Listen to the audio version of this article (generated by AI).
Hello, Reader.
Tom Yeung here with today’s Smart Money.
The computers that took Apollo 11 to the moon were a marvel at the time.
To create each 70-pound machine, women hired from textile mills thread copper wire around thousands of tiny magnetic rings. Through a ring meant a “1”; around it meant a “0.” Engineers called it “LOL memory,” for the “little old ladies” who built it, and this mesh of rings formed the basis of the Apollo Guidance Computers.
That meant each of these machines was extremely expensive. Each navigational computer cost NASA around $200,000 apiece, or $1.8 million in today’s dollars. Two flew to the moon, while many more were used on Earth.
Barely three years later, the cost of computing plummeted. In 1972, Bill Hewlett dared his engineers to fit HP’s 40-pound desktop calculator into a handheld device. The result was the HP-35, the world’s first scientific pocket calculator. It cost around $3,100 in today’s dollars, and it could do almost everything the Apollo computers could.
By the 1990s, a solar-powered Casio fx-260 could perform the same calculations for around $9.
The falling cost of computing has also been true with AI.
That means the pricey AI infrastructure could lose value the same way the Apollo Guidance Computers did as more efficient chips and models emerge.
So, in today’s Smart Money, I’ll explain why the falling cost of AI computing could threaten today’s biggest AI investments – and what history’s $9 calculator can teach us about finding winners in the AI trade.
The Hidden Risk of Falling AI Costs
We can see how the cost of AI compute is falling by looking at a metric known as “cost per million tokens.” This measures how much it costs to have an AI model process or generate roughly 750,000 words, or the length of the King James Bible.
In late 2022, it cost $20 to output a million tokens, according to research group Epoch AI. Dedicated AI chips were only starting to become available, so AI computing prices were still relatively high.
By October 2024, that figure had fallen to 7 cents per million tokens thanks to the release of dedicated AI chips and more efficient AI models.
Today, the cheapest mainstream AI models cost just 2.5 cents per million tokens, and analysts expect another 90% decline by 2030.
That raises a thorny question for “neocloud” AI data centers like CoreWeave Inc. (CRWV). These companies must buy expensive AI chips today for roughly $90,000 apiece (once they include the other data center hardware) and recoup that investment by selling AI computing power before the chips either:
- break down or
- cost more in electricity to run than they earn in revenue.
It’s the same problem faced by bitcoin miners, where newer chips make older ones obsolete. (To their credit, many neoclouds have a safety net: contracts that require customers like OpenAI and Anthropic to pay for computing power, even if they don’t use it.)
That’s why companies like Citadel Securities are starting to warn that there is precious little margin for error if AI spending doesn’t pay off. Data center companies all know that AI use will go up… but they have not budgeted much breathing room for if AI computing costs come down. This is particularly alarming because so many of these projects are funded by debt.
Put another way, today’s AI chips cost $90,000 each because they are so hard to build. But much like the computers of the 1970s, AI chip prices are coming down because we’re getting better at building them. That means today’s highest-end chips will seem quite ordinary in a couple of year’s time.
That spells trouble for the AI data centers buying $90,000 chips today. If competitors can buy the same chip for $50,000… $30,000… or $3,100 in a couple of years, they’re going to undercut today’s big spenders and drive a lot of incumbents out of business. In turn, competitors to those firms will buy even cheaper chips, creating a downward spiral.
In other words, falling AI costs could create a chain reaction. Data centers could struggle to earn back what they spend on infrastructure, leaving them less willing or able to pay top dollar for the next generation of chips. And that puts the chipmakers benefiting from today’s high prices in a potentially difficult position.
When AI Gets Cheaper, Who Wins?
Artificial intelligence will change the world… but computing is getting cheaper so quickly that the companies investing billions to build AI infrastructure may not earn enough money to justify their investments. And if computing prices keep falling, even kingmakers like Nvidia Corp. (NVDA) could get hurt.
It’s a big question mark that offers little room for error.
That’s why Eric and I have continued to sound so cautious about the AI trade, especially when every miracle machine eventually becomes a $9 calculator. The question for investors is who’s holding the $200,000 version when that happens.
Now, we’re not arguing that AI is a bubble or that AI adoption will stall.
But benefiting from AI means looking beyond the chipmakers and AI data centers spending enormous amounts of cash today. We have no idea whether these infrastructure builders and chip suppliers will earn their money back.
That’s why knowing what to buy – and what to avoid – matters more than ever.
In Fry’s Investment Report, Eric Fry looks beyond the AI hype to find opportunities the market may be overlooking, while staying cautious about companies priced for perfection.
Until next time,
Thomas Yeung, CFA
Market Analyst, InvestorPlace
