Nvidia Unveils Major Performance Leap With Blackwell NVL72

Nvidia (NVDA) reported that its Blackwell NVL72 rack system can accelerate many of today’s most advanced artificial intelligence models by up to 10×, marking one of the largest performance jumps in its AI platform history.

10× Speed-Up on Leading MoE Models

According to Nvidia’s announcement:

The NVL72 system dramatically boosts performance for mixture-of-experts (MoE) models including:
- Kimi K2 Thinking
- DeepSeek-R1 (DEEPSEEK)
- Mistral Large 3
These models run 10× faster on GB200 NVL72 compared to the HGX H200 platform
Nvidia noted MoE models “mimic the efficiency of the human brain,” making them ideal for large-scale reasoning tasks

Inside the GB200 NVL72 System

Features 72 Grace Blackwell GPUs tightly interconnected
Designed to handle massive training and inference workloads
Entered mass production in October
Represents Nvidia’s flagship system for enterprise-scale AI and next-gen compute clusters

Geopolitics: Jensen Huang Meets President Trump

Separately, Nvidia CEO Jensen Huang met with President Trump to discuss export controls, as lawmakers raise concerns over advanced AI chips potentially reaching China.

Discussion comes amid heightened scrutiny of GPU exports
Nvidia shares were only fractionally lower in midday trading despite the geopolitical backdrop