Will AI do to Nvidia what the dot-com boom did to Sun Microsystems? Analysts compare current hype to past ones.

By Wallace Witkowski

The dot-com boom and server demand drove the market cap of Sun Microsystems to $200 billion in 2000. By 2009, Oracle picked it up for $6 billion.

It's different this time: The two-syllable mantra that has echoed across every boardroom and C-suite this year, whipped investors into a frenzy and sent businesses chasing their own is going to make everyone tons of money.

If the year in question is 1999, that mantra is "dot-com," and it lost investors a lot of money. In 2023, the mantra is "AI." And because the dot-com frenzy fell well short of its hype and took the market down with it, analysts at Bernstein Research are asking whether artificial intelligence has the potential to be the same kind of expensive bubble. Or will it be different this time?

Ever since OpenAI, backed by billions of dollars from Microsoft Corp. (MSFT), released ChatGPT on Nov. 30, talk of generative AI and what it could do smoldered for the rest of 2022 and then broke out into a wildfire in 2023.

A team of Bernstein analysts led by Toni Sacconaghi and Stacy Rasgon examined three historical instances where hype and the promise of revolution loosened billions of dollars in capital investment -- server sales during the dot-com era, the huge buildout of telecom capacity in the 1990s, and the buildout of cloud infrastructure 10 years ago -- to gauge how the current rush to spend billions on hardware and services to support AI may turn out.

Could Nvidia be the next Sun Microsystems?

"While a number of observers have described Nvidia's (NVDA) growth acceleration as 'unprecedented', we look to server sales during the dot.com era as a precedent, best exemplified by Sun Microsystems," the Bernstein analysts wrote, noting that Sun Micro's market capitalization peaked at $200 billion during the dot-com boom, only to have the company acquired by Oracle Corp. (ORCL) in 2009 for under $6 billion

"Sun's experience helps highlight that upstream hardware vendors often see much more violent swings than end-user demand," the analysts said.

Another risk that has a "potentially eerily similarity to Nvidia today," Bernstein analysts said, is that much of the incremental server demand came from dot-com startups -- many of which proved to be unviable businesses.

Nvidia stands to clean up, with expected data-center revenue to grow $15 billion, nearly double year over year, with an additional $20 billion in 2024, Bernstein Research estimates. And as some analysts pointed out last week, even bullish estimates may be too conservative, Bernstein noted, given that Microsoft, Meta Platforms Inc. (META), Alphabet Inc.'s (GOOGL)(GOOGL) Google Cloud Platform and Beijing-based ByteDance Inc. alone might account for Nvidia's year-over-year guidance.

Read: Nvidia gets more good news from Big Tech, even as AI spending 'may not lift all boats'

"This risk appears particularly salient considering that major user-facing AI applications don't appear likely to scale before 2025, and a good amount of the AI cloud demand in the near term seems to be coming from pre-product, [venture-capital] backed startups and enterprises testing the viability of using Generative AI without budgets and likely corporate commitments to support full usage," the Bernstein analysts said.

Read: Amazon, Microsoft and Google cloud services bet heavily on AI, but do their customers even want it?

Lucent and the telecom buildout of the 1990s

A "more ominous comparative build of capacity" could be seen in the Lucent, Nortel and broader fiber buildout in the 1990s, the Bernstein analysts noted.

"Companies raced to build out fiber networks as quickly as possible to serve dial-up internet, outpacing demand, an issue which was exacerbated as cable TV companies, realizing they were well positioned to offer broadband that was superior to dial-up, began launching competing offerings," the analysts wrote.

"By 2001, companies had laid about $90 billion worth of fiber, and only 2.6% of that capacity was in use, and it is estimated that to this day there remains tens of thousands of miles of fiber that is still 'dark' or unused," according to Bernstein Research.

Partly as a result of the tech and telecom bubbles and overbuilds, only two of the top nine tech companies by market cap in 2000 have ever returned to their peak market cap, and neither of those -- Microsoft and International Business Machines Corp. (IBM) -- were involved in building fiber networks, the Bernstein analysts noted.

"While Nvidia's ecosystem is growing increasingly entrenched, given we are in the early stages of the transition to AI there remains a risk that workloads could ultimately shift to a different architecture (broader adoption of custom silicon in the cloud, for instance) just as how internet shifted from dial-up to broadband," the analysts said.

"We also have yet to see meaningful attention to optimizing generative AI usage, which could impact demand," they noted. "Lastly, we could see alternative technologies that require different or less technology supplanting generative AI (and thus the demand for GPUs)."

Microsoft's Azure cloud buildout, and negative margins circa 2013

Bernstein analysts also compared the current AI buildout to Microsoft's investment in Azure starting in 2013.

While Microsoft's cloud service -- and main competitor to Amazon.com Inc.'s (AMZN) Amazon Web Services -- ran as much as a 15% negative margin in the early years, Azure crept out of negative territory and by 2021 gross margins had soared to 66%, the Bernstein analysts noted.

"One interesting nuance is that the early build for Azure came largely in the form of buildings, as Microsoft first built out physical data centers globally to meet the demand of both Azure and Microsoft Office 365," the analysts wrote. "They then filled these data centers with servers to match capacity with demand as demand came online."

"There appears to be significant building and improvement additions required by Generative AI, as AI servers tend to run hotter and use more power than traditional servers, and thus require specialized data centers," they said.

Nvidia has fed into that, with Jensen Huang, its founder and chief executive, announcing a slew of products and services targeted at expanding AI development in March.

As for the current hype, Bernstein analysts said it all depends on how much risk one wants to take investing in AI.

Acknowledging that "gauging the magnitude and trajectory of capacity additions and AI demand may be foolish to predict," the Bernstein analysts estimate another 12 to 18 months or more of "unbridled AI infrastructure buildout, but things look murkier beyond that."

Read: Is AMD ceding ground to Intel? Wall Street will soon find out

"We believe that Nvidia and other AI infrastructure hardware vendors offer the most near term upside potential, but carry the most risk of an overbuild," the analysts said.

While Bernstein did not specifically mention it, Advanced Micro Devices Inc. (AMD), which held its AI product launch in June, is regarded as a distant second to Nvidia when it comes to AI data-center hardware market share.

AMD reports its quarterly results after the close of markets Tuesday, following Intel Corp.'s (INTC) relatively strong results last week.

AMD's shares are up 77% year to date, while the PHLX Semiconductor Index is up 52.5%, the S&P 500 is up 19% and the tech-heavy Nasdaq Composite Index has risen 37%. Nvidia shares have soared 220%.

A lower-risk approach to investing in AI is through the cloud-service providers, or "hyperscalers," as the industry refers to them, like Microsoft's Azure, Amazon's AWS or Google Cloud Platform.

Such a play will "likely offer similar longer term upside, with comparably lower risk, at lower valuation multiples," the analysts wrote. "We see their downside scenario being analogous to Amazon's delivery network overbuild in 2021-22."

-Wallace Witkowski

This content was created by MarketWatch, which is operated by Dow Jones & Co. MarketWatch is published independently from Dow Jones Newswires and The Wall Street Journal.

(END) Dow Jones Newswires

08-01-23 0759ET

Copyright (c) 2023 Dow Jones & Company, Inc.