The buildout of AI data centers in the United States is forcing a rethink of one of the most-used metrics for sizing up new sites: cost per megawatt (MW). A report from First Call Group argues that comparing projects on dollars per MW alone can mislead, because a traditional 480 V AC data center, a hybrid setup, and a future 800 V DC architecture can differ sharply in cost, density, cooling, and how long they take to build.
AI data center cost per megawatt in 20 seconds
- JLL expects a global average of $11.3 million per MW in 2026 for the building and basic infrastructure.
- In the U.S., major markets typically run between $10 million and $14 million per MW under the same method.
- AI technology equipment can add up to $25 million per MW.
- First Call Group estimates big differences among AC, hybrid, and future DC architectures.
- Available electricity, cooling, density, supply chain, and construction time can matter as much as the upfront cost.
The debate lands as project sizes in the U.S. change fast. Uptime Institute counted 181,209 MW of power tied to large data centers over 100 MW announced during 2025 alone, roughly double the year before. North America made up about 80% of that new capacity. Uptime does warn that many announced projects may never get built at their declared power.
AI is the main driver. Lawrence Berkeley National Laboratory estimated U.S. data centers used about 176 TWh of electricity in 2023, 4.4% of national consumption. By 2028 it projects a wide range, 325 to 580 TWh, or roughly 6.7% to 12% of all U.S. electricity, depending on accelerator growth, utilization, and cooling efficiency.
With projects reaching hundreds of megawatts and even several gigawatts, a small difference in unit cost can add up to hundreds of millions or billions of dollars. It also raises the risk of using that figure out of context.
A megawatt of data center capacity isn’t always the same product
It starts with a basic question: what’s actually included in cost per MW.
JLL estimates the average global build cost will reach $11.3 million per MW in 2026, up 6% from 2025. Between 2020 and 2025, costs climbed from $7.7 million to $10.7 million per MW.
But that figure describes a specific site: a single-tenant, 50 MW air-cooled data center. It covers shell and core, the building and its infrastructure, and leaves out land and active IT equipment.
In the U.S., the same method shows big regional gaps. JLL puts Chicago at $12 million to $14 million per MW, Northern Virginia at $11 million to $12 million, and Phoenix, Dallas, and Atlanta at roughly $10 million to $11 million. A liquid-cooled site can add about 10% to those build costs.
That points to the first big problem in comparing a conventional site with an AI factory.
JLL estimates the extra technology equipment the operator installs later can reach up to $25 million per MW in AI infrastructure: servers, accelerators, networking, and other parts that traditional construction metrics don’t usually count.
So a project can be fairly announced as an $11 million per MW data center yet end up well past $30 million per MW once the IT gear is in.
There’s no contradiction. They’re measuring different things.
First Call Group puts the spotlight on electrical architecture
The report Cost per Megawatt in the AI Factory Era from First Call Group pushes the critique further.
Its point: even two numbers that look like they cover comparable electrical infrastructure can hide completely different architectures.
The analysis lays out four 2026 U.S. configurations and estimates installed costs of $13-14 million per MW for a traditional 480 V AC architecture with double conversion, $15 million to $17 million for a hybrid mixing 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).
Treat these as First Call Group’s own estimates, not a general U.S. market rate. The industry is still mid-transition, and real costs hinge on scope, redundancy, location, and suppliers.
And higher upfront costs for 800 V DC don’t necessarily mean it’s more expensive over the whole lifecycle.
The interest in 800 V DC comes from trying to simplify the electrical chain that feeds ever denser racks.
Conventional architectures need several conversions from incoming AC power down to the processors. Each conversion adds equipment, space, and losses.
Future DC systems aim to cut some of those stages and deliver more power at less current, which lowers losses and copper use.
Getting there takes new equipment, protection systems, busways, standards, qualified suppliers, and compatible regulation.
So upfront costs can rise before any economies of scale show up.
The move to 800 V DC isn’t widespread yet
NVIDIA is one of the companies trying to speed up this shift.
It plans to use 800 V DC in next-generation AI factories, especially as racks head toward hundreds of kilowatts and eventually more than 1 MW.
The goal isn’t just to power stronger GPUs but to move vast amounts of electricity inside a site without endlessly multiplying cables, transformers, and conversion gear.
Still, 800 V DC isn’t the standard architecture for today’s data centers.
Rolling it out industry-wide needs mature solid-state transformers, DC distribution, protection solutions, connectors, energy storage, certifications, and supply chains.
First Call Group argues that cost per MW should come with what it calls an “architecture envelope,” a description of all the factors you need to understand what was actually built.
