3-Year GPUs

Research Affiliates CEO Chris Brightman believes the AI arms race behaves differently from traditional capital investment.

• AI capex surged from $250 billion in 2024 to $650 billion in 2026, based on Bloomberg estimates.
• On the surface it looks like data center expansion, but in reality it's closer to replenishment spending that rapidly swaps out GPUs and servers.
• According to Brightman, the economic lifespan of AI hardware is only about 3 years.
• As an example, the Nvidia H100 generated $36,000 in annual profit and 137% ROI in year 2, but by year 4 this sharply deteriorated to a loss of more than $4,400 and -34% ROI.

The core argument is not physical failure but the pace of generational turnover. As Nvidia, AMD, and others release better performance-per-watt and performance every year, hyperscalers must keep buying new equipment just to maintain the same AI capacity.

Because of this, Meta, Amazon, Microsoft, and Alphabet are structured less around making money from AI and more around absorbing losses to protect their existing core businesses. Amazon must defend cloud, Microsoft must defend Office/subscriptions, Alphabet must defend search advertising, and Meta must defend social advertising — the logic being that falling behind on AI features would undermine their competitiveness.

On the other hand, AWS CEO Andy Jassy has stated that he views the useful lifespan of server and networking equipment as 5 to 6 years, revealing differing views within the industry over depreciation and actual hardware lifespan.