The Trillion-Dollar Question: Can Big Tech’s Massive Bet on AI Data Centers Pay Off?

In the rapidly evolving landscape of artificial intelligence (AI), the construction and operation of data centers represent a significant financial commitment for large technology companies, or hyperscalers. A recent study by ScotiaBank Global Equity Research suggests that for a dedicated 100 MW inference campus based on GB200 NVL72 racks, the operational cost threshold is approximately $2.91 per GPU hour. This figure emerges from a detailed analysis of total expenses and potential revenue generated by large-scale inferencing operations. The study notes that constructing this data center costs around $4.03 billion for 100 MW of critical IT load, with 64% allocated to IT equipment (approximately $3 million per NVL72 rack) and 36% to facilities (about $14.5 million per MW). The majority of the financing is a debt at an interest rate of 8.0% over six years, with a residual value of 15% after six years. Despite these substantial costs, the study indicates that profitability can be achieved within the current market, considering rental indications that surpass this threshold. However, official numbers for token generation costs vary significantly between different models and platforms. The impressive size of these investments reflects the growing dependency on AI in the global economic fabric but also raises questions about long-term viability, especially in a volatile demand and technology landscape. Projections show that financial commitments by major tech players have already exceeded $2.588 trillion with another $779 billion pending on their balance sheets, not to mention additional ongoing guarantees. These numbers underscore the importance of strategic planning and financial risk management in this emerging sector.