A Data Centre Boom Built for Yesterday’s AI

The data centre boom may be structurally mispriced because cheaper small models and local AI could reduce the need for centralized, exponentially growing compute.

• US data centre owners would need total industry revenue equal to 9–14% of projected US GDP by 2032 to earn a 10% return, according to a report cited from Columbia Business School professor Stijn Van Nieuwerburgh.
• A Stanford study reported by the Financial Times found that smaller models can match or outperform larger models on many common chat queries, while continuing to improve on reasoning tasks.
• The IEA estimates that a small language model can use as little as 3% of the energy required by a large reasoning model; KPMG also found that 49% of business leaders had scaled back, postponed or paused AI initiatives because of rising expected costs.
• The Silicon Data LLM Token Expenditure Index has fallen to around half its May peak, largely because usage has shifted toward cheaper Chinese open-weight models.
• Local AI could bypass data centres by running on laptops, phones and other devices, offering lower marginal costs and stronger data privacy; Nvidia is already developing AI-capable computers with Dell and Lenovo.

Takeaway: Investors should question whether debt-financed data centre capacity planned today will match demand when facilities become operational in three to five years, especially as power constraints, overcrowding and uncertain AI-lab economics compound the risk.