|  | | | | AI: Real Deal or Mega Bubble? |
| In September of 1995, a few dozen of the world’s most successful business tycoons gathered at the luxurious and sprawling K Club in Kildare, Ireland, about 25 miles southwest of Dublin. The event was organized by Warren Buffett, who periodically arranged such get-togethers. The gentlemen spent a few days playing golf, eating steaks and trying to learn from one another how they could all get even richer. |
|  | The K Club - Kildare, Ireland |
| It was an inspirational and grandiose setting at a moment when tremendous technological changes were afoot. Just a year prior, Jeff Bezos had founded Amazon.com in his garage. It was the dawn of internet. Opportunity was knocking.
Today, the world is in a very similar place. Artificial intelligence, or AI, is widely considered the next chapter in the story of the computer age.
AI technology allows computers to perform cognitive tasks that were previously limited to human beings. These include reasoning, learning, problem-solving, perception, interpreting language and expressing ideas. Very importantly, these cognitive tasks can now be performed at a pace and on a scale that vastly exceeds the capabilities of any human brain.
AI has been developing in the lab for a long time. As far back as 1952, a computer was trained to play checkers. But thanks to a number of recent breakthroughs, AI is finally ready for large scale commercial applications.
On the software side, Learned Language Models (LLMs), such as those utilized by OpenAI’s ChatGPT, now provide the logical framework for computers to think like humans. On the hardware side, high-powered Graphic Processing Units (GPUs), such as those sold by NVIDIA (NVDA), support the vast number of computations that AI programs require.
In the eyes of many, the possibilities for AI are endless and the profit potential is unimaginably large. Investors are giddy, as billions of dollars are being spent to develop the world’s AI computing capability. In anticipation of future earnings growth, tech stocks have been flying, while the entire economy is benefiting at least indirectly from the ongoing investment boom.
But many observers think we have seen this movie before. Dotcom mania led to a massive investment spree that ultimately crashed and burned. While the internet would of course go on to become an incredible force, many early investors badly misjudged how quickly its potential would materialize and where the real profit opportunities would ultimately be found.
We are confident AI technology will have an enormous impact on the economy and will create substantial upside for many investors in a wide range of businesses. At the same time, there are potential excesses in the marketplace now which do present investors with genuine risks. The AI investment opportunity is real, but it needs to be approached with caution and selectivity.
To avoid the pitfalls, we have much to learn from the mistakes of those who participated in the late 1990s technology bubble. Especially interesting is the experience of the so-called “fiber barons,” who rapidly put billions to work in pursuit of internet riches that never came to them.
Buffet’s gathering in Ireland in fact set the stage for one of the most extraordinary technology misfires in American corporate history.
Among the invitees was Walter Scott, Jr. of Omaha, Nebraska, whose family owned a 100-year old construction firm called Peter Kiewet Sons’ Company. Because the business generated more cash flow than it needed, the family tended to make a wide range of diversified investments outside of the construction space.
Kiewet made one particularly important investment in the late 1980s, just a few years after Ronald Reagan’s DOJ broke up Ma Bell and deregulated the telecommunications industry. An ambitious executive named James Crowe guided Scott to buy a fiberoptic cable company known as Metropolitan Fiber Services (MFS).
Crowe was the son of a highly decorated World War II Marine Corps major. He arrived at Kiewet through the acquisition of another construction company. Crowe was not a military man, but he has been described as having the demeanor of one.
The MFS deal ended up being a homerun for Scott and his family. Thanks to Crowe, MFS would become one of the largest Competitive Local Exchange Carriers (CLECs) in the United States. The business was eventually spun-off with a valuation in the billions.
Both Omaha natives, Walter Scott and Warren Buffett knew each other well. As Warren explained: “We both had a crush on the same girl, Carolyn Falk. She was the daughter of my father's business partner. Walter won her, and they ended up getting married."
Warren eventually got over the loss and was happy to bring his hometown friend Walter into his golden circle of business titans. The group now included Microsoft founder Bill Gates, who also carved some time out of his busy schedule to join Warren and his friends at the Irish resort.
