<\/p>\n
(Bloomberg Opinion) — How AI will reshape our lives is clouded with uncertainty. Researchers are debating what it means to achieve artificial general intelligence, or AGI, or whether the neural-network systems that drive chatbots like ChatGPT can lead us to that big prize. But smart investors should avoid\u00a0trying\u00a0to predict future breakthroughs of a technology that is still at an evolutionary stage.\u00a0 <\/p>\n<\/div>\n
Rather than participating in OpenAI Inc.\u2019s expensive secondary share sales, they\u00a0can\u00a0get exposure to the AI theme using a time-tested approach. The supply chain is long, complex and prone to bottlenecks. At each chokepoint, there are\u00a0a few economic moats\u00a0\u2014\u00a0to use Warren Buffett\u2019s phrase \u2014\u00a0that deliver outsized earnings a lot faster and surer than ambitious, large language-model developers.\u00a0 <\/p>\n<\/div>\n
The most well-known constraint occurs in chip design. Demand for Nvidia Corp.\u2019s products is so strong\u00a0that OpenAI had to find alternative sources,\u00a0concluding a mega deal with rival designer Advanced Micro Devices Inc.\u00a0this week for inference functions. Inferencing, or\u00a0using pre-trained models to generate user content,\u00a0can make do with less advanced chips. <\/p>\n<\/div>\n
Dig deeper and there are\u00a0more obscure but equally interesting opportunities.\u00a0We are witnessing a\u00a0data-center boom, especially in the US. In July, construction spending reached $41 billion on an annualized basis, or 47% of the total outlaid\u00a0on offices. But developers fret over various bottlenecks, from procuring chips to securing power generation. <\/p>\n<\/div>\n
Try getting your hands on gas turbines, which\u00a0burn natural gas and turn a generator to produce electricity. It used to take around two\u00a0years to obtain a new one, now it\u00a0takes up to five or more. <\/p>\n<\/div>\n
This is in part due to an oligopoly market structure, dominated by\u00a0GE Vernova Inc. in the US, Germany\u2019s Siemens Energy AG and Japan\u2019s Mitsubishi Heavy Industries Ltd. After a gas turbine boom and bust in the early 2000s, the three producers have turned cautious and are\u00a0unwilling to expand capacity, fearing that demand from data centers may not stick. Meanwhile, startups are not trying to edge in either, because the\u00a0technological barrier to entry is high. In some cases, turbine designs took decades to test and tweak.\u00a0 <\/p>\n<\/div>\n
The market\u00a0for high-bandwidth memory, or HBM,\u00a0essential to support AI systems like those from Nvidia is another good example. South Korea\u2019s SK Hynix Inc. is the industry\u00a0leader, followed by Samsung Electronics Co. and Micron Technology Inc. The technology is so complex that Huawei Technologies Co. had to use Korean products in some of its leading Ascend AI processors, even though the Chinese\u00a0conglomerate would prefer to source domestically. Not surprising then that\u00a0OpenAI also inked\u00a0strategic partnerships\u00a0with Samsung and Hynix in an attempt to secure its supply chain and scale up before everyone else.\u00a0All three stocks are on fire this year.\u00a0 <\/p>\n<\/div>\n
To be sure, identifying AI winners\u00a0is not as simple as studying\u00a0tech entry barriers and market structures. Netherland\u2019s ASML Holding NV has a near-monopoly on the specialized lithographic equipment\u00a0used to make high-performance chips, but its stock price has missed this year\u2019s AI rally. One key reason\u00a0is a\u00a0concentrated customer base. Not being able to sell to the Chinese due to export restrictions, ASML unfortunately relies too much on foundry Taiwan Semiconductor Manufacturing Corp., which balks at purchasing the Dutch company\u2019s\u00a0ridiculously expensive\u00a0extreme-ultraviolet-light, or EUV, machines.\u00a0 <\/p>\n<\/div>\n
So what are the lessons? First, for investors without a computer-science background, identifying supply-chain bottlenecks is perhaps easier than asking if AI research can result in training models that exhibit human-like intelligence. After all, some money managers have already gone through a similar\u00a0exercise, looking at suppliers around the world during the smartphone era.\u00a0 <\/p>\n<\/div>\n
Second, the AI-fueled stock boom\u00a0is not limited to the US or China, because the supply chain is so complex that no country can claim to do all on its own.\u00a0 <\/p>\n<\/div>\n
Third, there have\u00a0been a lot of worries about an AI bubble \u2014 investors\u00a0are understandably wary when mega caps from\u00a0Oracle Corp. to\u00a0AMD can melt up\u00a0by more than 20% in a single trading day. But the divergent stock performance of, say, gas-turbine\u00a0makers and ASML\u00a0indicates that there\u2019s still rationality in this rally. Not all AI names get to shine, yet. <\/p>\n<\/div>\n
Most importantly,\u00a0in this AI race, the bulk of the\u00a0profit is probably\u00a0earned upstream, where the biggest bottlenecks are. By 2029, when OpenAI finally starts to break even, boring turbine engineering firms might have made a truckload of money already.\u00a0 <\/p>\n<\/div>\n
More From Bloomberg Opinion:<\/p>\n<\/div>\n
This column reflects the personal views of the author and does not necessarily reflect the opinion of the editorial board or Bloomberg LP and its owners. <\/p>\n<\/div>\n