Widely quoted estimates put the capital cost of 100 MW of data center capacity at $900 million to more than $1.8 billion depending on facility type, and they do not necessarily include servers and GPUs. Epoch AI’s May 2026 model implies the building is about 30% of the upfront capital for a GB200 NVL72 campus, while Cushman & Wakefield finds costs up 21% per MW in under two years. The ranges are plausible, but they describe one layer of the cost stack.
Epoch AI’s May 2026 model of a typical US hyperscaler AI campus puts upfront capital at $37.9 billion per gigawatt of IT capacity, of which the facility accounts for $11.4 billion, or 30.2%. Servers and networking make up most of the remainder, with land and utility works under 1%. Scaled to 100 MW, which Epoch notes is not fully linear, that is roughly $1.1 billion for the building and $3.8 billion for the whole campus. The commonly quoted ranges, which circulate without a named source, run $0.9 billion to $1.2 billion for hyperscale, $1.0 billion to $1.4 billion for colocation, $1.1 billion to $1.5 billion for enterprise, up to $1.5 billion for telecom, and $1.0 billion to $1.6 billion or more for AI and HPC, and they do not necessarily include servers, GPUs, and network equipment. That caveat carries more weight than the ranges: the gap between building and campus in Epoch’s model is about $2.7 billion per 100 MW, three times the entire spread across facility types.
Read per megawatt, the quoted hyperscale range of $9 million to $12 million lines up with construction-only benchmarks. JLL’s 2026 outlook, as cited by industry trackers, puts global shell-and-core cost at $11.3 million per MW, up 6% from $10.7 million in 2025 and 47% from $7.7 million in 2020, and Epoch’s facility cost works out to $11.4 million. Cushman & Wakefield’s September 2026 guide measures something broader: all-in greenfield development in the United States and Canada averages $17.6 million per MW for the most modern facilities, excluding chips and GPUs, or about $1.76 billion per 100 MW. That sits above every quoted range except edge, at $1.2 billion to $1.8 billion, which is a portfolio of sites rather than a single facility, and the open-ended top of the AI range. Turner & Townsend’s index, by contrast, covers construction only, from a baseline of a 30 to 50 MW air-cooled hyperscale build. The benchmarks are not in conflict, but a quoted figure is meaningless without knowing whether land, grid connection, and contingency are inside it. Cushman & Wakefield ranks power infrastructure as the largest category at 21% of all-in cost and contingencies second, while Turner & Townsend’s construction mix, as reproduced by industry trackers, puts electrical at 54% for an air-cooled build, so category shares from different sources cannot be compared directly.
Facility type matters less than the quoted ranges imply, and AI density less than its reputation. Turner & Townsend says colocation or enterprise facilities may move its indicated cost per watt by up to 15% from the hyperscale baseline. At the quoted bounds, the midpoints imply colocation about 14% above hyperscale, in line with that ceiling, but enterprise, AI, and HPC about 24% above, well outside it. For liquid-cooled AI halls, Turner & Townsend measured an average construction premium of 7% to 10% in the United States against air-cooled builds of similar IT capacity, with mechanical’s share of the cost mix rising from about 22% to 33% and electrical’s falling from 54% to 48%, according to figures reproduced from the index. The usual assumption that redundancy requirements push AI costs up also needs qualifying: Turner & Townsend notes that AI facilities may be able to minimise some of the redundancy cost that cloud data centers carry, depending on their reliability requirements. Applying the 7% to 10% premium to Frankfurt’s $11.6 per watt, a 100 MW build would add roughly $81 million to $116 million, a fraction of what the quoted AI range implies.
Location moves the number by more than cooling does. Turner & Townsend’s 2025 index ranks Tokyo at $15.2 per watt, Singapore at $14.5, Zurich at $14.2, Silicon Valley at $13.3, New Jersey at $12.9, London at $12.0, and Frankfurt at $11.6. The $3.6 per watt difference between Tokyo and Frankfurt is about $360 million on a 100 MW build, more than three times the liquid-cooling premium on a Frankfurt-cost project, though the index is based on 30 to 50 MW builds and is scaled here only for illustration. Timing matters as well. Turner & Townsend measured 5.5% cost inflation for traditional data centers in 2025, down from 9.0% in 2024, and JLL forecasts 6% in 2026, which compounds to roughly 12% on construction-only measures over two years. Cushman & Wakefield’s all-in figure rose 21% per MW between its late 2024 and 2026 guides, and powered land in primary US markets averaged $584,000 per MW, 51% higher year on year, which suggests that power equipment, land, labor, and contingencies are inflating faster than the building shell.
The denominator is the other unstated variable. Epoch defines its gigawatt as IT nameplate capacity and takes a power usage effectiveness of 1.14 from Lawrence Berkeley National Laboratory estimates for AI-specialised facilities, so 100 MW of IT load implies about 114 MW of facility demand. Project announcements and grid applications often use facility or grid megawatts instead, and a figure per MW of one cannot be applied to the other without that conversion.
Whole-campus estimates are scattered, which is why the facility benchmarks cannot stand in for them. Epoch’s $38 billion per gigawatt compares with Bernstein’s 2025 estimate of about $35 billion and the $50 billion to $60 billion Nvidia cited on its fiscal second quarter 2026 earnings call, a spread of $15 billion to $25 billion per gigawatt, or $1.5 billion to $2.5 billion per 100 MW, larger than the entire quoted building-cost range. Definitions of the gigawatt are not harmonised across these sources. The running costs invert the intuition behind the building figures: in Epoch’s model, servers are 60% of an $8.5 billion annualised cost, with energy at about $0.6 billion a year, roughly 7%.
Add EnergyNews.biz as a preferred source on Google