Gates made a big impression on Walter Scott. He spoke to the group about how revolutionary the internet was going to be and how it would disrupt the traditional telecommunications market.
When Scott returned to Omaha, he conferred with Crowe to figure out how to take advantage of Gates’ prophesy. Less than a year later, in April 1996, the two men decided that MFS should buy internet backbone provider UUNet for $2 billion, which was then a record price for an internet company.
The UUNet deal worked out well, because it transformed MFS into a true internet play. This almost immediately prompted the infamous Bernie Ebbers of WorldCom to acquire MFS. In August of the same year, WorldCom purchased MFS for $14 billion, another record-setting transaction.
If Crowe wasn’t in the big leagues before, he definitely was now. Around Omaha, he was the man with the Midas touch. Many Kiewet employees got quite rich thanks to him. Plenty of people were ready to back his next move, and with the internet in its infancy, he was just getting started.
Crowe was now head of all of Kiewet’s diversified activities, which were already publicly traded as an independent business unit. He hatched another business plan. He would build the premier fiberoptic network in the country. It would extend thousands of miles and be able to handle the explosive growth in internet traffic that was about to materialize. |
|  | James Crowe and Walter Scott, Jr. |
| Crowe renamed the business entity Level 3 Communications, which began trading on the Nasdaq under the ticker LVLT. (The high-tech sounding name was derived from terminology used by fiberoptic engineers.)
The business case behind Level 3 was quite simple. The internet was an incredible new technology. It was going to transform the world, and data traffic would be incalculably large. Level 3 was going to be the premier platform to support all of it.
Those who remember the late 1990s tech and telecom bubble might recall the name Jack Grubman. He was the Salomon Brothers analyst who later became the poster child for all the conflicts of interest that overtook Wall Street stock research. Grubman himself was making a fortune by helping his firm’s investment banking clients raise money through his wildly positive predictions about their business prospects.
Grubman became one of Level 3’s loudest cheerleaders. He promoted the narrative that the internet was a revolutionary technology with seemingly infinite future demand potential. In his initiation report on LVLT, Grubman seemed to imply that providers of internet bandwidth were, by extension, infinitely valuable. |
| | Level 3 is a great play on bandwidth...[which] will be chronically scarce. Capacity actually creates demand in this business...bandwidth-centric names are good values at any price since nobody can predict the true demand caused by growth…. Like the attic of a house gets filled, no matter how much bandwidth is available, it will get used. - Telecom analyst Jack Grubman |
| | Like almost all of the stocks under his coverage, with perhaps a few exceptions to create a veneer of credibility, Grubman aggressively recommended shares of LVLT from the get-go. For a while, he seemed to have it right. (To be fair, he was one of many Wall Street analysts vying to become captain of the cheerleading squad).
From its April Fool’s Day 1998 debut until shares peaked, along with the rest of the market, on March 10, 2000, LVLT investors made some 250% in less than two years. But like almost every other “New Economy” stock from that era, it was all down hill after that. |
|  | By the second half of 2001, shares of LVLT were trading down some 90% from their 1998 IPO price and about 97% from peak levels. Having borrowed billions of dollars from the junk bond market to fund the network buildout, the company did not die but was drowning in debt.
A cautionary tale
We could have chosen from scores of companies to highlight the perils of “irrational exuberance,” as former Fed Chair Alan Greenspan once described the dotcom era. We chose Level 3 specifically because of its direct relevance to the current environment.
Investors in Level 3 and peer companies in the late 1990s had wildly aggressive expectations around internet bandwidth demand. They funded business models that relied on these ambitious forecasts and applied valuations to stocks on the basis of these forecasts actually materializing.
When it came to Level 3, the basic idea supporting the business model was sound—in a sense. The world would need a lot of fiberoptic cable to accommodate eventual growth in internet traffic. But investors had gotten way ahead of themselves.
By the early 2000s, thanks to aggressive investment by companies like Level 3, Qwest Communications, Global Crossing and others, there was a glut of fiberoptic capacity. Analysts estimated low single digit utilization levels, as the world was still largely accessing the internet through old-fashioned copper wire dial-up technology. It would be many years until wireline and wireless broadband would be deployed at scale.
