BIG TECH´S CAPEX SPLURGE MAY BE IRRATIONALLY EXUBERANT / THE ECONOMIST

A trillion-dollar AI arms race

Big tech’s capex splurge may be irrationally exuberant 

Beware of overhype and overbuild

From the 19th-century railway mania to the telecoms boom at the dawn of the internet age, cautionary tales abound of over-investment in infrastructure fuelled by excitement over a new technology. 

With the rise of generative artificial intelligence (AI), history is repeating itself. 

In recent weeks four tech giants—Alphabet, Amazon, Meta and Microsoft—have pledged to spend close to a total of $200bn this year, mostly on data centres, chips and other gear for building, training and deploying generative-AI models. 

That is 45% more than last year’s blowout. 

Tech barons such as Meta’s Mark Zuckerberg admit that it may be years before this investment generates returns. 

It is an AI arms race.

The tech firms are not only buying infrastructure. 

In the past few years they have joined a stampede to put venture capital (VC) into Openai, Anthropic and other makers of foundational models. 

Traditional VC firms bellyache that they have not seen such corporate big-footing since the dotcom boom. 

The tech giants are flush with cash—they can afford to splash out. 

But, if the past is any guide, a bust is coming and the firms carry such weight in the stockmarket that, should their overexcitement lead to overcapacity, the consequences would be huge.

History is illustrative. 

In the early days railway track was laid for locomotives that were soon superseded by more powerful ones. 

As the rolling stock grew heavier, lines had to be replaced with sturdier stuff. 

During the 1990s telecoms firms increased capital expenditure by three-and-a-half times and laid 600m km of cable, according to Morgan Stanley, believing that people would be willing to pay high fees to go online. 

But users thought the internet should be free. 

The bank draws an analogy with today’s data centres, which may be ill-equipped to cope with increasingly powerful graphics-processing units, the chips used to train and run generative AI. 

The tech giants’ assumptions about people’s willingness to pay for chatbots and other whizzy “gen-ai” tools may be just as misplaced.

All the signs are that big tech has succumbed to irrational exuberance. 

Runaway spending is one of the risks. 

Wall Street is already pencilling in expectations that the four firms’ capex could come to an eye-popping $1trn over the next five years (Apple is taking a more cautious approach). 

Revenues may rise as a result, but so will costs. 

These include juicy salaries for brilliant engineers and mammoth electricity bills for data centres that can handle the heavy demands of generative AI. 

Even now, investors are lukewarm about such gambles. 

In recent weeks they have applauded Google’s capex plans but thrown cold water on Meta’s. 

Given the way Mr Zuckerberg and others have blown money on Pharaonic projects before, they have good reason to be jittery.

Another risk is that models will be commodified. 

The cloud-computing “hyperscalers”, namely Alphabet, Amazon and Microsoft, have built and invested in large proprietary AIS that are considered state-of-the-art. 

They also run smaller, open-source alternatives, including those made by Meta, which are getting better and cheaper. 

Hugging Face, a hub for ai enthusiasts, lists more than 650,000 models. 

The more market share these take from the large proprietary models, the lower the likely returns on investment.

A last risk is that scale brings diminishing returns. 

With more computing power and data, the giants’ models are getting bigger. 

But nobody should assume that they will get proportionally better as more money is thrown at them. 

On May 13th Openai launched a new version of its GPT-4 model, called GPT-4o. 

It is faster and its linguistic skills mean that you get more chat for your GPT. 

But it was not the brand new GPT-5 that some wonks had hoped for. 

Will ever-bigger models pass the bang-for-the-buck test?

As in any arms race, the driving force behind the spending is as much defensive as offensive.

None of the four wants to be left behind, lest it fall victim to disruption. 

Fortunately, the damage to society at large is likely to be limited. 

As with railway tracks and telecoms cables, overcapacity makes things cheaper. 

In many infrastructure booms, the benefits accrue to the users more than to those who lay the foundations. 

Do not be surprised if that happens again.