SemiAnalysis sees about 50 GW of new data centres in 2027
Research firm SemiAnalysis forecasts roughly 50 GW of new global data centre capacity in 2027, and says the bottleneck may shift from power to chips and memory.
What happened
On 19 October SemiAnalysis published "Datacenters 2027: The Year of Acceleration", forecasting roughly 50 GW of new global data centre capacity in 2027, reportedly about equal to all capacity built before 2024. The report body also cites 43 GW, and available summaries do not reconcile the two.
The numbers
More than 24 GW across 40 campuses is in advanced construction for 2027, mostly over halfway complete. Binding, non-cancellable leases signed this year total slightly under $1 trillion, with tenants including Meta, AWS, Microsoft, Google, Oracle, Nvidia, AMD, Anthropic and OpenAI.
Where the bottleneck moves
The report suggests shortages could move from power to chips, with memory and logic chips the next bottleneck in 2027-28. Another SemiAnalysis piece expects HBM to stay scarce through 2027, and Dylan Patel argues ASML's EUV tools become the ultimate constraint by 2030.
The sceptics
Altimeter Capital's Brad Gerstner called the 43 GW forecast too aggressive and expects about 25 GW to be built. The full report is paywalled, so its detailed modelling is not public.
Why it matters
This wave will set the cost and availability of AI services. More capacity may keep model prices falling, but chip and memory scarcity could raise them and power demand could lift electricity prices.
Is power solved?
Power has been the main constraint for two years. Companies increasingly build their own on-site generation, such as gas turbines and fuel cells, instead of waiting for grid connections, which helps capacity grow fast. That is not possible everywhere: Nscale's Loughton site in the UK shows grid limits still decide many projects. Microsoft's withdrawal there came the same weekend as the report, a reminder that forecasts depend on local conditions.
What it means for your business
- Avoid long fixed price commitments or add price-drop clauses.
- Time hardware purchases while GPUs and memory are scarce.
- Watch regional energy prices.
