Is running GPT-5 profitable? * Gemini is doing 10^15 tokens per month * Suppose OpenAI is 2x, that's ~0.8 billion per second * B200 chip can prob output over 3000 tokens/sec * So you need 310k B200s (at 80% util) * At $2.5/hour, that's $560m/month * Add 50% for staff etc, $840m/month * OpenAI's revenue is $1.6bn/month There's uncertainty, but it lines up with other estimates I've seen – running frontier models seems highly profitable. OpenAI is only loss making because 70% of the compute is invested in developing GPT-6. The key question is whether those investments will pay off.