What an AI data center costs per megawatt and per gigawatt

Published estimates for the building, power and cooling of a modern data center run from roughly 11 to 18 million dollars per megawatt, which is 11 to 18 billion dollars per gigawatt, before any chips. Adding the computers themselves can multiply that several times. The spread comes from what each number counts.

This post walks through the main public figures, who produced them and when, what they include, and why two honest numbers can differ by more than half. It is a guide to reading the figures, not investment advice and not a forecast.

Key numbers at a glance

• Cushman & Wakefield, 2026 Data Center Development Cost Guide (trade coverage dated 20 August and 3 September 2026): about 17.6 million dollars per megawatt, all in, for the most modern greenfield facilities in the United States and Canada, excluding chips and GPUs.
• Cushman & Wakefield, same guide: construction cost per megawatt up an average of 21 percent since the previous edition, published in the fourth quarter of 2024.
• Cushman & Wakefield, same guide: power infrastructure is the largest cost category, at an average of 21 percent of total greenfield development cost.
• Cushman & Wakefield, same guide: powered land in primary US markets at 584,000 dollars per megawatt in 2026 so far, 35 percent above the five year average and 51 percent above a year earlier.
• JLL forecast as reported by Datacentres.com (12 May 2026): global average construction cost of 11.3 million dollars per megawatt for 2026, up 6 percent on the year and 47 percent above 2020 (7.7 million dollars).
• Turner & Townsend, Data Centre Construction Cost Index 2025 to 2026 (5 November 2025): cost inflation of 5.5 percent in 2025 for traditional data centers across 52 markets.
• Turner & Townsend, same index: a 7 to 10 percent construction cost premium for AI data centers over traditional ones in the United States, for similar IT capacity.
• Turner & Townsend, same index, per watt by market: Tokyo 15.2, Singapore 14.5, Zurich 14.2, Silicon Valley 13.3, London 12.0, Frankfurt 11.6 US dollars per watt.
• Turner & Townsend survey, same index: 48 percent of respondents named power availability the biggest obstacle to delivering projects on schedule.
• Meta and Blue Owl joint venture for the Hyperion campus in Louisiana (announced 21 October 2025): about 27 billion dollars in expected total development costs for buildings and long lived power, cooling and connectivity infrastructure, for a campus initially planned at 2 gigawatts.
• Reported by Bisnow (May 2025, Abilene campus): about 15 billion dollars of financing across two phases for a 1.2 gigawatt campus.
• CBRE, North America Data Center Trends H1 2026 (8 September 2026): 7,481 megawatts under construction in eight primary markets, up 24.8 percent year on year.

First, what a megawatt means here

A megawatt is a million watts. A gigawatt is a thousand megawatts. In data center talk, a megawatt is a measure of electrical capacity, not of floor space or of how many computers fit. A building described as 100 megawatts can draw up to about that much power at full load.

There are two ways to count. One is "IT load", the power that reaches the servers and networking gear. The other is "facility power" or "total power", which also includes cooling, lighting and losses in the power equipment. For the same building the second number is larger. Some of the sources below do not say which one they mean. The Epoch AI tracker of OpenAI's Stargate sites states that its capacities are total facility power unless noted, which is a helpful habit. When you read any per megawatt cost, ask which megawatt it divides by.

Dividing by a larger megawatt count gives a lower cost per megawatt for the same building. That alone can explain a few percent of difference between sources.

The three kinds of cost people mix together

Almost every confusing comparison comes from mixing three layers.

• The shell and its systems: land, the building, the electrical gear that brings power in and distributes it, cooling, fire protection, and the people and permits needed to build it. This is what most "cost per megawatt" figures describe.
• The grid connection and any on site generation: the substation, the lines, and sometimes gas turbines or batteries. Some studies fold this in. Others leave it out because the utility pays for part of it.
• The computers: GPUs or other accelerators, servers, networking and storage. These are bought and depreciated very differently from a building, and they are the largest dollar item in an AI build.

The Cushman & Wakefield figure states that it excludes chips and GPUs. That makes it a clean number for the first and part of the second layer. It does not tell you what a finished AI cluster costs.

The all in figure: about 17.6 million dollars per megawatt

Cushman & Wakefield published its 2026 Data Center Development Cost Guide in the late summer of 2026. According to the Business Wire release carried by Finviz, the average all in cost to develop the most modern greenfield facilities in the United States and Canada is about 17.6 million dollars per megawatt, excluding chips and GPUs. Greenfield means a new build on undeveloped land, as opposed to filling an existing building.

