Inside the Dell Factory that Builds AI Factories

Have you ever wondered where the large-scale AI infrastructure is built? A few months ago, Michael Dell sent me a note saying that Dell makes AI systems and that we should take a look. A few months later, we ended up in Franklin, Massachusetts filming a tour of the enormous factory where Dell Technologies is building the large-scale AI factory racks that can scale to 100,000+ GPUs. While we were there, the company had several of its IR7000 NVIDIA GB200 NVL72 racks being assembled in a facility that has been operating for decades, but where cameras were never allowed into. That is, no cameras… until now.

This is one where you will want to watch the video:

We always suggest watching in its own tab, window, or app for the best viewing experience. Of course, we have to say this is being sponsored by Dell. This is also fun for history buffs as we also are showing the first public library in the United States.

Inside the Dell Factory that Makes AI Factories

Getting inside the factory, we only focused our cameras on certain sections that were related to making the AI systems, but this place is huge at roughly 700,000 square feet / 65,000 m^2 of manufacturing floors.

Our journey starts with the loading docks. To build high-end AI racks, you need many components. Something neat that we learned while there is that the loading dock doors had to be raised as we have transitioned from the era of 42U to 45U racks and into an AI era of taller 52U racks, sometimes with additional components installed on a cap.

Even the internal doorways of the facilities have had to increase in height to handle moving racks through the facility.

In that shipping and receiving area, the next step is removing components from their shipping packaging. Dell told us that it reuses or recycles all of the shipping materials.

One big area of reuse is that it keeps the packaging materials for the racks on-site. Racks are large, so they would generate a significant amount of waste if they were not re-used.

Instead, there is an area for storing these materials so the bare rack can enter the factory, get its components installed, tested, and then the packing materials are used to re-pack the completed racks as they are sent to customers.

Another area of reuse is the pallets. The fully loaded racks can weigh roughly 3000lbs or 1360kg. The systems are built on these pallets and transported around the factory. These are actually shock pallets have orange shock absorbers in them so that they can handle the stresses of moving these highly complex and highly valuable racks around the facility. As some frame of reference, a GB200 NVL72 rack currently sells for a few million dollars.

Aside from the racks, all of the components from the nodes, to the switches, to the liquid cooling CDUs, down to the power and networking cables need to be unboxed and brought to the assembly area.

We found a few racks that were in the early stages of being outfitted before making their way to the man assembly area.

The bare rack needs to be prepared with its basic infrastructure before the high-value components are installed.

That often means the racks need to be prepared and some of the base metal pieces are installed before assembly of the higher-value parts.

Something else that needs to be noted here is that not all of the Dell IR7000 racks, even the NVIDIA GB200 NVL72 versions, are exactly the same despite what you might see in marketing materials. Customers have different data center environments and capabilities as well as preferences. In the context of these large AI clusters that can often cost hundreds of millions and well into billions of dollars, customer needs need to be addressed beyond a simple one-size-fits-all solution. For example, if a customer wants to use a different rack vendor, power vendor, CDU, or other physical rack change, then those combinations can be built, tested, and prepared for a factory assembly at scale.

Next, let us get to assembling the systems.