How Much Does It Cost to Build a Data Center in 2026?

Data center construction cost per MW is the starting point for almost every number in the AI infrastructure economy, from the rent a colocation tenant pays to the debt a developer can raise. This guide, part of the Kurums Data Center Economy hub, breaks down what it costs to build a data center in 2026: the benchmark indices from JLL, Turner & Townsend and Cushman & Wakefield, how traditional and AI facilities differ, where the money goes inside the building, and why the GPUs often cost more than everything else combined. Once the facility is built, the bills keep coming, so read it alongside our breakdown of data center operating costs, power and PUE and our explainer on how data centers are financed.

How much does it cost to build a data center per MW in 2026?

A standard data center shell with power and cooling infrastructure costs roughly $10–14 million per megawatt of IT capacity in most major markets in 2026, excluding land and servers. Premium markets such as Tokyo exceed $18 million per MW, while some African and Indian markets come in below $11 million.

The industry quotes construction costs per megawatt of critical IT load, or per watt, which is the same number expressed differently: $12 per watt equals $12 million per MW. JLL’s 2026 Global Data Center Outlook forecasts an average shell-and-core cost of $11.3 million per MW in 2026, up 6% from $10.7 million in 2025 and from $7.7 million in 2020, a compound rise of roughly 7% a year. That figure excludes land and IT equipment.

Turner & Townsend’s Data Centre Construction Cost Index 2026-27, published on 22 October 2026, benchmarks 62 markets using data from more than 500 live or recent projects and a survey of 401 industry leaders. It puts the US country average at $12.08 per watt and says average cost per watt rose 13% year on year in North American and European markets, roughly double the 6% global rate. Cushman & Wakefield’s 2026 Data Center Development Cost Guide, released in August 2026, takes a broader, all-in greenfield view and reports an average of $17.6 million per MW for the most modern US facilities, with a range of $8.9 million to $23.3 million across US and Canadian markets, excluding chips and GPUs.

The gap between $11 million and $18 million is mostly about scope. JLL’s figure is the building and its core systems; Cushman & Wakefield’s includes site acquisition, sitework and large contingencies. When a developer, a lender and a hyperscaler compare numbers, the first question is always what is inside the figure.

Which cities are the most and least expensive places to build?

Japan is the most expensive place to build a data center for the fourth year running, with Tokyo at $18.28 per watt and Osaka at $17.15, according to Turner & Townsend. Zurich and the Nordic capitals follow, while Nairobi, Cape Town and Lagos sit near the bottom of the index.

Labour costs, seismic requirements, land scarcity and currency movements shape the ranking. Chicago climbed to 11th place at $13.28 per watt, a sign of how hard US Midwest labour markets have been hit by the AI build-out. At the other end, Lagos fell from 7th at its 2024 debut to 42nd, at $10.87 per watt.

Source: Turner & Townsend, Data Centre Construction Cost Index 2026-27 (October 2026). Cost per watt of IT load equals US$ million per MW.

In the United States, Cushman & Wakefield names Silicon Valley, Chicago, New York, Northern New Jersey, Toronto and Los Angeles among the priciest development markets, and Texas markets among the cheapest thanks to affordable land, labour and deep contractor networks. Silicon Valley paid the highest wages for engineering and skilled electrical and mechanical trades, while Chicago led on general construction wages. JLL’s market estimates point the same way: roughly $12–14 million per MW in Chicago, $11–12 million in Northern Virginia, and about $10–11 million in Phoenix, Dallas and Atlanta.

In Asia-Pacific, Cushman & Wakefield’s 2026 guide ranks Japan, Singapore, South Korea, Australia and Hong Kong as the dearest markets, with costs up about 10% in 2025. Cheaper markets help explain hyperscaler expansion elsewhere in the region, as covered in our guides to India data center investment and Asia-Pacific data center investment.

How much more does an AI data center cost than a traditional one?

The AI premium on the building is modest, about 10% in the US according to Turner & Townsend, putting AI facilities at $14–16 per watt. The real difference is the fit-out: JLL says AI tenant technology can cost up to $25 million per MW, often more than the building.

An AI hall differs in three expensive ways. Rack density: JLL notes AI training runs at roughly 40 kW per rack and above, so each megawatt needs heavier electrical distribution. Cooling: liquid loops supplement air handling, and JLL estimates liquid cooling adds about 10% to build costs. Schedule: AI tenants pay for speed, which means premium labour and parallel procurement.

Electrical architecture is becoming a cost variable in its own right. First Call Group’s 2026 estimates range from about $13–14 million per MW for a traditional 480-volt AC design to $18–20 million or more for an 800-volt DC design, the architecture Nvidia plans for next-generation AI factories.

Where does the money go inside a data center build?

Power infrastructure is the single largest cost. Cushman & Wakefield’s 2026 guide attributes 21% of a US greenfield budget to power infrastructure, 17% to core-and-shell construction and sitework, 16% to contingencies and only 7% to site acquisition.

The rest covers mechanical systems, fire protection, controls, fees and owner costs. The 16% contingency stands out: developers carry large buffers because equipment prices and delivery dates remain volatile. Powered land is getting dearer, though. Cushman & Wakefield reports powered land in primary US markets at roughly $584,000 per MW, up about 51% year on year.

Electrical: switchgear, transformers, UPS and generators

Everything between the utility substation and the server rack is electrical: medium-voltage switchgear, transformers, uninterruptible power supplies, batteries, power distribution units, busways and standby generators. For a GPU campus, Bernstein’s analysis estimates that mechanical and electrical systems make up close to one-third of total capex including IT, with generators and turbines around 6%, transformers around 5% and UPS systems around 4%.

