What it is and the problem it solves
An AI data center is a specialised facility for training and running AI/ML models. It solves the problem of executing massively parallel AI workloads—but only at the cost of extreme power density and memory scarcity.
How it works
AI data centers use AI accelerators and high-speed interconnects to handle parallel AI workloads. They draw ~60 kW per rack—six times the power load of general-purpose data centers, which use ~10 kW per rack.
What works
The hardware stack—accelerators, interconnects, and high-bandwidth memory—enables viable AI model training and inference at scale. Power and memory demand are not theoretical; they are measured, deployed, and constraining.
What does not
They do not scale without straining global memory supply. Their construction since 2024 has exhausted global computer memory supply. HBM production consumes three times the wafer capacity per bit of DDR5, worsening the bottleneck.
What it changes
They shift memory production: 70% of global computer memory output in the 2026 fiscal year is allocated to AI data centers. The Stargate project alone takes ~40% of global DRAM wafer output—900,000 wafers per month.
Is it worth your time
Yes—if you manage memory supply chains, power infrastructure, or semiconductor fabrication. No—if your work is unaffected by DRAM allocation, wafer capacity, or rack-level thermal design.
