The expansion of AI data centers in the United States is forcing a reevaluation of one of the most commonly used metrics for assessing new facilities: cost per megawatt (MW). A report from First Call Group argues that comparing projects solely based on millions of dollars invested per MW can lead to misleading conclusions, because a traditional 480 V AC data center, a hybrid setup, and a future 800 V DC architecture can have very different costs, densities, cooling requirements, and deployment timelines.
The key points of AI data center cost per megawatt in 20 seconds
- JLL predicts a global average cost of $11.3 million per MW in 2026 for building the facility and basic infrastructure.
- In the U.S., major markets typically range between $10 million and $14 million per MW using the same methodology.
- Technology equipment for AI can add up to $25 million per MW.
- First Call Group estimates significant differences among AC, hybrid, and future DC architectures.
- Available electricity, cooling, density, supply chain, and construction timeline can be as important as the initial cost.
This discussion comes at a time when the scale of projects in the U.S. is rapidly changing. Uptime Institute identified 181,209 MW of power associated with large data centers over 100 MW announced solely during 2025, roughly double the amount from the previous year. North America accounted for about 80% of this new capacity. However, Uptime warns that a substantial portion of announced projects may never be fully built with the initial declared power.
The driving force is primarily AI. The Lawrence Berkeley National Laboratory estimated that U.S. data centers consumed around 176 TWh of electricity in 2023, accounting for 4.4% of national consumption. By 2028, they project a very wide range—between 325 and 580 TWh—equivalent to approximately 6.7% to 12% of all electricity in the U.S., depending on accelerator growth, utilization, and cooling efficiency.
With projects reaching hundreds of megawatts and even multiple gigawatts, a seemingly small difference in unit cost can amount to hundreds of millions or billions of dollars. But it also increases the risk of misusing that figure out of context.
A megawatt of data center capacity no longer always represents the same product
The issue begins with a fundamental question: what exactly is included in the cost per MW.
JLL estimates that the average global construction cost of a data center will reach $11.3 million per MW in 2026, a 6% increase from 2025. Between 2020 and 2025, costs rose from $7.7 million to $10.7 million per MW.
However, this figure pertains to a specific installation: a single-tenant, 50 MW air-cooled data center. It essentially includes shell and core, that is, the building and associated infrastructure, but excludes land and active IT equipment.
In the U.S., the same methodology shows significant regional differences. JLL places Chicago between $12 million and $14 million per MW, Northern Virginia between $11 million and $12 million, and Phoenix, Dallas, and Atlanta roughly between $10 million and $11 million. A liquid-cooled facility can add around 10% to these construction costs.
This highlights the first major challenge in comparing a conventional facility with an AI factory.
JLL estimates that the additional technology equipment supplied later by the operator can reach up to $25 million per MW in AI infrastructure. This can include servers, accelerators, networking, and other components that are not typically counted in traditional construction metrics.
Therefore, a project may be correctly announced as an $11 million per MW data center but ultimately require well over $30 million per MW when IT infrastructure is added.
There is no contradiction. Different things are being measured.
First Call Group emphasizes the electrical architecture
The report Cost per Megawatt in the AI Factory Era by First Call Group takes this critique further.
Their argument is that even two figures that seem to include comparable electrical infrastructure can conceal architectures that are completely different.
The analysis references four configurations for the U.S. in 2026 and estimates installed costs of $13-14 million per MW for a traditional 480 V AC architecture with double conversion, between $15 million and $17 million for a hybrid combining 480 V AC and ±400 V DC, $16-$18 million for direct bipolar ±400 V DC distribution, and $18-$20 million or more for architectures using 800 V DC with solid-state transformers (SST).
These figures should be treated as estimates from First Call Group itself, not as a general U.S. market rate. The industry is still undergoing technological transition, and actual costs depend heavily on scope, redundancy, location, and suppliers.
Furthermore, initially higher costs for 800 V DC architectures do not necessarily mean they will be more expensive over their entire lifecycle.
The interest in 800 V DC stems from efforts to simplify the electrical chain powering increasingly dense racks.
Conventional architectures require multiple conversions from incoming AC power to processors. Each conversion adds equipment, space, and losses.
Future DC systems aim to reduce some of these stages and deliver more power with less current, thus decreasing losses and copper requirements.
However, reaching this point requires new equipment, protection systems, busways, standards, qualified suppliers, and compatible regulation.
Hence, initial costs may increase before any potential economies of scale appear.
The shift to 800 V DC is not yet widespread
NVIDIA is among the companies trying to accelerate this transition.
The manufacturer plans to use 800 V DC in next-generation AI factories, especially as racks evolve toward hundreds of kilowatts and eventually over 1 MW.
The goal is not just powering more powerful GPUs but also moving vast amounts of electricity within a facility without multiplying cables, transformers, and conversion equipment endlessly.
But 800 V DC should not yet be considered the standard architecture for current data centers.
Deploying this industry-wide requires mature solid-state transformers, DC distribution systems, protection solutions, connectors, energy storage, certifications, and supply chains.
First Call Group advocates that cost per MW should be accompanied by what they call its “architecture envelope”, which is a description of all the factors needed to truly understand what has been built.
