It does have a large initial cost. It also has a large ongoing cost. GPU time is really, really pricey.
Even putting aside training and infrastructure, OpenAI still loses money on even their most expensive paid subscribers. While guys like Deepseek have shown ways of reducing those costs, they’re still not enough to make these models profitable to run at the kind of workloads they’re intended to handle, and attempts to reduce their fallibility make them even more expensive, because they basically just involve running the model multiple times over.
It does have a large initial cost. It also has a large ongoing cost. GPU time is really, really pricey.
Even putting aside training and infrastructure, OpenAI still loses money on even their most expensive paid subscribers. While guys like Deepseek have shown ways of reducing those costs, they’re still not enough to make these models profitable to run at the kind of workloads they’re intended to handle, and attempts to reduce their fallibility make them even more expensive, because they basically just involve running the model multiple times over.