Via City AM,

• Big Tech's AI spending has exploded, with Google, Microsoft, Amazon, and Meta collectively investing hundreds of billions of dollars in infrastructure.

• Rapid technological change may shorten the economic life of AI servers and GPUs, increasing depreciation and replacement costs.

• The long-term profitability of AI will depend not only on demand growth but also on whether companies can justify the enormous ongoing capital requirements.

The eye-watering capital expenditure plans of Big Tech has been one of the year’s biggest stories. 

Google, Meta, Amazon, and Microsoft have all splurged to secure a podium spot in the race to build out the infrastructure that will run the artificial intelligence (AI) revolution.

Total capex by these four firms is expected to reach $750bn (£560bn) this year, around half the annual spending of the entire UK government. It is much higher than this high-tech quartet has budgeted for before. And it is expected to be even higher next year.

Shareholders are on board with the plan, up to a point.

Since 2023, the average share price across the four firms has doubled. But that hasn’t kept pace with the average quarterly capex budgets, which have roughly quadrupled over the same period.

These trillion-dollar businesses can’t be too far away from hitting a ceiling on growing their computing power. 

Firstly, because of physical constraints – things like the supply of chips and the availability of power and water infrastructure – with the latter beginning to come under genuine constraint in some parts of the developed world.

Secondly, because of the sheer build cost, given that most AI projects are far from hitting profitability, and there isn’t enough cash flow elsewhere to fill the hole.

Alphabet, Google’s parent company, has raised $85bn on its own in debt over the past year. It plans to raise another $80bn in equity over the coming months – an unprecedented fundraise and not something it can keep doing forever.

Getting older faster

Most of the focus has been on data centre build-out. But there is also another major factor, and one in danger of being overlooked: maintenance.

The cost of keeping AI running once the infrastructure is in place will be vital. 

Data centre servers tend to last in the region of three to six years before they have to be replaced.  Given the speed of innovation and intensity of compute needed for AI, you can expect that to skew towards the lower end of the range for the hyperscalers. 

The kit inside AI data centres accounts for as much as two-thirds of the build cost. Add replacement costs onto the capex projections over the next few years, and things start to look scarily expensive.

Annual depreciation of property and equipment across the four firms has almost doubled over the past two years to $116bn.  You can expect that to accelerate, given how much equipment has been added to their balance sheets over the past 18 months.

Last year, Amazon cut the expected useful life of its data centre assets from six years to five, a move which it said was “due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning.” 

So far, Meta, Microsoft, and Alphabet have yet to follow suit, sticking with six years, but it seems like only a matter of time before they capitulate and cut this back, pushing up depreciation costs even further.

Something has got to give – sooner or later. Or am I missing something?