Fiber supply overwhelmed demand. As a result, there was a plunge in the unit pricing that these fiberoptic backbone carriers could charge their customers, which included Internet Service Providers like AOL.
Not only was the system overbuilt, the incremental cost of building internet backbone capacity started to collapse thanks to technological advances. The billions that companies like Level 3 plowed into fiberoptic networks had to be written down dramatically. New capacity could be created just a few years later at a fraction of the cost.
Is the AI boom just like the fiber glut?
There are numerous potential parallels between the late 1990s fiber investment wave and the ongoing AI investment boom. In a very broad sense, they are premised on the same idea: enormous growth in data consumption. A distinction would be that in the late 1990s, the emphasis was on massive growth in data traffic, whereas today, the focus is on enormous growth in data processing, specifically for AI applications.
In the 1990s, the idea was that those who create the capacity that will transmit all the data will earn enormous profits. Today, the thinking is that those companies who provide the capacity to process all the data will see tremendous profit growth—and their shares are now priced accordingly.
Do the valuations now attached to AI stocks make any sense? There may be no more important question facing all investors right now—not just shareholders of direct AI plays like NVDA. |
| Thanks to many years of outperformance, over 30% of the S&P 500 Index is now comprised of stocks classified as Information Technology, whereas the average weighting of the other ten sectors is close to 6%. Moreover, many large capitalization stocks that most people might think of as “tech” are classified under other sectors.
In fact, only three stocks within the Magnificent Seven are, technically speaking, tech stocks. Amazon (AMZN) and Tesla (TSLA) are classified as Consumer Discretionary, while Alphabet (GOOG) and Meta Platforms (META) are considered Communications Services. All of them are perceived by investors as AI beneficiaries, if not leaders.
In the context of the broader market, tech and technology-related stocks are as important as ever. And investors have priced in AI-driven growth, with tech sector P/E multiples having expanded significantly over the past five years. By contrast, the earnings multiple of the S&P 500 on an equal weighted basis (which dilutes the impact of tech within the index) is largely in-line with historical levels. |
|  | Many companies outside of the technology space have benefited from the AI investment boom as well, as has the entire economy. As we saw after the tech bubble started to burst in 2000, negative wealth effects and plunging investment demand can have significant impacts on the broader economic picture.
The macroeconomic story of the early 2000s is complicated by the 9/11 attacks and subsequent wars and geopolitical turmoil. But the post-dotcom recession began in March 2001. Ten-year Treasury yields were as high as 6.7% in January 2000 and fell to 4.9% in August 2001, just prior to the terrorist attacks.
Investors may recall as well the enormous shift within equity markets from New Economy growth to Old Economy value stocks. Many boring, low-tech companies that had been abandoned by investors because they lacked a clear internet angle later benefited from the economic stimulus that followed the tech crash.
In the early 2000s, Greenspan’s Fed unleashed tremendous monetary stimulus. The Fed Funds effective rate was brought down from over 6.5% at the height of the tech bubble to 1% in 2004. This set the stage for the eventual housing bubble, which propped up many brick-and-mortar sectors along the way. Value outperformed growth substantially. |
|  | The implications of whether or not we are on the precipice of crash in AI stocks, reminiscent of what we experienced in 2000 to 2001, therefore go beyond the direct impacts on business models connected with AI spending. The fate of AI stocks could have a major impact on key macroeconomic variables like interest rates.
Tech stocks have been on a tear since bottoming at the end of 2022 after investors priced in Fed tightening, which was prompted by the post-Covid, post-Ukraine inflation shock. The recovery has largely been fueled by enthusiasm around AI, particularly the mind-boggling success of NVDA.
By year-end 2023, the tech-heavy Nasdaq Composite Index had already surpassed its 2021 highs. Between year-end 2023 and the market peak on July 10, 2024, the Nasdaq gained another 24%, with NVDA shares up more than 170% in this time frame.
