AMD is neck-and-neck with Nvidia on inference performance while offering lower total cost of ownership, making deals with Anthropic, Meta, and OpenAI economically rational.

SambaNova's SN50 chip enables 1-2,000 tokens/sec on frontier models with 10T parameters on a single air-cooled 20kW rack, delivering the speed and efficiency General Compute needs…

Nvidia's healthy margins and circular financing deals subsidize demand but its Vera Rubin/Groq architecture requires 80 back-and-forth trips per token for large models, making it…

Groq's attention-FFN disaggregation forces 80 round-trips per token on 80-layer models, creating insurmountable latency for large model inference despite fast single-chip performa…