That envelope includes the voltage, the number of conversion stages, rack density, cooling method, redundancy level, UPS or battery technology, supplier availability, and regulatory requirements.
Two 100 MW projects can turn out to be very different products, even when both are sold as AI data centers.
Cooling changes the math too
You can’t look at electricity apart from cooling.
A traditional data center with a few kilowatts per rack could mostly rely on air cooling. New AI servers push density to the point where direct liquid cooling of the chips is becoming common on specific platforms.
Change the thermal system and you change the electrical infrastructure, the internal distribution, and the auxiliary space needed.
That’s why cost per MW gets less comparable when one project uses air cooling and another is designed for liquid from the start.
Higher density also brings a result that seems contradictory.
A much more powerful rack needs pipes, coolant distribution units (CDUs), heat exchangers, and more demanding electrical systems. At the same time, it can pack more compute into less space.
Cost per MW can go up while the cost per unit of useful compute goes down.
For an AI factory, that second metric can matter more.
Time to get power carries a cost too
Another cost that rarely shows up in dollar-per-MW tables is time.
JLL now treats the speed of securing electric supply as the top factor in choosing a data center location.
It’s easy to see why.
A $1 billion site that can start operating a year earlier than a cheaper one can hold a big economic edge if it’s already earning revenue while the other still waits on a grid connection.
The International Energy Agency (IEA) estimates the U.S. will add more than 420 TWh of electricity consumption over the next five years, with about half of it from data centers.
That growth is also geographically concentrated, which makes it harder to absorb than demand spread across millions of users.
The IEA projects U.S. data center electricity use will rise about 240 TWh between 2024 and 2030, a 130% jump, with the U.S. and China accounting for roughly 80% of the global increase.
For operators, that turns substations, transformers, transmission lines, turbines, and permits into economic variables.
Cheap land without a ready power connection can cost far more in the long run than a pricier site with assured capacity.
Gigawatt-scale projects change the financial rules
The AI factory idea is also making some sites look more like large energy or industrial projects than traditional data centers.
A 1 GW campus equals a thousand megawatts.
At a build cost of $12 million per MW, a simple extrapolation suggests $12 billion just for the physical parts at that scale. Adding AI technology can push total investment well past that.
Don’t read those extrapolations as real budgets, since economies of scale, construction phases, shared infrastructure, and scope differences all shift the cost. But they show why early architecture choices are so important.
Uptime Institute tracks more than 350 announced projects over 100 MW since 2021. In 2025 proposals, North America alone accounts for 144,411 MW, which shows the size of the pipeline, even if Uptime doubts all of it gets built.
At this scale, an architecture that moves costs 10% up or down is a major decision.
Cost, speed, and efficiency together
Maybe the most useful part of First Call Group’s approach isn’t crowning a winning electrical architecture. The market will settle that.
It’s asking whether a single number can capture the economics of an AI factory.
Cost per MW still helps when you compare projects on identical assumptions.
It works for two 50 MW air-cooled data centers with similar redundancy and included costs.
It gets shakier when the projects differ in liquid cooling, batteries, substation integration, or technology scope.
New AI campuses may need several metrics at once.
Beyond cost per installed MW, it matters how long it takes to make that MW available and how much useful compute it can produce over the asset’s life.
Indicators like tokens per second, tokens per watt, cost per token, accelerator utilization, and availability come into play.
A pricier data center can end up cheaper overall if it comes online sooner, keeps accelerators busier, and uses less electricity per inference.
On the other hand, an advanced architecture can be a bad investment if it depends on hard-to-source parts, delays construction, or brings operational risks that aren’t sorted out yet.
So for large U.S. operators, the question moves from “how much does each megawatt cost?” to something bigger: how much does it really cost to have a megawatt that can be 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 major U.S. markets run between $10 million and $14 million per MW for a 50 MW air-cooled data center. That covers shell and core construction and excludes land and active IT equipment.
What could a MW built for AI cost?
There’s no single figure. JLL notes that AI data center technology equipment can add up to $25 million per MW on top of the building and infrastructure costs.
Is 800 V DC cheaper than a traditional AC architecture?
Not universally, not yet. First Call Group estimates the upfront CAPEX for 800 V DC architectures is higher, due to new equipment and less mature supply chains. Its potential advantage is fewer conversions, lower losses, and certain operational savings as the technology matures.
Why can cost per MW mislead?
Because two projects can include very different things. Voltage level, cooling method, redundancy, rack density, substations, batteries, external site works, and technology equipment can all swing budgets, even when they’re all expressed in dollars per MW.
