The extraordinary sums committed to AI infrastructure suggest an industry unconstrained by capital, but money is proving easier to find than many of the things needed to turn it into compute. Across North America, access to power, skilled labour and viable sites is beginning to redraw the data centre map while pushing the cost of every new megawatt sharply higher.
There is a revealing contradiction in the economics of the AI infrastructure boom. Investors are prepared to commit hundreds of billions of dollars to new capacity, yet a project with financing in place can still be held back by an electrical connection, a shortage of specialist contractors or equipment that will not arrive for more than a year. The constraint on growth is increasingly less about willingness to invest than whether the physical economy can keep pace with the extraordinary appetite for compute.
The scale of that collision is captured in Cushman & Wakefield’s North America Data Center Development Cost Guide 2026, which examines development economics across the United States and Canada. The report estimates a $2.3 trillion pipeline of data centre investment, but only $492 billion of that is currently under construction. The remaining $1.8 trillion sits in planned and precommitted development, leaving an enormous gap between the infrastructure the market wants to build and the capacity being delivered.
AI is central to that expansion, but it is also making the underlying constraints harder to solve. Facilities designed for AI and high-performance computing demand more sophisticated power and cooling infrastructure, while the rush to build capacity has intensified competition for the equipment and people required to deliver it. Construction costs per megawatt have risen by an average of 21 per cent since the previous edition of the report in late 2024, far ahead of inflation over the same period.
John McWilliams, Head of Data Center Insights at Cushman & Wakefield, describes an industry entering a different phase. “Access to power, land, and skilled labor are increasingly determining where growth can occur,” he says. “While development costs have risen significantly, demand for AI and cloud infrastructure continues to reshape investment patterns across North America. The result is an unprecedented pipeline of development that is extending beyond traditional hubs and into emerging markets with the scale and resources needed to support the next generation of digital infrastructure.”
Power changes the value of land
Land has always mattered to data centre development, but the value of the ground itself is becoming inseparable from what can be delivered to it. A cheap parcel with no realistic prospect of securing sufficient electricity may have little relevance to an AI operator, while an apparently expensive site with committed power can remove years of uncertainty. That distinction is creating a market in which developers are increasingly paying for certainty rather than acreage.
Cushman & Wakefield attempts to bring some definition to the increasingly loose industry term “powered land”. Its model begins with raw land and progressively adds secured grid capacity, planning entitlement, network connectivity and acceptable ground conditions. At the top sits what it calls “true powered land”, where the major development risks have already been addressed and power can potentially be delivered within 12 to 18 months. Such sites are scarce enough to attract competitive bidding.
That scarcity is now visible in prices. Powered land in primary US markets averaged $584,000 per megawatt during 2026 to date, 51 per cent higher than a year earlier. Yet the more significant consequence is not the price itself but the behaviour it encourages. Hyperscalers can pay a premium to remove power risk, while developers and investors are acquiring cheaper raw land in the hope that securing electricity and permissions will transform its value. The report notes that this speculative activity has contributed to so-called “phantom” data centres, projects occupying positions in power queues without necessarily having an end user ready to build.
The result is beginning to change where infrastructure goes. Established markets still possess network connectivity, expertise and mature supply chains that are difficult to replicate, but those advantages cannot manufacture electricity or suitable land. Development is consequently radiating beyond traditional clusters as operators search for places where very large campuses remain physically possible.
Building the facility is becoming harder
Finding somewhere to put the data centre solves only part of the problem. Once construction begins, developers enter another market in which AI infrastructure is competing for scarce resources. Mechanical and electrical trades are particularly important to data centres, and the report describes an ageing workforce and insufficient new entrants colliding with exceptional levels of digital infrastructure construction.
The implications extend beyond higher wages. A market can offer abundant land and power but still struggle to deliver projects if it lacks enough electricians, mechanical specialists and technical workers. Importing labour adds cost and logistical complexity, while multiple hyperscale developments arriving simultaneously can overwhelm even apparently healthy regional workforces. Cushman & Wakefield highlights Houston, Denver, Dallas-Fort Worth and Austin-San Antonio as relatively strong across several relevant labour pools, illustrating why site selection can no longer be reduced to electricity and real estate alone.
The shortage of skilled workers is also changing how data centres are built. The report identifies growing use of off-site modular and manufactured mechanical, electrical and plumbing systems, allowing substantial parts of a facility to be prefabricated elsewhere before being transported and installed on site. This does not eliminate the need for specialist labour, but it reduces the number of workers required on site and can make development more practical in markets where the local skills base would otherwise struggle to support a major project.
Even when land, power and labour are secured, the supply chain can still determine when a data centre actually opens. Transformers, generators and switchgear are not simply becoming more expensive; some now carry procurement times long enough to reshape project schedules. The guide records estimated lead times reaching 113 weeks for pad-mounted transformers and 100 weeks for generators, while switchgear prices have risen 60 per cent since 2021. General construction materials increased much more modestly over the same period, suggesting that the infrastructure boom itself is creating acute pressure around the specialist components on which data centres depend.
Those pressures explain why the headline construction cost is considerably more complicated than inflation. For modern greenfield facilities, the report puts average all-in development costs at $17.6 million per megawatt, excluding the chips and GPUs that ultimately make an AI facility valuable. Depending on resilience and workload requirements, the range extends from $8.9 million to $23.3 million per megawatt. Power infrastructure alone accounts for around a fifth of greenfield development cost, while contingency has become the second-largest category because uncertainty itself now needs a substantial budget.
The data centre map moves outward
These constraints are producing a geographic shift that may prove more consequential than the rising cost of individual projects. Primary markets are not disappearing, nor are their advantages becoming less important. The problem is that the scale of anticipated AI demand can no longer be accommodated within them alone.
Cushman & Wakefield expects primary markets to retain enormous investment, but their share of planned and precommitted capital expenditure falls as development spreads outward. Frontier markets currently account for 18 per cent of capital spending on projects under construction but are projected to capture 35 per cent of planned and precommitted investment. Their anticipated capital expenditure reaches $631 billion, almost seven times the amount currently being deployed there.
Texas demonstrates why that migration is not simply a search for cheap land. Several of its markets sit towards the lower end of the report’s construction-cost rankings because relatively affordable development is combined with lower labour costs and an established ecosystem of contractors and operators. The emerging winners of the AI infrastructure race may therefore be locations capable of assembling the complete development proposition rather than those offering the cheapest single ingredient.
There is another constraint that cannot be solved through engineering alone. Community opposition and tighter regulation are becoming material development risks as data centres consume more land and power and move into places with little history of hosting them. The report argues that community engagement and entitlement strategy now affect site viability and schedule certainty, meaning social acceptance increasingly belongs in the same development calculation as grid capacity and construction cost.
That changes the meaning of the $2.3 trillion pipeline. It is tempting to read the figure as evidence that an immense wave of infrastructure is inevitable, but planned capital is not the same thing as delivered capacity. Every proposed gigawatt still has to find electricity, navigate approvals, secure equipment and be physically built by a workforce already under pressure.
The AI infrastructure race is therefore entering a less glamorous but more decisive stage. The availability of GPUs and the ambitions of model developers may determine how much compute the market wants, but they cannot determine how quickly North America can build it. Increasingly, that answer will be found in substations, construction sites and the communities being asked to accommodate an infrastructure expansion on a scale the digital economy has never attempted before.
