Your conclusion about training being the cost factor that will eventually align with profitability in the inference phases relies on training new models not being an endless arms race.