By Tae Kim
Two years into the generative artificial-intelligence revolution, investors have shifted their attention from the technology's benefits to its sky-high cost. It's an inevitable part of the hype cycle.
Big Tech firms will spend more than $200 billion combined in capital expenditures this year, primarily to build out data centers, buy chips needed to train their AI models, and power their continuing use. The spending is real but so are the benefits, which could still surprise investors in the months and years to come.
"We're going to depreciate and wipe out the old infrastructure and build a new one," Philippe Laffont, who runs tech hedge fund Coatue Management, told attendees at his firm's annual conference this past July.
Over the past year, Laffont has gone beyond his traditional tech domain to buy utilities and industrial stocks. AI has jolted those businesses out of a long slumber.
"If you had told me five years ago that, as a high-beta alpha investor, I would now be a top investor in utilities, energy, and industrial, I would definitely bet a lot of money that would never happen," he says.
For this story, Barron's interviewed more than a dozen senior executives, including CEOs, chief technology officers, data center architects, and policy experts. They all said that AI was already changing work processes and knowledge retrieval, and increasing employee productivity. They portray the wave of spending by tech companies as necessary -- and justified.
Worries and flashbacks to the dot-com bust are understandable, but the AI data-center buildout isn't this century's fiber optics. AI data centers are more akin to highways and electrical transmission lines -- ultimately, they'll create a more productive economy. Investors shouldn't shy away from the opportunity.
Today's data centers that power cloud-computing platforms typically use about 50 megawatts of power. The AI data center will require 10 times that much.
Giordano Albertazzi, CEO of Vertiv Holdings, a leading provider of power and cooling infrastructure equipment for data centers, says his company's customers are building an increasing number of data centers with capacities of 500 megawatts or more. "What we see in the market is continued acceleration," he says.
One industry executive tells Barron's that he has already seen plans for at least 10 different data center projects exceeding 500 megawatts of power -- all slated for construction within the next three years.
The technology underlying AI continues to see the benefits of "scaling laws," which means that AI models continue to improve in line with how much computer power and data are used to train each successive model. Thus far, "there seems to be no end to scaling," says Microsoft Azure Chief Technology Officer Mark Russinovich.
To be sure, the costs to build new models are soaring. OpenAI's current GPT-4 model cost $300 million to train, according to estimates from Bernstein, with the next two models expected to cost several billion dollars and then more than $25 billion, respectively.
In early September, Elon Musk's xAI deployed more than 100,000 Nvidia Hopper graphics processing units. That GPU bill alone comes to several billion dollars. Musk said his AI start-up plans to double its capacity to an effective 200,000 GPUs in the coming months.
Oracle recently announced that it would deploy the largest-ever GPU cluster, available in the first half of 2025 and delivering up to 131,072 Nvidia GPUs for customers to train and run AI workloads. Not to be outdone, Meta Platforms CEO Mark Zuckerberg announced last month that his company is training Llama 4 on over 100,000 Nvidia GPUs. Meta's current Llama 3 model was trained on only 16,000 Nvidia GPUs.
The wave of one-upmanship spending, where each model training generation grows by a factor of 10, could make current predictions about data centers look quaint. BofA Global Research estimates that data center capex could rise by 14% a year, from $215 billion in 2023 to $311 billion in 2026. Look for those forecasts to go higher.
The arms race has been paired with plenty of skepticism, with some questioning the return on AI infrastructure. But there are two important innovations the skeptics may be underestimating: the ability of AI models to understand natural-language queries and the ability to apply AI computing power to derive insights from the vast pools of unstructured data inside companies.
Making employees more productive by providing the right information hidden away in internal databases quickly and easily doesn't drive chatbot-style attention, but it's no less disruptive and revolutionary.
"If you think about the world's data, the majority of it is sitting in cold tiers and archives," says Arthur Lewis, president of Dell Technologies' infrastructure solutions group. "Enterprises are now using that data to feed these models and realize the full value of the data."
Partha Ranganathan, a technical fellow at Google Cloud, says he's seeing more customers using AI to synthesize and analyze a large amount of complex data using a conversational interface.
