Morgan Stanley Projects U.S. Data Center Costs to Reach $49 Billion per Gigawatt

Morgan Stanley has released a report estimating that the cost of constructing U.S. data centers will climb to approximately $49 billion per gigawatt of capacity by July 2026. This figure represents a significant increase of $20 billion over the past three years, primarily influenced by the growing demand for AI infrastructure.

According to analyst Brian Nowak, the rising costs are attributed to several factors, including increasing GPU and memory prices, extended delivery times, material shortages, and a scarcity of skilled labor. Additionally, off-rack costs, which encompass expenses outside of the computing racks, have surged from $11–13 million per megawatt to $16–19 million per megawatt.

The report also highlights a shift towards behind-the-meter power setups, which generate electricity on-site rather than relying on the public grid, as a contributing factor to the rising costs. Despite these challenges, major cloud providers such as Meta, Amazon, and Microsoft continue to invest heavily in AI infrastructure, with Morgan Stanley raising its capital expenditure forecasts for these companies by 9-10%.

Furthermore, the report indicates that U.S. data center power demand could reach between 97–116 GW by 2028, with a potential shortfall of up to 70 GW without mitigation measures. Even with mitigation, a net shortfall of 33–43 GW is expected.

Interestingly, the report mentions former Bitcoin mining sites as part of the strategies to address the electricity gap. While this is seen as a marginal solution, it underscores the existing infrastructure that could be leveraged for AI development.

For investors and builders, these rising construction costs mean that the returns on AI services must increase significantly to justify the investments. Suppliers of GPUs, memory, and construction services are likely to benefit from the escalating costs, while energy access remains a critical factor in determining which projects will proceed as planned.