Training costs are not the same as inference costs. DeepSeek (or anyone hosting DS largest model) will still need a lot of money and a bunch of GPU clusters to serve the customers.