Core Scientific CEO: AI Data Center Build Costs Climb Toward $12M Per MW

Construction costs for AI data centers have climbed sharply as labor inflation, equipment bottlenecks and changing technical requirements reshape the economics of large-scale development, Core Scientific (NASDAQ: CORZ) CEO Adam Sullivan told investors at TD Cowen’s 54th Annual Technology, Media & Telecom Conference.

Sullivan estimated that data center build costs have risen from roughly $8 million per megawatt several years ago to about $11.5 million to $12 million per megawatt today, depending on design. Labor is one of the main drivers: overall labor costs are up about 20% year over year, while electricians are up about 30%.

Labor once accounted for about 30% of total build costs, according to Sullivan. Today, that share is moving closer to 40% or higher.

The comments offer a detailed look at the cost pressures facing companies racing to deliver AI-ready capacity for hyperscalers and GPU cloud providers. Core Scientific, historically known as a bitcoin miner, is now expanding aggressively into high-density data center infrastructure.

Some relief may come from design changes tied to future GPU architectures. Sullivan pointed to Nvidia-related design shifts, the move toward 800-volt systems and efforts to remove long-lead components from data center builds. UPS systems could be one of the next components redesigned out of certain configurations, he suggested.

Those changes could eventually bring build costs closer to $10 million per megawatt, though supply-chain constraints remain significant. Medium-voltage switchgear is currently one of the hardest items to source, with lead times around 100 weeks. Generator sets, by comparison, are taking about 14 months.

Core Scientific has tried to get ahead of the bottlenecks by pre-buying medium-voltage equipment. Because some of that gear can be shifted between projects, the company has the flexibility to redirect capacity across sites such as Dalton, Pecos and Muskogee.

The company is preparing five new development sites as part of its AI data center expansion. Sullivan described Core Scientific’s recent capital raise, which put about $3 billion on the balance sheet, as an important step toward prefunding the equity needed for those projects and securing contractors, long-lead equipment and land.

Customer requirements are also moving quickly. The first data halls across Core Scientific’s five new sites are being designed around Nvidia GB300 systems, while future buildings are expected to move toward 800-volt architectures. Sullivan said that shift could become a requirement for many 2028 deliveries.

Meanwhile, power strategy is changing alongside the technical requirements. Hyperscalers have become more willing to rely on behind-the-meter generation, moving from what Sullivan characterized as a “very low double digit” cap to more than 50% of site power. That change has influenced Core Scientific’s land acquisition strategy and its planning around natural gas pipeline access.

At Pecos, Texas, the company had to build a concrete plant on site because the location was too far from the existing supply. In Muskogee, Oklahoma, labor competition is intense: Core Scientific is building a 70-megawatt data center for CoreWeave (NASDAQ: CRWV) while Google is developing a 600-megawatt data center next door.

For customers, speed to power has become more important than the specific source of electricity, Sullivan indicated. Core Scientific can deploy hundreds of megawatts of behind-the-meter generation within 18 months, though distribution-level pipeline access remains a key constraint.

The rising cost environment is also affecting deal economics. Sullivan estimated that hyperscale data center deals are likely to settle around 12% to 14% development yields, with 10% unlevered yield on cost representing the upper end in some cases. Higher returns may be available from “neo cloud” customers, but those deals come with more counterparty risk.

Core Scientific remains focused on full turnkey, GPU-ready facilities, though it has examined hybrid structures that combine powered shell and turnkey delivery. Such structures may reduce development risk by shifting some final fit-out work to customers.

The company is also watching for potential consolidation. Sullivan expects execution problems over the next year to pressure valuations for some public and private data center developers, potentially creating acquisition opportunities for stronger operators.