The construction of new AI data centers is currently lagging behind the rising demand for inference capacity. Delays in power access, capital availability, and legislative permitting are forcing chipmakers to seek space in existing telecommunications and colocation facilities.

Anton McGonnell, vice president of core product at SambaNova Systems Inc., notes that many older sites lack the infrastructure to support modern liquid-cooled GPU racks. These systems often require reinforced flooring and power densities ranging from 120 to 250 kilowatts per rack.

SambaNova is positioning its air-cooled SN50 fifth-generation chip as a solution for existing data centers. The hardware is designed to handle the decode phase of inference, avoiding the heavy power and cooling requirements associated with newer GPU-based racks.

The company, which reached a valuation of $11 billion in July, is collaborating with partners to manage the prefill phase of inference. At the SC26 event, SambaNova demonstrated a heterogeneous setup using Intel D-Series GPUs for prefill alongside its own SN50 racks.

Industry-wide, the procurement of power and cooling equipment faces lead times of 18 to 24 months. These supply chain constraints, combined with rising construction costs, continue to limit the speed at which new AI-ready data center capacity can be brought online.