[Training a 200B LLM from Scratch] 11/25: How Do We Prepare a Data Center for These GPUs?

The procurement plan in Part 3 includes 264 servers. Now we will give those machines a specific home: choose a data center, put two servers in each cabinet, bring them in over two deliveries, and then demonstrate that they can work together continuously.

This article follows that example plan from start to finish. We will use Flexential's Atlanta-Douglasville 2 facility in Georgia as the site for design discussions and a quotation. We will start with an APC 42U cabinet, place 264 servers in 132 compute cabinets, and retain the earlier plan's 4.5 MW of IT capacity for the hall. Each choice below explains how the pieces fit together and how a delivery of servers progresses toward a working training system.

Choose a hall that can accommodate the whole system

Our 2,048 training GPUs need to exchange results frequently, so we will keep the main training cluster in one data hall. This makes equipment distances, installation, and later maintenance easier to coordinate.

Flexential's official Douglasville 2 specifications list 4.5 MW and approximately 24,000 square feet, or 2,230 square meters, per hall. That capacity matches the plan in Part 3. We will therefore build this example around a request to lease one hall for compute, networking, and storage, with the corresponding power and cooling services.

IT capacity: The electrical power a facility can supply to servers, networking, storage, and other computing equipment. Cooling and other facility systems consume additional power.

This gives us a specific operator and site to approach. An official announcement on July 8, 2026 still described Douglasville 2 as on track for completion. As of October 27, 2026, we have not obtained confirmation of available capacity or a delivery date for this plan. Before scheduling shipments, the delivery agreement will identify the hall, the date installation can begin, and the power and cooling conditions. The machines will then have a definite destination.

First, look at the whole building: yellow lines carry power to the cabinets, while blue and red arrows show cool air entering and heat leaving; the cabinet outlined in gold, Rack A, will be enlarged in the next illustration.

Why plan the entire hall now? Once the first delivery is installed, the second still has to join the same cluster. Reserving space and capacity for the complete system lets us expand into assigned positions. Next, we can look at the smallest installation unit in that hall: one cabinet.

Put two servers in a cabinet, then account for the remaining space

The Exxact TS4-142487900 selected in Part 3 is an eight-GPU B200 server. It is 8U high and its chassis is 930 mm deep. U measures the height of rack-mounted equipment; 8U is approximately 35.6 cm.

U: A unit of height for rack-mounted equipment. One U is approximately 4.45 cm.

We will use the APC NetShelter AR3350B2 as the candidate cabinet. It provides 42U of mounting space in an enclosure 750 mm wide and 1,200 mm deep. Our plan places two servers in the lower 16U, covers the remaining unused openings with blanking panels, and routes power and cables at the rear. Keeping heavy equipment low also helps with installation and maintenance.

Why fill only 16U of a 42U cabinet with servers? Power and heat set the limit. Using Part 3's planning allowance of 14.25 kW per server, two servers require 28.5 kW. If they sustain that power for an hour, the servers alone consume approximately 28.5 kWh and produce heat that must be removed continuously. Two servers per cabinet is the density we choose for this plan, making power, cooling, and maintenance easier to arrange at the cabinet level.

Kilowatt (kW): A unit of electrical power.

Kilowatt-hour (kWh): A unit of energy consumed. Operating at 28.5 kW for one hour consumes 28.5 kWh.

Our cooling example uses air-cooled equipment with a contained hot aisle. Cool air enters the front of the cabinet, passes through the servers, and carries heat out through the rear. Hot air collects in a return path for the facility's cooling system to remove. Blanking panels cover unused equipment openings to reduce hot air recirculation toward the front. The facility then designs cooling for at least the planned 28.5 kW heat load of each two-server cabinet, with engineering margin based on the final equipment requirements.

Hot aisle: The passage behind cabinets where hot exhaust air collects. Separating it from the cool-air side reduces mixing.

Putting this air-cooled plan into practice requires the equipment vendor to confirm that the server supports the specified cooling conditions. Exxact's current public page does not provide all airflow, server weight, and rail requirements, so the procurement documents must supply those details before the cabinet and cooling design are finalized together.

Now follow Rack A inward: one cabinet contains two servers, and opening one server reveals the eight GPU modules inside.

Planning illustrations of the facility and hardware hierarchy; internal equipment layouts are simplified.

On October 27, 2026, Exxact's configuration page displayed an estimated power of approximately 14.15 kW. We retain Part 3's 14.25 kW here to keep the planning basis consistent. The final electrical design will use confirmed requirements for the ordered configuration. Cabinet dimensions need the same care: a 1,200 mm enclosure depth does not make all of that space available for rails and cables. The integrator should complete one trial cabinet installation before repeating it across the others.

Which numbers should we check for the trial cabinet?

APC's manufacturer specifications for the AR3350B2 list an enclosure 1,991 mm high, 750 mm wide, and 1,200 mm deep, with a maximum mounting depth of 1,048 mm and a static load capacity of 1,814.4 kg. The Exxact chassis is 930 mm deep, but rail positions, plugs, and cable bend space still need to be checked against the actual installation. Total cabinet weight must include the cabinet, both servers, power distribution equipment, and cables, then be checked against floor and transport-route load limits.

The Exxact configurator lists 12 power supplies with 6+6 redundancy. Our plan asks the vendor for a wiring diagram that matches the facility's A and B power feeds, identifying the outlets and current ratings required on each feed and how the server operates if one feed fails. Power-supply ratings are used to design the power paths; they are not added together as normal power consumption.

