🧠 Food for thought

Implications, context, and why it matters.

GPU utilization and Return on Investment (ROI) timelines will show if there is a bubble

• Hyperscalers plan about $300 billion of 2025 spend on data centers and GPUs, with over $1 trillion by 2027 1. Payback hinges on GPUs staying busy enough for 18 to 24 month amortization periods, the time over which hardware costs are spread and paid back 1.
• Profit per unit of compute depends on price per token (small chunks of text processed by AI models), variable costs, tokens served plus idle time 1.
• Teams shift focus from brute-force pre-training to post-training and test-time inference scaling, which adds compute when a model answers a query, such as longer context or ensembles 1.
• The pace holds if AI adoption meets plans 2. Shortfalls raise risk of stranded assets, or capacity that sits idle and cannot earn returns 2.

Colocation and infrastructure vendors can benefit from power and cooling retrofits

• Rising GPU rollouts push data centers to add liquid cooling and power upgrades, opening work for retrofit suppliers and colocation operators, third-party data centers hosting customers’ servers 3.
• Liquid cooling, which routes coolant to hot parts, is set to rise from about 10% in 2024 to over 20% in 2025, aided by modular gear that scales upgrades at lower cost without full rebuilds 4.
• Grid bottlenecks and four year waits for new lines shift site picks to places with available power, favoring vendors that boost capacity at current sites 3.
• Schneider Electric, an industrial energy and data center supplier, is positioned for this wave 5. Edged, a data center developer focused on sustainable designs, uses closed-loop waterless systems (recirculating coolant without consuming new water) that can save over 277 million gallons a year at one campus 6.