The same release says construction costs per megawatt rose an average of 21 percent since the previous edition, and that power infrastructure is the largest cost category at an average of 21 percent of the total. Contingencies, the money set aside for surprises, ranked second, though the public release gives no percentage for them. The firm's head of data center insights, John McWilliams, is quoted saying that access to power, land and skilled labor are increasingly determining where growth can occur.

Two details are worth noting. First, powered land, meaning land with a credible path to electricity, is priced at 584,000 dollars per megawatt in primary US markets in 2026 so far. That is the firm's figure for powered land in primary markets, and it is 51 percent higher than a year earlier. Land with power is now a scarce product, and its price is part of why the all in number has risen. Second, the guide covers primary, secondary, tertiary and frontier markets, and says Texas ranks among the lowest cost environments. The full report is sold separately, and the public release does not break the 17.6 million dollars into a full list of parts.

Applied to a gigawatt, 17.6 million dollars per megawatt is 17.6 billion dollars. That is simple arithmetic on the published average, not a figure the firm gave for a gigawatt campus, and individual projects will land above or below it.

The global average: 11.3 million dollars per megawatt

Datacentres.com Research, in an article dated 12 May 2026, reports a JLL forecast that global data center construction will average 11.3 million dollars per megawatt in 2026. It describes that as 6 percent above the prior year and 47 percent above 2020, when the figure was 7.7 million dollars, a compound annual growth rate of about 7 percent. It says Silicon Valley and Tokyo can exceed 15 million dollars per megawatt, while India and Southeast Asia stay below 8 million.

Treat this one with care. The article attributes the headline to JLL but is itself a secondary report, and it does not define what the 11.3 million dollars covers. It does not say whether land, fit out or soft costs such as design fees are included. It also quotes several figures with no named source, such as transformer prices and wage growth, and those are not repeated here.

Even so, the number is useful for one thing: it shows that a global average, which includes many cheaper markets and many older style buildings, sits well under the US all in figure for the newest facilities. The two are not in conflict. They measure different building types in different places with different definitions.

A second Datacentres.com page gives a rough per megawatt breakdown for a standard build: shell and core about 3 to 5 million dollars, mechanical and cooling 2 to 4 million, electrical distribution 3 to 5 million, and fire suppression and controls 0.5 to 1 million. It adds land at 50,000 to 500,000 dollars per acre, utility interconnection from 1 million to more than 10 million dollars, and permitting from 200,000 dollars to 2 million. That page gives no sources for the ranges, so read them as an orientation to the shape of the budget, not as quotes. The shape is consistent with the Cushman finding that electrical and power items weigh heavily.

The index view: cost per watt by city

Turner & Townsend takes a different approach. Its Data Centre Construction Cost Index for 2025 to 2026, published on 5 November 2025, compares 52 markets and states costs in US dollars per watt. A dollar per watt is the same as a million dollars per megawatt, so 15.2 dollars per watt means 15.2 million dollars per megawatt.

Its published ranking for traditional data centers has Tokyo at 15.2, Singapore at 14.5, Zurich at 14.2, Osaka at 14.1, Silicon Valley at 13.3 and New Jersey at 12.9. London is listed at 12.0 and Frankfurt at 11.6. The firm says overall cost inflation averaged 5.5 percent in 2025 for traditional, air cooled facilities.

Two findings stand out for anyone thinking about AI specifically. The firm measured, for the first time, a construction cost premium of 7 to 10 percent for AI data centers over traditional ones in the United States, comparing projects of similar IT capacity. The reasons it gives include the move to high density, liquid cooled designs and greater technical complexity. It also reports survey results: 83 percent of respondents said local supply chains are not well prepared for the advanced cooling that high density facilities need, and 48 percent said power availability is the biggest obstacle to delivering projects on schedule.

A 7 to 10 percent premium may look small next to the Cushman 21 percent rise. They answer different questions. The first compares AI buildings to ordinary ones today. The second compares today's costs to those of the previous edition, about two years earlier.

From megawatts to gigawatts: whole campus examples

Per megawatt averages are tidy, but campuses are priced as one project. Two public examples give a feel for whole project numbers. Both are about financing and expected cost, not audited final spend.

The Hyperion campus in Richland Parish, Louisiana is being developed by Meta with funds managed by Blue Owl. The announcement language, as reproduced by Data Center Frontier, describes about 27 billion dollars in expected total development costs for the buildings and long lived power, cooling and connectivity infrastructure. Bisnow reports the campus as about 2,250 acres, nearly 4 million square feet across at least nine buildings, initially 2 gigawatts of AI computing and potentially above 5 gigawatts according to Meta's chief executive. Dividing 27 billion dollars by 2,000 megawatts gives about 13.5 million dollars per megawatt. That is simple division on the announced figures, and it assumes the 2 gigawatt starting figure. If the campus ends up larger, the per megawatt number falls, and if the 27 billion dollars changes, it moves too.

