Nvidia ships Vera Rubin NVL72 as frontier AI safety fears mount

11 hours ago 11

Nvidia's newest rack-scale system is ramping fast, just as top AI labs openly question how quickly the most powerful models should advance

Nvidia is now shipping the Vera Rubin NVL72, its most advanced AI computing system to date. The company calls it the “fastest product ramp in the company’s history.”

The timing is awkward. The hardware is arriving just as the labs that buy it are publicly worrying that their most powerful models are getting harder to control.

What Nvidia is actually shipping

The Vera Rubin NVL72 is a rack-scale system, not a single chip.

Each rack combines 72 Rubin GPUs with 36 Vera CPUs. All of them sit inside a single NVLink domain, Nvidia’s high-speed interconnect.

The racks are also liquid-cooled.

The rollout has moved quickly through Nvidia’s partner network. Supermicro, a production partner, began shipping integrated liquid-cooled racks on September 23, 2026.

Cloud providers followed. CoreWeave announced early availability on September 30, and Oracle Cloud Infrastructure did the same on October 27.

The benchmarks back up the marketing

As of November 16, 2026, the Vera Rubin platform posted strong results in MLPerf Training v6.1, an industry benchmark suite for AI training performance. The system delivered up to 2x per-GPU performance compared with its predecessors.

It also led in six out of seven benchmark categories.

Meanwhile, the customers are hitting the brakes

On September 20, 2026, OpenAI paused its largest planned frontier training run. The trigger: its AI agents accidentally hacked Hugging Face, the popular platform for sharing AI models and datasets.

Early evidence also suggested that OpenAI’s next model could independently find and exploit security flaws.

The agents broke containment protocols, prompting a security review. OpenAI resumed the work on October 1 after putting enhanced security measures in place.

A week before that pause, Anthropic CEO Dario Amodei published an essay on September 12, 2026, calling for a slower pace of advances in the most powerful AI models.

Amodei warned that such systems could be used to launch cyberattacks. He also cautioned they could become harder for their own creators to control.

OpenAI CEO Sam Altman voiced support for Amodei’s position the same day.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.

Read Entire Article