I don't understand the logic around excluding the training costs and pointing to that as profitability. FOSS models are not that far behind, the day that OpenAI announces that they're done training models is the day that everyone focuses their energy on getting as close to SOTA output on free models.
These companies have to keep training because that's their moat, hyping the next big model is only thing keeping the market from asking the real questions about growth capability.
Training free models also costs money. As long as a dozen small companies can raise money to train models they give away for free, OpenAI can also do so, except they charge a pretty penny, so they can even throw in some of their own cash.
The world where OpenAI stops training because it's too expensive is a world where everyone else has given up even earlier.
> The world where OpenAI stops training because it's too expensive is a world where everyone else has given up even earlier.
With the commodification in this space, I don't see how that ever happens. The allure of free, local AI that competes with SOTA is enticing to all of the same people that are currently willing to pay for it. With governments wanting to take stakes in these companies and fears of widespread job loss, there's always going to be energy to pry these tools away from the few who currently control them.
Whether they succeed is a different question, but the pressure to keep improving will not stop anytime soon and will continue to force frontier labs to keep throwing money into the furnace. And for as long as that remains true, excluding training costs from expenses is disingenuous.
These companies have to keep training because that's their moat, hyping the next big model is only thing keeping the market from asking the real questions about growth capability.