GPUs and RAM Are in Short Supply, but the Real Bottleneck for AI Is Electricians
8 hours ago
- #Energy Demands
- #Datacenter Expansion
- #AI Infrastructure
- AI and HPC demands are driving massive datacenter expansions, with companies like TeraWulf pivoting from Bitcoin mining to AI infrastructure.
- TeraWulf's Lake Mariner site is scaling up to 750 MW by 2025, featuring buildings like CB-4 (200 MW) with liquid cooling and heavy-duty floors supporting up to 10,000 lbs per rack.
- The site uses a closed-loop cooling system without water in normal operations, with fluid lasting 10–15 years, and relies on dual 345 kV power feeds for redundancy (89% zero-carbon grid mix).
- Key partners include Schneider Electric for electrical infrastructure and Motivair for liquid cooling, with clients like Core42 and Fluidstack already operating on-site.
- Infrastructure costs for AI/HPC are high, around $7M–$10M per megawatt, partly due to liquid cooling and mechanical rooms that occupy over half the footprint.
- A major bottleneck for datacenter projects is the shortage of electricians, with 650–800 needed at Lake Mariner, sourced partly from local projects like the Buffalo Bills stadium refurbishment.
- Connectivity requirements have eased; low latency is less critical for training sites, and building connectivity is simpler than before.
- TeraWulf operates as a colocation provider, with customers installing their own compute racks and paying average U.S. rates of ~$140/kW/month on 10–15-year contracts.
- The company has invested $290M in infrastructure and holds 3 GW of capacity across sites, leveraging existing power assets like a former aluminum plant in Kentucky.