NVIDIA Just Compressed the Internet Into 60 Pounds — Meet the Future of AI

At GTC Paris, NVIDIA unveiled a jaw-dropping innovation: NVL72, a new system that links 72 Blackwell GPUs through a groundbreaking MV-Link spine capable of 130 terabytes per second of bandwidth.

This “spine” is built entirely from copper coaxial cables, directly connecting all 144 Blackwell dies (packed in 72 GPU modules) with no data blocking. Every GPU talks to every other one. All at once.

Here’s the mind-blowing part: 130 TB/s is more than the peak of the entire global internet traffic. And all of that fits into a chassis that weighs just 60 pounds (27 kg).

Shrinking the Internet for AI

NVL72 isn’t just fast — it changes the game. It acts like a single, unified brain, letting massive AI models exchange data in real-time with virtually no latency.

This setup paves the way for next-gen AI workloads involving trillions of parameters, giving researchers and developers an unmatched speed boost.

In short, NVIDIA didn’t just build a machine — they’ve built the foundation for the future of AI infrastructure.