Hewlett Packard Enterprise CEO Antonio Neri sees customers selecting a large-language model and fine-tuning the model with their own data that is "unique and verticalized to their needs."
"No. 1, AI is a tool for enhancing business productivity," he says.
"Cat pictures and funny essays are not the exciting part of AI to me," says Jon Lin, an executive vice president at data center operator Equinix. "Fundamental materials discovery, drug discovery, and improvement to health is the potential. That's what gets me excited about doing the things we do every day."
Bristol Myers Squibb purchased a large Nvidia AI system a year ago, and the company says it's already paying dividends. "It has accelerated our capability," says Greg Meyers, Bristol's chief digital and technology officer. "Finding new medicines to tackle really challenging diseases is as much a computational problem as it is a scientific problem."
Meyers says that AI has allowed Bristol's drug discovery scientists to iterate and research additional permutations of its drugs. It also allowed the company to take advantage of data compiled from prior clinical trials to better design future ones. "We're on track as a result of the AI work to reduce our clinical trial cycles time by almost two years," he says.
Megacap technology firms are touting the unprecedented growth trajectories of their new AI business ventures. Amazon.com CEO Andy Jassy recently told investors that the company's generative AI business is on track for a "multibillion-dollar revenue" annual run rate. Year over year, the business is more than doubling at a growth rate surpassing the early stages of its Amazon Web Services cloud business.
Generative AI "is a really unusually large, maybe once-in-a-lifetime type of opportunity," Jassy said on Amazon's recent earnings call.
Microsoft CEO Satya Nadella has similarly touted the supercharged growth of its cloud business. He said last month that the company's AI business is on track to exceed a $10 billion annual run rate this quarter, making it the "fastest business in our history to reach this milestone."
Meanwhile, Big Tech CEOs are also realizing significant productivity gains for their own businesses through the use of AI. Meta's Zuckerberg told investors last month that AI-driven feed and video recommendations have led to an 8% increase in time spent on Facebook and a 6% increase on Instagram this year. More than a million advertisers have used the company's Gen AI tools to create ads, and early results show a 7% increase in conversions to revenue, he said.
Alphabet CEO Sundar Pichai said on the company's recent earnings call that more than a quarter of new code at Google is now initially generated by AI and then reviewed by staff. Seven of the company's major products -- with more than two billion users -- have incorporated Google's AI Gemini model, Pichai said. "Our long-term focus and investment in AI are paying off and driving success for the company and our customers."
In the latest September quarter, Alphabet, Microsoft, and Amazon all boosted their capital expenditures by 62%, 51%, and 81%, respectively, versus the prior year. The total capex bill across those companies, plus Meta, could top $230 billion this year.
Microsoft said it has been capacity-constrained for Azure AI services, meaning it doesn't have enough AI GPU servers to meet current demand. It plans to increase AI infrastructure capacity in the first half of 2025, while Alphabet and Amazon vowed to spend more next year. Meta also said it expects "significant capital expenditures growth" for 2025.
Each company is chasing one product: Nvidia's GB200 NVL72 AI system. The AI-server incorporates 72 GPUs, linked together inside one server rack roughly two feet wide, four feet deep, and seven feet tall. That small package of computing power differentiates Nvidia from its rivals.
Vertiv's Albertazzi calls the NVL72 an "awesome value proposition."
Customers are clamoring for the NVL72 because it's far more efficient and powerful than prior models, saving companies money on overall AI model training and queries. Nvidia says the GB200 NVL72 provides up to 30 times the performance of its prior H100 GPUs for large-language model inference, the process of generating answers from AI. It's four times faster for training AI models, the company says.
The buzz surrounding the GB200 is so intense that Microsoft, OpenAI, and Google have all featured photos of the Nvidia server on their social-media accounts in the past month, seeking to bask in some of Nvidia's innovation afterglow.
To be sure, AI's early profits won't be evenly spread around the market. The arms dealers in the AI race remain the best way to play the opportunity, with three stocks particularly well positioned: Nvidia, cooling specialist Vertiv, and cloud provider Oracle.
(MORE TO FOLLOW) Dow Jones Newswires
November 15, 2024 02:00 ET (07:00 GMT)
Copyright (c) 2024 Dow Jones & Company, Inc.
Comments