# To Build a Data Center – Intro to Diesel Generation > Source: > Published: 2026-08-18 22:58:12+00:00 Everyone knows that data centers need some form of power generation in addition to GPUs. But there are a bunch of other parts, which vary wildly in price and scarcity. I'm going to walk through a full setup (close to a "turnkey data center"), including: - Rental generators - Fuel & delivery - Transformers - GPU racks & cooling systems - Networking ### Generation: Rental Units A data center can be used for training or for inference. Traditionally, training required hundreds of megawatts colocated, with gigawatts preferred. For inference, some neoclouds and inference providers are fine with a setup as small as 50 MW. Since inference requires less power, it's easier to build a data center for, so I'll use that as the reference point. Compared to other sources of generation, diesel units are easy to combine with each other. Each unit comes with multiple "ratings" (standby, prime, continuous), which specify how much power the unit can output under various conditions. In reality, the output of the generator is not constant, but we'll simplify and assume we're continuously outputting ~1-2 MW, though it varies depending on factors like temperature and altitude. So it'd take around 40 diesel units to hit our desired output of 50 MW. Maybe 1-2 of those are backup generators. Each unit is the size of a 20 foot shipping container and weighs ~20,000 pounds. The availability and cost of rental diesel units varies month-to-month. A rental diesel unit produces some number of megawatts, and it's rented for some period of months by vendors like Aggreko or United. An estimate for the cost of renting a diesel unit is somewhere around $40,000 per MW-month. So 50 MW of diesel generators costs around $2M/month, not counting fuel costs or anything else. Combining diesel units requires "paralleling switchgear," which synchronizes the generators and allows the GPUs to split their load across them. Fortunately, rental diesel generators come with this switchgear; otherwise it'd take a year to get them. ### Fueling the Generators To use the rental generators, we need to burn diesel. You can review the basics of heat rate calculations in Power 2026. How much diesel is required, and how much does it cost? The heat content of diesel is around 130,000 Btu/gallon at best (0.13 MMBtu/gallon), and a typical heat rate for a diesel generator is 9 MMBtu/MWh (though in reality, heat rates aren't constant). If we're producing 50 MW, a single day consumes the following: (50 MW * 24 hours) * (9 MMBtu / MWh) / (0.13 MMBtu/gallon) = 83,000 gallons/day, or ~2.5M gallons/month Diesel prices vary, but a defensible placeholder is around $5/gallon. Using the 2.5M gallons/month from above, that's $12.5M/month. At this level of consumption, it's wise to have a fuel supply contract in place. Note that generators are spread across several acres. One option to get the fuel from the fuel tank to the generators is to pay a guy to drive around all day and do it, a serious operation at this scale. Another option is an actual fuel delivery system. That includes pumps and smaller tanks that sit near each generator called "day tanks." When those day tanks get low, a level controller will fill them up from diesel sitting in a "bulk tank" where most of it is sitting. One detail worth mentioning at this point: diesel isn't shelf-stable. Bacteria and algae or whatever will grow if the diesel is just sitting around. There's usually some filtering system in place to prevent that. All-in, the fuel delivery system will cost around low six-figures/year, which is a rounding error compared to the diesel itself. You might have heard about transformers before, since there's a huge bottleneck in the supply chain right now. Thankfully our design doesn't require them. But for context: why are transformers ever needed? The 480 V power that immediately comes out of the generator is difficult to ship over hundreds of feet, requiring tremendous current. There's no way around this for a several hundred megawatt power plant, since it's not possible to fit all of the GPUs right near the generator. The way to handle it is with a step-up transformer to convert it to a voltage that's easier to transport, and a step-down transformer to convert the voltage back down to 480 V where the power is consumed. The advantage of small rental units is that you can create 10 blocks of 5 MW each called a "block design." Placing the GPUs near each generator block avoids transporting the power too far, so transformers are no longer needed. "Busways" are used to split the power within each 5 MW block to its GPUs. There aren't really bottlenecks to obtaining busways, and I'd ballpark the cost as low single-digit millions to install. There's another type of transformer called a "grounding transformer," which also isn't needed for this setup. Without getting into the details, there's a safety concept called "grounding" for generators. Rental generators already come with some sort of grounding system built-in. ### Powering the GPUs Once power is generated by the rental units, it's time to connect some GPUs. GPU racks constitute the bulk of the cost for the whole operation. #### How does GPU cooling work? Every MW consumed is eventually released as heat. Regular A/C uses air to transport the heat away. In contrast, modern GPUs need what's called liquid cooling directly on the chip. At a high-level, a metal plate sits right on top of the chip. The chip heats that plate up. A "coolant distribution unit" (CDU) controls the systems and pumps coolant through pipes that flow through the metal plates. Finally, the heat must be released from the coolant somehow. The "dry cooler" is a big fan that releases the heat into the atmosphere. The cooling system itself consumes some power, about 10% of the total power consumed (rule of thumb). So in reality, the GPUs consume more like 45 MW, and the last 5 MW is required to cool the system. CDUs are the biggest bottleneck among the cooling equipment. It's hard to get CDUs, but they'll often come bundled with a GPU rack configuration upon request, adding about 5-10% to the total order price. Getting a CDU without the rack is much harder. There's a more temporary way to set this up, but it has its own complexities. Vendors like United, Sunbelt, and Aggreko sell rental chillers. They are available pretty much immediately, and upper bound the cost at $20K/MW-month, so $1M/month for our 50 MW setup. The caveat is that a rental chiller alone won't work for GPUs that require liquid cooling, like GB300s. A CDU is still required, which is the tough part, so it's better to buy the CDU with the GPU rack. #### What's the going rate for GPU racks? There are many different options and configurations for GPU racks. One of the most common is the GB300 NVL72 rack, which comes with 72 Blackwell Ultras. The vendor will provide the anticipated power consumption per rack, which is around 150 kW. 50 MW of power is closer to 45 MW after cooling, which can support a maximum of 300 racks, or around 21K GPUs. One option for GPUs is buying them. I help to broker GPUs, so I can quote these directly. The fastest path to get a GB300 rack is around 2 months from now, and it'll cost ~$5.5M (depending on the configuration). It's cheaper to wait longer, and there are bulk discounts. B300s have a lower entrypoint and are faster to get, more like 1 month to delivery at $600K for an 8 GPU B300 server. So in short, 300 racks is around $1.65 billion. This is a capital expense, not a monthly operating expense, and the residual value of the GPUs on project completion will drastically impact the final economics. Another option is leasing. Lease arrangements are often customized and vary a lot based on the downpayment and lease term. ### Intranet and Internet One nice thing about the NVL72 racks is that the 72 GPUs come all connected to each other within each rack. The issue is that the racks don't come connected to each other. The cabling and switches are surprisingly expensive, around 10% of the entire cost of the GPUs. The good news is that NVIDIA has secured priority access to the scarce components, so if you order the cabling/switches alongside your GPUs, it'll all arrive at the same time. The GPUs also need to connect to the internet. This part doesn't really have any supply chain issues. It's just more of a constraint on where it's possible to build the data center. The idea is that you bury plastic tubes ("duct") underground and feed the fiber cable through it. There are different choices for exactly which fiber cable you use. The hard part is making sure there's actually a route where to install the fiber. Passing through a road requires a permit. Passing through a railroad? Permit. Wiring it overhead instead? Permit. A contractor can conduct a route survey and work out the details of exactly where the wire is going. As a placeholder for cost, I'll use $100K/mile, though actual construction costs vary. After installing the fiber, it's wise to run a bunch of tests to make sure the fiber works correctly and hand that over to whoever is renting the data center.