Repeat the cabinet 132 times and place them on one floor plan

At two servers per cabinet, 264 servers require 132 compute cabinets. Of these, 128 cabinets hold 256 training servers, and four hold eight spares. To match the two deliveries, we divide the compute area into Zones A and B. Each zone has 66 cabinets containing 128 training servers and four spares: 64 training cabinets and two spare cabinets.

Our example layout gives each zone three rows of 22 cabinets. Networking and storage occupy a separate area nearby, with room for access and the second delivery's connections. The actual hall's columns, doors, and facility interfaces will determine the construction layout.

First, calculate the cabinet footprint to understand the scale:

132 × 0.75 m × 1.2 m = 118.8 square meters.

People need access, hot air needs a path out, and equipment must be removable from the cabinets. In a separate area sketch with six rows arranged one behind another, adding aisle space gives a compute area of approximately 345 square meters. This estimate helps establish how much space we need.

How do we get approximately 345 square meters?

Arrange six rows one behind another, with 22 cabinets in each row. Each row of cabinets is 22 × 0.75 = 16.5 m long. Allowing 1.5 m at each end gives a total length of 19.5 m. The six rows themselves occupy 6 × 1.2 = 7.2 m in depth. Seven aisles, between the rows and at the front and rear, each add 1.5 m, giving 17.7 m in total. The sketch therefore occupies 19.5 × 17.7 = 345.15 square meters.

The 1.5 m aisle width is a choice for this teaching sketch. The facility will determine final widths from maintenance, fire safety, and cooling requirements. These are neither the scale dimensions of the illustration above nor the operator's actual hall layout.

Beyond the compute area, we need room for networking, storage, installation staging, and facility systems. The hall's 2,230 square meters provides an area reference. The next step is to place these areas on the operator's actual drawing, accounting for columns, entrances, service access, and facility connections.

Once the equipment has positions, we can allocate power within the same plan:

Megawatt (MW): A unit of power. One MW equals 1,000 kW.

The 0.300 MW for networking and storage is an initial allocation in this plan. We will check it against the detailed equipment list. If that list exceeds the allocation, we adjust the distribution and site plan. Each part then has both a location and a power allowance, before installation reveals that a particular row needs more power than it can receive.

Cooling follows the same table. The compute cabinets contribute a planned heat load of 3.762 MW to the cooling design, with networking and storage heat included as well. The facility's own cooling equipment also consumes power, which is arranged on the facility side. The 3.3 MW used in Part 3's electricity calculation is an operating-load budget for estimating bills. Facility capacity, planned equipment power, and operating electricity costs answer different questions.

Install Zone A first, then connect Zone B

The first delivery now has a specific destination: the 66 cabinets in Zone A. At two servers per cabinet, that is 132 servers, including 128 for training and four spares. Zone B reserves the same 66-cabinet layout for the second delivery.

Before the first delivery arrives, Zone A's cabinets, power distribution, and cooling must be ready, along with the shared infrastructure needed to connect both zones. The network backbone and shared storage in Part 3's plan must also support the eventual combined system. When the second delivery arrives, the main job is connecting equipment in positions already reserved for it.

We retain Part 3's target sequence: receive the first delivery in month 4, the second in month 6, and complete full-scale acceptance in month 7. The equipment supplier and facility use this sequence to coordinate their schedules. Once the operator supplies an actual site-access date, the delivery schedule fixes shipment, installation, and acceptance dates for each batch together.

We rehearse equipment movement in the same order. First, take one server through unloading, doors and elevators, entry into Zone A, rack installation, and power connection. Confirm that tools, rails, and working space are suitable, then proceed with the remaining equipment. Installed cabinets also need clear identifiers. For example, the first cabinet in the first row of Zone A can be labeled A-R1-C01, with the serial number and position of each server recorded. Maintenance staff can then locate a specific machine when an alert arrives.

Redundancy: Extra capacity or an alternative path prepared for a failure. For example, power feed B can carry the agreed load when feed A fails.

Each cabinet's two power feeds follow the equipment vendor's confirmed wiring arrangement, and the hall follows the operator's failure-response design. Acceptance tests verify that equipment meets the agreed operating target after a feed fails. Two feeds provide an alternative path during a failure; normal electricity consumption still follows the equipment's actual load.

Use an acceptance example to turn “installed” into “ready to operate”

When this equipment is handed to the training team, one of the most useful deliverables is an acceptance record with clear decisions for each item. Here is the example used in our plan:

Specific temperature and performance thresholds are agreed for this equipment in the procurement and integration contracts and entered into the test sheet before delivery. Staff on site then have explicit pass criteria. A failed item can be traced to the relevant equipment, circuit, or service, repaired, and tested again.

NVIDIA's DGX B200 guide also places health checks and load tests before operational use. Our acceptance tools will come from the Exxact equipment vendor, with records collected in stages from one server to one cabinet, the first delivery, and the complete cluster.

The resulting plan is concrete: a 4.5 MW hall at Douglasville 2 as the site to discuss with the operator, 132 two-server compute cabinets, staged installation in Zones A and B, and a separate area for networking and storage. Power and heat removal are designed from the same equipment list, and acceptance follows one set of records. Taking this example to the operator and equipment supplier gives both sides a specific basis for turning equipment placement, capacity, dates, and pass criteria into an engineering plan they can deliver.