Networking is the New Bottleneck for Mixture-of-Experts AI Workloads
| While Mixture-of-Experts (MoE) architectures have drastically reduced compute costs, they have exposed a critical networking bottleneck that GPU investment alone cannot fix.
Unlike the predictable, choreographed communication of dense models, MoE creates “improvisational” and unpredictable traffic patterns that often lead to significant GPU underutilization. |
|
 
|
This white paper explores how industry leaders like DeepSeek AI and Meta are already reporting that communication latency can account for up to 50% of training time or 30% of serving latency.
|
