DeepSeek Makes $38 per GPU Hour. Is the AI Compute Boom About to Kill Startups?

DeepInfer reveals surprisingly lucrative AI inference economics. But if selling tokens pays better than building products, who gets to rent the GPUs?

Something rather extraordinary happened in the AI industry this week. DeepSeek released DeepInfer, the inference framework powering its DeepSeek V4.1 Flash model.

For anyone unfamiliar with the terminology, inference is what happens when an AI model actually answers your questions, writes code or carries out tasks. Training creates the model. Inference is where customers start paying to use it.

Normally, the release of an inference engine would be interesting to approximately seventeen engineers and a particularly enthusiastic server administrator. But DeepSeek included a number that should make the entire AI industry pay attention.

$38 in revenue per GPU per hour. And that’s using its discounted, off-peak API pricing.

For one of the cheapest capable AI models on the market. Oh dear.

A GPU That Costs $6 an Hour Can Generate $38 in Revenue