Since July 10, we have seen significant volatility in NVDA and other tech stocks, which we have covered. A number of factors have contributed to this retrenchment, including rather stretched valuations across the tech sector. As indicated above, the earnings multiple of the SPDR Tech ETF (XLK) briefly surpassed 40 times, a historically steep level.
Skeptics emerge
The investment community has also been alarmed by a number of reports from respected sources suggesting AI-related stocks are overvalued, potentially by a large margin, and that expectations around AI revenue cannot possibly materialize.
On June 20, David Cahn of Sequoia Capital, a premier venture capital firm, published an update to his note from September 2023, in which he raised the question of whether sufficient revenue is being generated by AI applications to justify the massive investment that is taking place. |
| | At that time, I noticed a big gap between the revenue expectations implied by the AI infrastructure build-out, and actual revenue growth in the AI ecosystem, which is also a proxy for end-user value. I described this as a “$125B hole that needs to be filled for each year of Cap Ex at today’s levels.” This week, Nvidia completed its ascent to become the most valuable company in the world. In the weeks leading up to this, I’ve received numerous requests for the updated math behind my analysis. Has AI’s $200B question been solved, or exacerbated? If you run this analysis again today, here are the results you get: AI’s $200B question is now AI’s $600B question. - David Cahn, Sequoia Capital |
| | Cahn’s conclusion was that the further run-up in AI stocks has only enlarged an already tremendous gap between AI revenue expectations and infrastructure investment costs. |
| The conversation continued online when the venture capitalists who host the widely viewed All-In Podcast, including David Sacks, tackled the issue at length. For humorous and accessible commentary on key issues affecting the tech sector, as well as politics, we highly recommend the All-In Podcast in general and this episode in particular (the AI discussion begins just before the 21 minute mark). Every tech-focused investor we know follows the show—some quite religiously. |
|  | | Another expression of skepticism around AI valuations came from hedge fund Elliott Management, which released a private letter to investors that became a major story on August 2 in the Financial Times: Elliott says Nvidia is in a ‘bubble’ and AI is ‘overhyped’: Hedge fund tells clients many supposed applications of the technology are ‘never going to actually work’.
We got a hold of a copy of this short letter, which offered some back of the envelope criticism of AI stock valuations. In particular, Elliott took aim at NVDA, whose Graphic Processing Units (GPUs) were described as being aggressively bid-up as a result of an “arms race” among the big four “hyperscaler” AI cloud providers (META, GOOG, MSFT and AMZN).
Due to the absence of a “killer app” for AI, as well as serious privacy concerns, Elliott appears to view the high revenues and high margins associated with NVDA’s data center business as unsustainable. According to Elliott, the current AI bubble is like the previous tech bubble, just “larger and more entertaining.”
One crucial point, which Elliott alluded to in passing, is that, this time, the key players are enormous and highly profitable. They are funding their investments in AI capacity through the massive cash flow they generate and their pristine balance sheets. This is a marked contrast with the late 1990s tech bubble, which on the dotcom side was largely funded by investors in profitless start-ups, including many retail investors, and on the telecom/network side by enormous junk bond borrowing.
Elliott estimates that the big four hyperscalers are spending $160 billion this year on data center cap ex. While $160 billion is clearly a large figure, it is worth remembering that the market capitalization of those four companies is roughly $8 trillion. So the investment they are making is only about 2% of their collective equity value. In a real sense, the money they are investing in these projects is almost immaterial to their current valuations.
It is also worth emphasizing who is allegedly misallocating the capital now. In the late 1990s, we had a speculative frenzy, with retail investors bidding up ridiculous business models in hopes of getting rich quick. Venture capitalists took advantage of a white hot IPO market, while telecom entrepreneurs like James Crowe borrowed billions to fund their grandiose plans.
Today, we have the largest and most successful businesses in the history of the world driving these AI investments with their own money. They may indeed be overzealous, but these are the multi-trillion dollar companies that more or less gave us the internet revolution in the first place. It’s possible they know what they are doing.
Navigating the risks
From an investment perspective, we are less concerned about low potential returns on data center investments, in the sense of potentially squandered capital. Rather, we are more concerned about the valuations now applied to these leading AI plays, which seem to be predicated on very robust future cash flow growth, largely related to the investments they are currently making.
