NVIDIA’s Vera Rubin NVL72 Compute Tray Claims Cable-Free Assembly in Under One Minute

Picture a server tray the size of a large kitchen worktop, packed with 72 of NVIDIA’s most powerful GPUs, sliding into a rack with no cables to plug in, no hoses to connect, and no fans whirring inside. According to NVIDIA, that tray is ready to go in sixty seconds flat.

That’s the claim the company posted this week about its Vera Rubin NVL72 compute tray — and it’s a striking one, even by the standards of an industry that rarely undersells itself.

What Is the Vera Rubin NVL72?

The NVL72 is NVIDIA’s rack-scale AI compute platform, built around the company’s new Vera and Rubin chip generations. Each rack holds 18 compute trays and 9 NVLink switch trays. Inside every compute tray sit 72 Rubin GPUs and 36 Vera CPUs, alongside ConnectX-9 SuperNICs and BlueField-4 DPUs — a dense stack of silicon designed to run the kind of large-scale AI workloads that data centres are scrambling to support.

For their part, the system sits on NVIDIA’s third-generation MGX NVL72 rack design, which the company says now has over 80 ecosystem partners building compatible hardware around it. That’s a broad manufacturing base, and it matters: the more partners building to the same standard, the faster customers can actually get systems deployed and generating a return.

Cooling is handled entirely by liquid, with NVIDIA specifying a 45°C inlet temperature for the liquid-cooling system. There are no fans in the compute tray itself. The idea is that removing air cooling from the tray reduces complexity, cuts points of failure, and makes the whole assembly easier to service.

The Cable-Free Design

The headline engineering claim is the cable-free modular layout. Traditional high-density AI server racks are a tangle — power cables, data cables, cooling hoses, all of which have to be connected by hand, tested, and managed over the life of the system. A single misconnected cable can cause hours of downtime in a production environment.

NVIDIA says the NVL72 tray sidesteps this entirely. The design uses a midplane-based modular layout, meaning components connect through the rack’s internal structure rather than through external cabling. Slide the tray in, and the connections are made. No plugging. No routing. No cable management.

Jensen Huang, NVIDIA’s chief executive, has repeatedly framed the company’s AI infrastructure products around the concept of the “AI factory” — a data centre that runs continuously, generating revenue like a production line. The NVL72’s design philosophy fits squarely into that framing. Faster assembly means faster deployment. Faster deployment means the customer starts earning sooner.

The One-Minute Claim

One minute. That’s what NVIDIA says it takes to assemble the compute tray, and it’s the number that will raise eyebrows.

To be clear: this is NVIDIA’s own claim, and it hasn’t been independently verified by third parties in publicly available testing. That doesn’t make it false — manufacturing automation has come a long way, and a modular, cable-free design genuinely does lend itself to rapid assembly. But until someone outside NVIDIA times it with a stopwatch, it’s a vendor claim, and should be read as one.

Third-party analysts and hardware commentators have broadly accepted the cable-free, liquid-cooled architecture as a genuine engineering step forward. The assembly time figure, however, has attracted less scrutiny than it perhaps deserves.

Vera Rubin Follows Grace Blackwell

The Vera Rubin generation succeeds NVIDIA’s Grace Blackwell platform — itself only recently deployed at scale — and represents the company’s continued push to make AI infrastructure faster to build, easier to run, and more reliable over time.

NVIDIA’s own blog describes the compute tray as designed for “density, reliability and ease of operation.” The MGX ecosystem, with its 80-plus partners, is central to that ambition. The more the rack design is standardised, the more competition there is among suppliers, which in theory should help on price and availability — though NVIDIA’s AI hardware has hardly been cheap.

The NVL72 is positioned as the compute engine for what NVIDIA calls AI factories: large-scale facilities where AI models are trained and inference is run at industrial scale. Demand for that kind of infrastructure has been intense, and shows no sign of slowing.

What This Means for Kent Residents

There’s no direct Kent-specific impact from this announcement — no local data centre, NHS contract, or council deployment is linked to the Vera Rubin NVL72 at this stage. The broader picture for UK consumers is that faster, more efficient AI infrastructure tends to feed through over time into the cloud services and AI tools that people use every day, from productivity software to online search — though the timeline between a new server tray and a cheaper or faster consumer product is rarely short.