in production environments has peaks and troughs, and idle compute at 3 a.m. will not turn itself into money.

At the end of his tweet, Yingjun Wu asked: what are others seeing in production environments? That question is more interesting than the conclusion. The same card, running different models, taking different traffic, and priced differently, produces a completely different set of books. When someone says it can be profitable, it is probably true; when someone says it cannot break even, it is probably also true.

The difference is whether you have factored in the parts hidden underwater: idle rate, operations and maintenance costs, and depreciation.

Original tweet link: Yingjun Wu on X