The Abilene, Texas campus built by Crusoe for Oracle and OpenAI is planned at about 1.2 gigawatts. Bisnow reported financing of about 4 billion dollars for the first phase and 11.6 billion dollars in debt and equity for the second, and cited about 15 billion dollars for the project overall. That is roughly 12.5 million dollars per megawatt if you divide 15 billion dollars by 1,200 megawatts. Financing raised is not the same as cost incurred, so this is a rough check only. In both examples, the buildings and their power and cooling land in the same broad range as the global and US averages above, below the 17.6 million dollar figure for the most modern builds.

What the computers add

The computers are the reason the headline numbers for a gigawatt get so large. According to Interesting Engineering, in coverage dated 22 September 2025 of the Nvidia and OpenAI announcement, Nvidia's chief executive told investors earlier that one gigawatt of data center capacity costs 50 to 60 billion dollars, of which about 35 billion dollars goes to Nvidia chips and systems. That is a statement by a chip supplier, relayed by a trade outlet, and it is a 2025 statement. The outlet does not link a transcript, so treat it as a relayed remark rather than a company disclosure. Higher per gigawatt figures circulate in secondary coverage without a clear origin, and they are not used here.

Put the layers side by side, using the figures above and the caveats stated for each. The building and power side runs roughly 11 to 18 billion dollars per gigawatt. The all in AI figure quoted by the supplier runs 50 to 60 billion dollars. The difference is the equipment. For a customer trying to understand what a lab's multi gigawatt plan implies, that is the key point: most of the money in a gigawatt of AI capacity is chips, and chips are bought on a different schedule and wear out faster than concrete. Our explainer on OpenAI compute commitments and Stargate shows how those multi gigawatt numbers are announced, and the one on Anthropic compute commitments covers the contract side.

Why costs keep rising

The sources agree on the broad causes, even where they disagree on the size.

• Power equipment and connection. Cushman ranks power infrastructure as the largest category. Turner & Townsend's respondents put power availability first among obstacles.
• Scarce powered land. The 51 percent rise in the price of powered land is the clearest single figure.
• Higher density cooling. Turner & Townsend links its AI premium to liquid cooled, high density designs, and finds local supply chains unprepared.
• Labor and materials. Cushman cites rising skilled wages and supply chain constraints.
• Competition for the same sites. CBRE reports 7,481 megawatts under construction in eight primary markets as of H1 2026, with more than 80 percent already preleased, up from 74.3 percent a year earlier, and primary market vacancy at a record low of 1.4 percent. When almost everything under construction is already spoken for, builders have little reason to cut price.

What is firm and what is uncertain

Firm, in the sense that a named source states it with a date: the Cushman 17.6 million dollar figure and its exclusions, the 21 percent rise, the Turner & Townsend per watt ranking and AI premium, and the Hyperion expected development cost.

Uncertain: what the JLL figure includes, how any average maps onto one particular site, and whether a given "gigawatt" is IT load or total power. The Nvidia per gigawatt figure is a supplier statement relayed second hand. Reported project financing is a proxy for cost, not a measure of it. Private company figures are unaudited. Finally, none of these numbers is a prediction of future price.

What to watch next

• The next editions of the Cushman and Turner & Townsend studies, to see whether the AI premium widens or narrows.
• Whether powered land prices keep rising faster than construction costs.
• Actual final cost disclosures from large campuses once they complete, which are rare and more informative than announcements.
• CBRE's next half year report for preleasing and vacancy.

Quick check

If you are reviewing the software and cloud components behind your own AI workloads, Safeguard offers SBOM and dependency scanning, plus a free scan at https://gold.safeguard.sh/scan.

Sources

• Cushman & Wakefield 2026 Data Center Development Cost Guide, Business Wire release via Finviz, 3 September 2026
• Cushman & Wakefield guide coverage, IREI, 20 August 2026
• Datacentres.com Research, Data Centre Construction Costs Rise to 11.3M per MW in 2026, 12 May 2026
• Datacentres.com, How much does it cost to build a data centre
• Turner & Townsend, AI data center rush raises alarm over construction supply chain, 5 November 2025
• Data Center Frontier, Meta Hyperion campus financing
• Bisnow, Meta pushes its largest data center project off its books, October 2025
• Bisnow, Primary Digital Infrastructure and the Stargate Abilene financing, May 2025
• Epoch AI, OpenAI Stargate: where the US sites stand, updated 23 April 2026
• CBRE, North American data center demand continues to outpace supply, 8 September 2026
• Interesting Engineering, Nvidia and OpenAI 100 billion dollar deal, 22 September 2025

Last reviewed: 19 October 2026.