Market research firm Factorial Funds recently released a report stating that OpenAI Deploy the text-to-video model Sora, which is required during peak periods 720,000 NVIDIA H100 AI accelerator cards.
If we calculate based on the cost of $30,000 per Nvidia H100 AI accelerator card, 720,000 cards would cost $21.6 billion (currently about RMB 156.168 billion). And this is just the cost of the H100 accelerator card. Running it will inevitably consume a lot of electricity. Each H100 consumes 700W, so 720,000 cards would consume 504,000,000W.
Factorial Funds estimates that training Sora for one month would require 4,200 to 10,500 NVIDIA H100 AI accelerator cards. If a single H100 accelerator card is used, it would take 12 minutes to render a one-minute video.