When it comes to the Magnificent Seven, as we noted about a month ago as markets peaked (see video below), these stocks now represent a huge exposure for most index investors. And with elevated earnings multiples that rely on AI growth materializing, the Mag 7 are vulnerable to unique valuation risks. |
|  | | In the bearish scenarios described above, in which current levels of investment in AI data centers turn out to be excessive, NVDA strikes us as one of the more vulnerable valuation situations. NVDA is a business that we have approached with cautious admiration. (In case you missed it, we wrote about NVDA here and had a conversation about it here.)
NVDA shares have surged because the company’s profitability has seen explosive growth and is expected to continue to grow at a healthy rate in the years to come. Abnormally high profits are wonderful but also dangerous in that they may not be sustainable.
In AI today, we essentially have a bottleneck. NVDA is the leading provider of GPUs, which function as the brains within AI data centers. There is some competition on the horizon, but NVDA GPUs are the best and most coveted. In essence, NVDA is now able to charge an arm and a leg for its GPUs, which translates into high prices, which translates into high revenues and high profit margins.
Analyst forecasts call for approximately $80 billion in operating profit for NVDA this year, which represents an approximately 65% operating margin. As a thought experiment, imagine NVDA had a serious competitor and had to drop unit prices by 25% (while operating costs remained the same). Roughly speaking, we calculate its operating profit would fall 35% to $50 billion.
AI winners and losers
Investors bullish on NVDA insist that its intellectual property and technological edge will allow the company to preserve its premium pricing and enormous profitability. But to the extent NVDA pricing is eventually squeezed by competition, it would actually be a boon for NVDA’s hyperscaler customers, who would simply be able to buy more GPUs for less. (This potential dynamic can be compared with internet service providers in the early 2000s, who were able to lease oversupplied internet backbone capacity at lower rates).
Pushing this scenario forward, if GPUs become more commoditized and cheaper, this frees up capital to spend on other aspects of AI data center development and operation. Companies that in one way or another help bring AI services to end users will be able to deliver more computing power for less money. End users themselves will be able to consume more computing power because it is less expensive to access.
If NVDA’s profitability is indeed unsustainable, this would be bad for NVDA shareholders. But it should be accompanied by falling AI computing costs, which will only stimulate more AI demand and innovation.
We have profiled a number of stocks in the 76report, which we hold within our Model Portfolios, that we believe are solid long-term AI beneficiaries. Each of these stocks stand to benefit from potential longer term commoditization of GPUs.
Oracle Corporation (ORCL): ORCL’s growth strategy is largely based on its cloud infrastructure business, which unites its core enterprise software operations, built around its leadership in databases, with cloud hosting and AI processing capabilities. ORCL’s Larry Ellison famously reported that he had to beg NVDA’s Jensen Huang for GPUs. As a GPU customer, the less ORCL needs to pay for GPUs, the more value ORCL can bring to its own customer base.
Texas Instruments (TXN): TXN is the leading supplier of analog semiconductors that are largely used in industrial and automotive applications that will proliferate as AI-driven technologies become embedded across multiple sectors. AI is unlocking the potential of the “Internet-of-Things.” The more accessible and prevalent AI becomes, the more demand there will be for sensors and other systems that cannot function without analog chips.
Digital Realty (DLR): DLR is one of the leading global operators of data centers, which are becoming a scarce form of real estate as they get increasingly repurposed for AI computing. Falling GPU prices should only increase demand for data centers in that tenants can purchase more GPUs for less money. An analogy can be made to falling automobile prices benefiting owners of parking lots.
Freeport-McMoRan (FCX): FCX is among the largest copper miners in the world and a beneficiary of the electrification theme. Lower GPU pricing means more AI, which means more demand for electrical wiring. More demand for electrical wiring means more demand for copper, which is extremely supply constrained.
Demand will come
One element of the Elliott later that we think will not age particularly well is the curmudgeonly attitude towards AI applications. In the letter, generative AI is said to have “few real uses aside from areas such as summarizing notes of meetings, generating reports and helping with computer coding.”
