Nvidia Says It Ships Over 1,000 Racks a Week; Hyperscalers Reportedly Deploy 72,000 Blackwell GPUs Weekly

Nvidia says it is shipping more than 1,000 MGX racks every week. The number came from the company's AI Infrastructure account on X.

The figure is worth attention because AI demand is usually counted in chips. A rack is where chips become working computing power: a cabinet of servers, networking gear and cooling equipment. MGX is Nvidia's modular design for those cabinets, built from standardized parts that partners assemble into different configurations.

The weekly numbers

Nvidia says its MGX ecosystem spans 350 factory sites across 30 countries, counting partners.

On the customer side, hyperscalers are reportedly deploying roughly 1,000 NVL72 racks a week. Those racks add up to 72,000 Blackwell GPUs every week. That figure is reported, not confirmed by the cloud companies themselves, and Nvidia has not broken it out by customer.

The pace reportedly follows the recent resolution of manufacturing and integration issues. No named source has laid out what those issues were or when they were fixed.

Nvidia describes MGX systems as engineered for power densities above 120 kW per rack, with room for future platforms such as Vera Rubin. The pitch to buyers: the frame outlives the chips inside it.

Rubin is already live at rivals

The next platform is moving fast too. Oracle made Vera Rubin NVL72 systems and Vera CPUs generally available on Oracle Cloud Infrastructure on October 27, 2026. Nvidia says Vera Rubin is in full production, and that the NVL72 ramp has reached five clouds.

Amazon is further back on the public timeline. On August 26, 2026, AWS and Nvidia announced that AWS will deploy 2 million additional Nvidia GPUs across its global infrastructure in 2027 and 2028. The two companies framed the build-out around agentic AI and physical AI, and around "AI factories" for enterprises, startups and governments.

The 2 million figure is not a Rubin-only order. AWS named Blackwell Ultra, Rubin and Rubin Ultra for the 2027 to 2028 window. It also sits on top of an earlier pledge, made at Nvidia GTC 2026, to add more than 1 million Nvidia GPUs starting in 2026. AWS said demand had run ahead of that earlier plan.

Amazon has not published a general availability date, instance type or price for any Rubin-based service.

Pods and power

In October 2026, Nvidia unveiled DSX Pod reference designs. Each pod is a shipping-container-sized unit built to house seven NVL72 racks. Nvidia says the design can cut deployment times from six months to as little as 14 days. Deliveries are expected in early 2027.

The constraint is not only factory output. Racks drawing more than 120 kW each need buildings that can supply and cool that load, and hardware can go live only as fast as the facilities allow. Nvidia's 350 partner sites can build to its designs, but the hardware can only be deployed as fast as customers' buildings can handle it.

Nvidia's weekly rack figure is the company's own number, announced by its own channel, and the 1,000-per-week NVL72 deployment pace is reported rather than confirmed by the hyperscalers.

What comes next

Three markers will test the claims. Whether the 1,000-rack weekly pace holds or climbs. Whether DSX Pod deliveries arrive in early 2027 as Nvidia expects. And whether AWS names a date and pricing for its first Rubin-based service, which Amazon has not yet published.