The ambitious expansion of artificial intelligence infrastructure is encountering a critical obstacle: a pronounced scarcity of gas turbines. Major global manufacturers are experiencing unprecedented demand, with production schedules extending well into the next decade. This supply chain crunch is impacting the ability of data centers to secure the necessary power generation equipment, leading to escalating costs and raising concerns about the stability of electricity grids. The situation highlights a growing disconnect between the rapid pace of technological advancement in AI and the more deliberate, capacity-limited nature of heavy industrial manufacturing.
The demand for power, particularly from U.S. data centers, is projected to surge dramatically. A Goldman Sachs report indicates that power consumption could more than double by 2027, climbing from 31 gigawatts in 2025 to 66 gigawatts. This translates to an increase in capacity additions from 8.5 GW last year to an anticipated 36.3 GW in 2027. Consequently, data centers would account for a substantial 8.5% of the total U.S. peak summer electricity demand, a significant jump from today's 4.1%. This aggressive growth trajectory places immense pressure on an already constrained power generation equipment market.
Leading gas turbine manufacturers are facing immense backlogs. GE Vernova, for instance, reported 116 GW of gas power equipment backlog and slot reservation agreements by the end of Q2, 2026, with deliveries stretching to 2031. Siemens Energy concluded its fiscal Q3 with a 69 GW backlog, while Mitsubishi Heavy Industries noted a 35 GW backlog for large-frame turbines, with orders scheduled for 2028-2030 delivery. Although the reporting methods for these figures vary, the consistent theme across all major players is a booked-out schedule for years to come. This means that even if a data center project were initiated today, the essential power generation components would not be available for several years.
The global manufacturing capacity for gas turbines, estimated by Wood Mackenzie at 60 to 70 GW annually, falls significantly short of the current order influx, which reached a record 38 GW in Q2 alone, with half coming from the U.S. This imbalance is exacerbating the supply challenges. Furthermore, the specialized nature of turbine components, such as hot-section castings, and the shortage of skilled labor for assembly and installation further complicate efforts to ramp up production. The average lead time for a new combined-cycle power plant has consequently expanded from three and a half years to roughly five, and even seven years for certain heavy-duty models.
The financial implications are also substantial. PJM's capacity auction on July 14, 2026, cleared at the maximum allowable price of $325 per megawatt-day, yet still fell short of reliability requirements by nearly 7 GW. This indicates a willingness to pay premium prices for power, underscoring the severity of the supply crunch. Turbine prices themselves have soared, with BloombergNEF reporting average combined-cycle project costs rising to $2,157 per kilowatt last year, up from under $1,500 in 2023. Wood Mackenzie anticipates turbine prices to increase by 195% from 2019 levels by the end of 2027. This escalation in costs is not driven by fuel prices, but by the bottleneck in manufacturing and skilled labor.
In response to these challenges, some states and utilities are beginning to reassess data center growth projections. Exelon, for example, reduced its "high probability" data center load from 18 GW to 11 GW, while Texas initiated an audit of its data center interconnection queue after requests reached an astonishing 474 GW. New York also paused new approvals. These actions suggest a recognition that the unchecked growth of AI infrastructure could strain existing power grids beyond their limits. The resolution of this bottleneck will likely involve policy decisions and regulatory interventions rather than a rapid increase in manufacturing output, potentially leading to a re-evaluation of data center development timelines.