Variables in this envelope include the voltage used, number of conversion stages, rack density, cooling method, redundancy level, UPS or battery technology, supplier availability, and regulatory requirements.
Two 100 MW projects may end up being technically very different products, even if both are marketed as AI data centers.
Cooling also impacts the calculation
Electricity cannot be analyzed in isolation from cooling.
A traditional data center with a few kilowatt racks could primarily rely on air cooling. But new AI servers are increasing density to levels where direct liquid cooling of chips is becoming common for specific platforms.
Changing the thermal system also modifies the electrical infrastructure, internal distribution, and auxiliary space requirements.
This explains why the cost per MW becomes less comparable when one project uses air cooling while another is designed from the start for liquid cooling.
Furthermore, increased density may produce a seemingly contradictory effect.
A much more powerful rack requires pipes, refrigerant distribution units (CDUs), heat exchangers, and more demanding electrical systems. But at the same time, it can allow more computing power in less space.
The cost per MW might rise while the cost per unit of useful computational power drops.
For an AI factory, this second metric can be more relevant.
Time to get electricity also has a cost
Another cost that doesn’t usually appear in dollar-per-MW tables is time.
JLL now considers the speed of obtaining electric supply the primary criterion for data center location selection.
It’s easy to see why.
A $1 billion facility that can start operations twelve months earlier than a cheaper one can have significant economic advantages if it begins generating revenue while the other is still awaiting grid connection.
The International Energy Agency (IEA) estimates that the U.S. will add over 420 TWh of electricity consumption over the next five years, with roughly half coming from data centers.
Growth is also geographically concentrated, making it harder to absorb than dispersed consumption across millions of users.
The IEA projects that U.S. data center electricity use will increase by about 240 TWh between 2024 and 2030—a 130% increase, with the U.S. and China accounting for about 80% of the global rise.
For operators, this makes substations, transformers, transmission lines, turbines, and permits economic variables.
A cheap land without a ready electrical connection can be much more expensive in the long run than a pricier location with assured power capacity.
Gigawatt-scale projects change financial rules
The concept of AI factories is also causing some facilities to resemble large energy or industrial projects more than traditional data centers.
A 1 GW campus equals a thousand megawatts.
At a construction cost of $12 million per MW, a simple extrapolation would suggest an infrastructure cost of $12 billion solely for the physical components at that scale. Adding AI technology components can increase total investment well beyond that.
Such extrapolations shouldn’t be taken as actual budgets, since economies of scale, construction phases, shared infrastructure, and scope differences all influence costs. Nonetheless, they highlight why initial architecture decisions are so critical.
Uptime Institute tracks over 350 announced projects since 2021 exceeding 100 MW. In 2025 proposals, North America alone accounts for 144,411 MW, showing the potential size of the pipeline—even if Uptime considers it unlikely all will be fully built.
At this scale, choosing an architecture that increases or reduces costs by 10% is a significant decision.
Cost, speed, and efficiency: a combined approach
Perhaps the most valuable insight from First Call Group’s approach is not declaring a winning electrical architecture—that’s still up to the market.
Instead, it’s questioning whether a single number can adequately capture the economics of an AI factory.
Cost per MW remains useful for comparing projects when using identical assumptions.
It helps compare two 50 MW air-cooled data centers with similar redundancy and included costs.
But becomes less precise when comparing projects with liquid cooling, batteries, substation integration, or technology scope differences.
In new AI campuses, multiple metrics may be necessary simultaneously.
Beyond cost per installed MW, it’s also crucial to know how long it takes to make that MW available and how much useful computing it can generate over the asset’s lifetime.
Indicators such as tokens per second, tokens per watt, cost per token, accelerator utilization, and availability come into play.
A more expensive data center might ultimately be more cost-effective if it comes online earlier, maintains higher accelerator utilization, and consumes less electricity per inference.
Conversely, a technically advanced architecture may be a poor investment if it relies on hard-to-source components, delays construction, or introduces operational risks that are not yet resolved.
Thus, for large U.S. operators, the question shifts from simply “how much does each megawatt cost?” to:
How much does it really cost to have a megawatt capable of being installed, cooled, connected, maintained, and used for years as sellable computing capacity?
Frequently Asked Questions
How much does it cost to build a data center per MW in the U.S. in 2026?
JLL estimates that major U.S. markets range between $10 million and $14 million per MW for a 50 MW air-cooled data center. This calculation covers shell and core construction and excludes land and active IT equipment.
What could a MW prepared for AI cost?
There is no single figure. JLL notes that AI data center technology equipment can add up to $25 million per MW to the building and infrastructure costs.
Is using 800 V DC cheaper than a traditional AC architecture?
Not yet applicable universally. First Call Group estimates initial CAPEX for 800 V DC architectures to be higher due to new equipment and less mature supply chains. Its potential advantage lies in reducing conversions, losses, and certain operational costs as the technology matures.
Why can cost per MW be misleading?
Because two projects might include very different elements. Voltage level, cooling method, redundancy, rack density, substations, batteries, external site works, and technology equipment can all significantly alter budgets, even if expressed in dollars per MW.