While the commercialization of AI is certainly in very early stages, it strikes us as short-sighted—and inconsistent with the experience of the last three decades—to discount the transformative potential of AI.
As a technological innovation, AI itself truly has no boundaries. It is simply about using advanced computing capabilities to access, manipulate and generate information in all its forms. Use cases for AI are not even limited by the human imagination, as AI itself has the potential over time to expand its own capabilities.
Just in the production of this report, we have begun implementing new AI tools with which we are starting to familiarize ourselves. Much of our research was done through a recently introduced AI-based search engine called perplexity.ai, which we have begun to use and recommend.
We are genuinely benefiting from this enhancement of traditional internet search, because of the ability to access information in a narrated and sourced way, rather than as a list of relevant website links. Similarly, as you may have noticed, we are experimenting with generative AI graphics tools, which are extremely easy to use.
A friend of 76research recently sent us the video below in which he recorded his experience in a self-driving vehicle on a recent visit to San Francisco and Silicon Valley. Self-driving vehicles, which are deeply reliant on AI technology (as well as analog chips), are a reality that has already arrived. |
|  | An AI-powered self-driving vehicle |
| The history of computing and internet data usage is one of uninterrupted and persistent growth, as technological advancement makes everything better, faster and cheaper. Betting against innovation in AI strikes us an incomprehensibly myopic.
It is certainly plausible, however, that the current data center investment boom, like the late 1990s fiber boom, is getting ahead of itself. The result could be a period of excess capacity. But as we saw in the prior era, the excess capacity was harmful to the providers of that capacity but helpful to the overall development of the industry, which did not stop. The dotcom era was a financial disaster for many, but over the next quarter century, the internet genuinely did transform the global economy and society.
Choose wisely
Understanding valuation is critical. Nobody knows with certainty how AI will develop. Valuation is our chief concern when it comes to many large-cap technology stocks. NVDA’s profit margins could be squeezed by price competition, while the Mag 7 hyperscalers who are investing aggressively in AI at the moment may not achieve the financial results in the next few years that the market is counting on.
If AI growth disappoints, these stocks will not go to zero, but their multiples could come down meaningfully. Investors will not be destroyed as LVLT shareholders were, since these companies have minimal debt levels and other profit streams. But they will not be happy.
As always, we seek attractive growth prospects and alignment with important long-term trends. But we much prefer stocks with valuations that do not require us to accept heroic assumptions about future profitability.
While staying mindful of worst case scenarios and understanding that AI will develop in fits and starts like other technological revolutions, we continue to view AI as a theme to which investors should seek exposure. The open-ended upside optionality is too great to ignore.
A technological revolution does not automatically translate into financial success, however. As Buffett himself has always emphasized, competitive advantage is a necessary ingredient. In the end, Level 3 never really had one. But it did have a whole lot of debt.
Level 3 was eventually acquired in 2017 by telecom peer CenturyLink, which has since renamed itself Lumen Technologies (LUMN). Sadly, an investor in LVLT at the end of 2001, when the shares were already down 97% from peak levels, would have lost 30% in the buyout, some 16 years later.
Walter Scott, Jr. passed away in 2021. James Crowe died in 2023. Warren Buffett is still alive. His holding company Berkshire Hathaway (BRK) recently disclosed it now has a $277 billion cash position.
Much of Berkshire’s cash was generated from selling down its enormous stake in Apple (AAPL). A competitively advantaged tech company if there ever was one, Apple became a disproportionately large position within the Berkshire portfolio thanks to extraordinary price appreciation.
Buffett may have missed his chance with Walter Scott’s future bride, but he was apparently never afflicted with fear of missing out on Level 3. While the rest of Omaha was piling in, Buffet reportedly declined to invest in his friend’s company.
When it comes to investing, especially in technology, Warren Buffett has always understood that not all boats are lifted by a rising tide. Investors need to pick their AI spots carefully. |
| |  | | | FOR SUBSCRIBER USE ONLY. DO NOT FORWARD OR SHARE. |
| | | | |  | |
|
|
|
|
|
