By James Mackintosh
Where will the money be made next in artificial intelligence?
Bankers are gearing up to sell SpaceX, Anthropic and OpenAI to investors in a series of giant IPOs. Should investors just buy anything with a hint of AI, or try to find the parts of the AI supply chain, the "stack, " where bottlenecks capture more of the value?
Searching out the bottlenecks looks tempting because the returns so far for anyone who got it right have been extraordinary.
First up, after the launch of ChatGPT in 2022, was Nvidia. Its graphics-processing units, or GPUs, turned out to be the best way to train the large language models driving the new AI. Soaring demand meant Nvidia could keep jacking up the price of new chips, and its shares are up 15-fold since November 2022 as profits rocketed.
This year, new bottlenecks emerged even as Nvidia slowed, facing competition from chips developed by Amazon, Alphabet and Cerebras, which recently went public. Makers of the high-bandwidth memory needed for AI data centers have beaten Nvidia hands-down this year, with Micron Technology tripling as, again, chip prices rocketed. Using AI also requires more central processing units, or CPUs, compared with GPUs, and CPU veteran Intel has more than doubled in two months.
Meanwhile, the bubble-era excitement that fueled developers of fuel cells in 2000 and 2021 has returned as Bloom Energy won data-center supply contracts. Its shares have beaten Micron. On the flip side, AI software firm C3.ai has given back all its gains after the 2023 excitement that propelled it to an almost fivefold gain in six months.
The basic argument for spreading bets is that AI is changing so fast that it's hard to anticipate the bottlenecks.
It was only a year ago that the memory stocks emerged from the doldrums, while as recently as September Intel shares were lower than when ChatGPT launched. Bloom, meanwhile, was languishing below its 2018 IPO price at the end of May last year.
Tom Slater, head of U.S. equities at Baillie Gifford, argues for spreading bets across the stack. This ranges from the layer of software applications at the top, through the providers of AI "agents" and the model developers, down to data management, chip providers and data centers, all the way to the suppliers of electricity.
"Some layers of the stack will be vastly overvalued at today's prices, and some will turn out to be vastly undervalued," he says. "The idea is that the ones that are vastly undervalued make up for the rest."
He remains a stock picker within the layers. Baillie's flagship London-listed investment trust, Scottish Mortgage, has almost a fifth of its assets in still-private SpaceX. The trust bought in eight years ago when the idea of data centers in space -- founder Elon Musk's latest pitch -- was pure science fiction (skeptics argue it still is). Elsewhere in the stack, Baillie owns model-developer Anthropic and Databricks, which makes software to support AI data management.
Others are trying to decide whether AI models will be a winner-takes-most market or a highly competitive commodity, how long chip bottlenecks will endure, the possibility of pricing power for data centers and the rewards for the usually dull business of supplying power.
The alternative is to try to own everything. As more AI stocks tap public markets to satisfy their vast capital needs, a widely diversified AI portfolio becomes possible.
This guarantees owning the winners, but also the losers. The history of the phenomenal gains and losses in the dot-com bubble suggests what mattered most was buying at the right moment -- the big winners came early and were cheap as a result.
If you put $100 into the S&P 500 at the time of each of the 490 internet IPOs between 1996 and September 2000 tracked by Jay Ritter, emeritus professor at the University of Florida, and reinvested the dividends, you'd have $479,000 today.
If you had put $100 into Amazon in 1997, you'd have made almost half as much from it alone, $280,000. Some of the other 489 IPOs worked out nicely, too, but the overall picture of a diversified internet portfolio was dismal, despite Amazon's phenomenal performance.
Investors today frequently talk of finding the next Amazon. If you picked the actual Amazon and stuck with it, you're probably reading this from your private island.
But even for Amazon, what mattered was your timing. In just over two years, from its 1997 IPO to its peak in December 1999, it turned $100 into more than $5,000.
If instead you put $100 in at the peak, the gain has been about the same, but over 26 rather than two years. That still crushed the wider index, as Amazon has proved one of the best investments of all time.
But from December 1999 to February 2009, an investor had to spend almost 10 years trailing far behind the S&P. To hold on required a lot of faith that the online bookseller would expand to become the world's everything store, mutate into a cloud provider, and then profit from AI.
Good luck picking the next one, and picking it at the right moment.
Slater thinks a new set of application-layer disrupters is on the way that will use AI to overthrow the existing tech leaders. "The companies that will define the next decade are yet to be decided," he said at the Quality-Growth Investor Conference in London.
Only one thing is certain. Wherever they sit in the stack, they will be hard to tell apart from eventual losers, and doubly hard to buy before everyone else notices.
Write to James Mackintosh at james.mackintosh@wsj.com
(END) Dow Jones Newswires
May 30, 2026 23:00 ET (03:00 GMT)
Copyright (c) 2026 Dow Jones & Company, Inc.
Comments