Nvidia and Cerebras: Selling Performance Customers Will Likely Never See

At the Hot Chips conference, Nvidia announced its Groq-3-based LPX racks are now in production, achieving an impressive 3,400 tokens per second with Gemma 4 31B, four times faster than competitor Cerebras. However, both companies are essentially disputing figures that customers will likely never experience in real-world applications. The focus should be on how efficiently these systems can scale performance and their cost-effectiveness. Nvidia and Cerebras need to provide metrics that reflect actual usage scenarios rather than just top-line performance numbers.