Part 5/11:
While Nvidia's H100 and other AI training chips dominate training workloads, Tesla believes the lion’s share of AI deployment—particularly in applications like autonomous vehicles and robots—is inference. Musk states AI5 will be among the best inference chips for models under 250 billion parameters. For larger models, such as ChatGPT or Gemini, AI6 will likely outperform competitors by being far more efficient and cost-effective.
Manufacturing and Massive Scale-Driven Economics
Tesla plans to have Samsung manufacture the AI6 chips at their massive Texas fab, with an initial deal valued at around $16.5 billion. Musk emphasizes that this figure is a bare minimum; actual production and deployment are likely to be several times higher—potentially reaching hundreds of billions in revenue.