TL;DR
SemiAnalysis's "Datacenters 2027: The Year of Acceleration," released October 25, projects about 51.2 GW of new data center IT capacity outside mainland China in 2027, more than double 2026. The report's key turning point: as power and construction constraints ease, chips and HBM will return as the primary bottleneck for AI data centers in 2027-2028.
Background
Throughout 2026, debate over whether AI data center buildout is overheated spanned the industry. Hyperscalers and AI labs kept signing gigawatt-scale leases, while figures such as Altimeter Capital founder Brad Gerstner publicly questioned delivery capability. In this context, SemiAnalysis released "Datacenters 2027: The Year of Acceleration" on October 25, 2026, with a quantified forecast: about 51.2 GW of new data center IT capacity outside mainland China in 2027, versus about 25 GW in 2026 and 15.6 GW in 2025.
Three Pillars Behind the 51.2 GW Forecast
SemiAnalysis supports its forecast with three verifiable indicator classes:
- Satellite imagery: About 40 campuses worldwide, totaling over 24.3 GW, are in late-stage construction, with most projects more than half complete and expected to deliver within 2027. The report calls this the most direct evidence that planned capacity is actually being built.
- Lease contracts: Binding, non-cancellable data center leases signed globally in 2026 YTD total slightly below $1 trillion. Meta, AWS, Microsoft, Google, NVIDIA, AMD, Anthropic and OpenAI are all seeking gigawatt-scale leases; Oracle completed about 4.5 GW in the second half of 2025.
- Self-build starts: Quarterly megawatt starts by U.S. hyperscalers are growing in a parabolic pattern, indicating demand has not stalled.
The Construction Model Is Changing
The report details how AI data center delivery cycles have shortened significantly:
- Modular construction has cut the delivery cycle of a 50 MW AI data hall from 22 months to 14 months; DayOne has compressed cycles from over 18 months to 9 months, cutting on-site labor by up to 63%.
- Behind-the-meter power has grown from near zero in early 2024 to 75 GW of reservations by August 2026, with 22 equipment suppliers.
- Microsoft abandoned a 50 MW data center lease near London because the local grid is not expected to be available until the 2030s, showing grid access remains a regional hard constraint.
Geographic Expansion
Non-U.S. markets grow faster in 2027: non-U.S. capacity is projected to rise from 6.9 GW in 2026 to 13.2 GW in 2027, nearly doubling; global capacity excluding mainland China reaches about 38 GW in 2027. Northern Europe and the Middle East, with available power and land, absorb more hyperscale projects.
The Next Bottleneck: From Power to Chips
The report's most important turning-point argument: completing a data center does not mean filling it with GPUs on day one. From 2024 to 2026, power access and grid interconnection were the primary bottlenecks. As construction cycles shorten and behind-the-meter power scales up, SemiAnalysis expects chips and high-bandwidth memory (HBM) to become the primary bottleneck again in 2027-2028. Even if data halls are delivered on schedule, GPU and HBM supply will determine how fast actual compute capacity comes online.
Data Sources
- SemiAnalysis "Datacenters 2027: The Year of Acceleration" (published October 2026), as reported by Xinzhi Times via Sina Finance on October 25, 2026 link