Mechanical and cooling

Chillers, cooling towers or dry coolers, computer-room air handlers and, increasingly, coolant distribution units and liquid loops. Bernstein puts thermal management at about 4% of an AI campus’s total cost, a small share of the whole but a large share of the building. Our explainer on air, direct-to-chip and immersion cooling covers the technology choices in detail.

Shell, site and soft costs

The envelope, structure and sitework are the least differentiated items. Soft costs such as design, permitting and construction financing rise when projects slip, and JLL reports that 57% of projects saw schedule delays in 2025.

How much do the servers and GPUs add to the bill?

For an AI data center, computing hardware usually costs more than the building. Bernstein estimates about $35 billion per gigawatt all-in, Epoch AI models roughly $37–38 billion for a 1 GW hyperscaler campus, and Nvidia has cited $50–60 billion per gigawatt.

These figures move the conversation from $11 million per MW to $35–60 million per MW. Bernstein’s November 2025 analysis of an Nvidia GB200 NVL72 deployment put a single rack at about $5.9 million, of which $3.4 million was compute hardware and $2.5 million physical infrastructure; GPUs alone accounted for about 39% of total capex and networking about 13%. Epoch AI’s May 2026 model of a stylised 1 GW US hyperscaler facility, built entirely with GB200 NVL72 systems, estimates up-front capex of roughly $37–38 billion, with the chip stack accounting for more than half and land and grid connection under 1%.

Nvidia’s own framing is higher. On its fiscal second-quarter 2026 earnings call, the company cited $50–60 billion per gigawatt of AI data center capacity, and in an October 2026 podcast interview Jensen Huang said Nvidia’s own revenue per gigawatt had risen from about $18 billion to more than $40 billion as it sells networking, CPUs and other systems alongside GPUs. Estimates differ by chip generation and scope.

This matters for anyone reading capex headlines. When a company announces a multi-gigawatt AI campus worth tens of billions of dollars, most of that money buys hardware with a useful life of perhaps five years, not concrete that lasts decades. That is the core of the depreciation debate we cover in our analysis of the AI data center bubble question, and it explains why Big Tech capex has grown so much faster than construction spending alone.

Why are data center construction costs still rising?

Costs are rising because demand for the same scarce inputs, electrical equipment, copper and specialist trades, is outrunning supply. Cushman & Wakefield reports a 21% rise per MW since late 2024, and Turner & Townsend found North American and European costs up 13% in a year.

Cushman & Wakefield tracked equipment price increases from December 2021 to early 2026: switchgear up 60%, copper up 46%, generators and motors up 33%, cooling equipment up 32% and transformers up 29%. Lead times compound the problem: the guide cites 68 to 113 weeks for pad-mounted transformers, 60 to 100 weeks for generators and 38 to 63 weeks for medium-voltage switchgear. A project that cannot get its transformers on time pays for idle crews, extended financing and, in the worst case, lost tenants.

Labour is the second pressure. In Turner & Townsend’s survey, 81% of respondents said skilled labour is constrained, with electrical and mechanical trades most sought after, and roughly one in five reported 2026 cost increases above 15%. JLL’s construction research, reported by Data Center Knowledge in October 2026, puts general escalation at 3–6% a year, against 2–3% before 2020.

Does building bigger lower the cost per MW?

Not reliably any more. Turner & Townsend’s 2026-27 report concludes that larger projects are no longer a dependable route to lower costs, because gigawatt campuses concentrate demand for transformers, labour and grid capacity in one place and often require on-site power generation.

Scale spreads design and substation costs, but a 1 GW campus also needs thousands of workers in one county and often its own generation.Uptime Institute counted about 181 GW of large data center projects announced in 2025, roughly double 2024, with about 80% in North America, and has cautioned that many may never be built. Our construction pipeline statistics track how much of that capacity is actually under way.

What still cuts costs is repetition: standardised designs, prefabricated power and cooling skids and framework agreements that let developers pre-order equipment. Turner & Townsend advises securing critical resources early and building closer supplier partnerships.

How do construction costs affect rents and returns?

Construction cost sets the floor for rent. Developers price leases to earn a target yield on total development cost, so a 21% rise in cost per MW must be passed through in higher rents, accepted as lower returns, or offset by cheaper financing.

So far the market has absorbed the increase. JLL reports global occupancy approaching 97% and forecasts rents rising about 5% a year through 2030, with AI-optimised space commanding lease rates up to 60% higher than traditional space. Digital Realty, for example, told investors in its second-quarter 2026 materials that it had about 1.4 GW under construction at a total cost of roughly $20 billion, about 63% pre-leased, with an expected stabilised yield of around 11.5%. Developers with pre-leased capacity can lock in returns before the concrete is poured; speculative builders carry the cost risk themselves. We explain these contract structures in our guide to data center business models.

JLL’s Matthew Giddings warned in October 2026 that higher asset values cannot close the gap on their own if power delays stretch timelines. Time to power, pre-leasing and tenant credit matter as much as cost.

How should a company budget for a data center project in 2026?

Start from the IT load and rack density, not the floor area, then build up three layers: powered land and grid connection, shell and core with mechanical and electrical systems, and tenant fit-out plus IT. Add a contingency in the mid-teens percent, in line with current market practice.

In practice: confirm utility capacity and the connection date first; fix density and cooling design early; pre-order transformers, switchgear and generators as soon as the design allows; benchmark against two indices with the same scope; and price the cost of delay at your financing rate. Then decide whether to build, lease wholesale or buy colocation.

Frequently Asked Questions