SambaNova used a Computex 2026 session on June 4 to make its most public case yet that the GPU-only approach to AI inference is hitting a fundamental wall — and to demonstrate, live on stage, an alternative architecture it calls disaggregated inference running in a production data center.