AI companies could have substantially more data center capacity available in 2027, according to SemiAnalysis’s October 19 report. The research firm forecasts approximately 50 gigawatts of additional global capacity next year, based on construction already underway and long-term leasing commitments.
If much of this capacity enters service, AI companies could expand model training and inference. Their ability to do so would still depend on access to advanced processors and high-bandwidth memory, which could become more restrictive as new facilities await equipment.
- Construction progress supports the 2027 forecast
- Long-term leases help developers finance construction
- Prefabrication can reduce time spent onsite
- Power supply remains a potential source of delays
- Chip shortages could leave new capacity underused
- More compute could reduce prices while testing project economics
Construction progress supports the 2027 forecast
SemiAnalysis reports that more than 24 GW of data center capacity across 40 campuses has reached advanced stages of construction, with much of it expected to be delivered in 2027. The firm also identifies approximately 60 GW under construction worldwide. These figures cover different portions of the development pipeline and should not be added together.
Projects already under construction are more likely to finish than those still in the planning stage, though delivery depends on equipment availability and grid connections. Facilities must also complete final testing before they can begin operations.
The forecast describes electrical capacity in gigawatts, while terawatt-hours measure electricity consumption over time. Neither indicates how much AI computing power a facility can deliver, since actual performance depends on the hardware installed and its utilization.
Long-term leases help developers finance construction
Long-term leases help data center developers secure construction financing by giving lenders evidence of future rental income. With a tenant committed to the facility, lenders can better assess whether the project will generate enough revenue to cover debt payments.
An increase in leasing commitments can indicate that more projects are moving toward construction. However, lease values represent payments expected over the contract period, rather than money already spent on development or capacity available for use.
SemiAnalysis reports a substantial increase in leasing activity, although its aggregate dollar figures differ and the available text does not clarify what each includes. The findings indicate that leasing commitments have grown, but they do not establish a single confirmed total.
These agreements also carry long-term financial risks, as The Neuron's coverage of OpenAI's compute commitments explains. Developers rely on future rental income to repay construction loans, while tenants must generate enough revenue to meet their payment obligations throughout the lease.
Prefabrication can reduce time spent onsite
SemiAnalysis credits modular construction and prefabricated electrical equipment with shortening data center building schedules. By assembling components in factories while site preparation continues, developers can complete multiple stages of construction at the same time.
The report cites timelines of less than 10 months for some hyperscaler projects and approximately nine months for DayOne developments. These figures reflect specific projects and should not be treated as typical construction periods across the industry.
Using standardized designs across multiple campuses can also reduce the need for custom assembly. However, faster construction does not guarantee an earlier opening if critical equipment, such as transformers, arrives late or the facility has yet to secure an operational electricity connection.
Power supply remains a potential source of delays
SemiAnalysis argues that onsite power generation and alternative electricity arrangements have eased supply constraints for data center developers. However, the International Energy Agency’s April assessment reports continuing shortages of turbines and transformers, as well as delays in securing grid connections. Its satellite analysis also found that many onsite gas-generation projects were still in early development.
Although the two assessments were published six months apart, the IEA’s findings highlight potential obstacles to bringing new facilities online. Onsite power plants require reliable fuel supplies and specialized equipment, while grid-connected data centers depend on utilities completing the necessary infrastructure.
The scale of future demand could add pressure to electricity supplies. According to Lawrence Berkeley National Laboratory’s June 2026 update, U.S. data centers could consume 649 terawatt-hours of electricity by 2030 under its reference scenario, equivalent to approximately 11.8% of national electricity consumption. Other scenarios project between 521 and 843 TWh, representing 9.5% to 15.3% of the total.
These estimates depend on how quickly data centers expand and how much electricity their equipment consumes. Actual demand could fall anywhere within or outside the projected range as construction schedules and computing requirements change.
Chip shortages could leave new capacity underused
Long-term leases help data center developers obtain construction financing by giving lenders evidence of future rental income. With a tenant committed to the project, lenders can better assess whether the facility will generate enough revenue to cover its debt.
An increase in leasing commitments can indicate that more projects are moving toward construction. However, the value of these agreements reflects payments expected over the lease term, rather than money already spent or computing capacity available to customers.
SemiAnalysis reports a substantial rise in leasing activity, although its aggregate dollar figures differ and the report does not clearly explain what each includes. The figures suggest that developers are securing more tenant commitments, but they cannot be combined into a single confirmed total.
These agreements also carry financial risks, as The Neuron's earlier coverage of OpenAI's compute commitments explains. Developers depend on rental payments to service construction debt, while tenants must generate sufficient revenue to meet their obligations over the life of the lease.
More compute could reduce prices while testing project economics
If available computing capacity expands faster than demand, AI providers could face pressure to reduce inference prices. They could also use the additional resources to support longer reasoning tasks or run more AI agent sessions simultaneously.
Lower inference costs could encourage greater use of computationally intensive applications. AI agents often require multiple model calls to verify results or retry unsuccessful actions, which increases the compute required for each completed task. As individual calls become cheaper, more of these workloads could become financially viable.
For data center developers, rapid expansion introduces financial risks. Construction debt must still be repaid if customers postpone deployments or consume less capacity than anticipated. The IEA also warns that growing reliance on capital markets makes data center development more sensitive to borrowing costs and expected investment returns.
SemiAnalysis’s forecast suggests that accelerated construction could increase the infrastructure available for AI workloads in 2027. However, lower inference prices and more capable services will depend on how quickly new facilities receive power and computing equipment. Developers must also secure sufficient customer demand to meet their financial obligations.
