Goldman Sachs research director condemns "AI bubble": This bubble may be more severe than the 2000 Internet bubble

Nasdaq has repeatedly hit new highs, as star tech stocks take turns to refresh records. At this moment, is there a huge crisis hidden in the AI bubble, and when will it collapse? Allison Nathan, a senior strategist at Goldman Sachs Global Macro Research, raised a crucial question in the latest "Top of Mind" report: Is there too much investment in AI with too little return? Although Goldman Sachs' report did not provide a clear answer, it shared a few expert interviews and explicitly presented a pessimistic view. It stated that tech giants plan to spend $1 trillion on AI capital expenditures in the coming years, but there is almost no substantial, visible evidence to prove that these investments are worthwhile. In the "Top of Mind" report, author Allison Nathan elaborated on the current trends in AI technology: Generative AI technology is believed to be able to change companies, industries, and society, so many large companies plan to invest $1 trillion in AI-related things in the coming years, such as data centers, chips, and grids. However, so far, apart from slightly improving the efficiency of developers, this money has not seen any other significant results. Even the stock price of NVIDIA, which benefits the most from it, has fallen. We asked some industry and economic experts to see if these huge expenditures will bring benefits and returns for AI, and discussed the impact on the economy, companies, and markets whether they bring returns or not. Furthermore, Jim Covello, Goldman Sachs' head of stock research, is very pessimistic about the current AI bubble, believing that this bubble may be even more serious than the dot-com bubble at the end of the last century. He pointed out: The cost of developing and operating AI technology is very high, estimated at around $1 trillion. To make this investment worthwhile, AI needs to solve very complex problems, but AI cannot do that yet. Disruptive technologies like the Internet could replace high-cost solutions with low-cost solutions even in the early stages. AI is expensive now and cannot provide cheaper alternatives. Moreover, Covello doubts whether the cost of AI can be reduced enough to make large-scale automation cheap, as the initial cost of AI is high, and key components (such as GPU chips) are complex to produce. This complexity may also limit competition in the AI field. He believes that AI is unlikely to significantly increase company valuations because the efficiency gains brought by AI are likely to be quickly caught up by competitors, and how AI actually brings revenue growth is also unclear. Finally, Covello questions whether AI can truly replicate the most valuable abilities of humans, as AI is trained based on historical data. He believes that AI will not reach the level of humans in these areas. Among the many experts interviewed by Goldman Sachs author Nathan, the most notable is Daron Acemoglu, a professor at MIT, who also holds a skeptical attitude towards AI. Acemoglu estimates that in the next decade or even longer, the driving force of generative AI technology on productivity and economic growth in the United States may be less than many people expect He believes that only about a quarter of tasks can be automated by AI, which means that AI will only affect 5% of all tasks. Although technology will become more advanced over time and costs will decrease, Acemoglu believes that the progress of AI models will not be as fast or impressive as many people imagine. Furthermore, he questions whether AI will create new job tasks and products. He believes that these impacts are not "natural laws," and one cannot expect AI technology to automatically bring about a large number of new jobs and products. Therefore, he predicts that the actual impact of AI on the economy in the next decade will be very limited, with AI only increasing U.S. productivity by 0.5% and GDP by only 0.9%. This could result in hundreds of billions of dollars in investment being wasted, and the tens of trillions of market value obtained by the "Big Seven" in the U.S. stock market may be the largest bubble in history. Information shows that the interviewee, Daron Acemoglu, a professor at the Massachusetts Institute of Technology, has several works, including "Why Nations Fail: The Origins of Power, Prosperity, and Poverty" and his latest work "Power and Progress: The Thousand-Year Struggle for Technology and Prosperity." The following excerpt from the interview with Acemoglu and Covello by author Nathan is said to help you realize the full extent of the AI bubble earlier than others, potentially avoiding huge investment losses in the future: Allison Nathan: Goldman Sachs economists predict that AI will increase productivity by about 9% in the next 10 years, and GDP will grow by 6.1%. However, you predict that AI will only increase U.S. productivity by about 0.5% in the next 10 years, and GDP will only increase by about 1%, which may be much less optimistic than many forecasters (including Goldman Sachs). Why are you less optimistic about the potential economic impact of AI? Daron Acemoglu: The difference in predictions seems to revolve more around the time AI will impact the economy rather than the ultimate prospects of the technology. Generative AI has the potential to fundamentally change some areas, but these changes will not occur in the next 10 years. Current generative AI is mainly improving existing processes by automating certain tasks or enhancing worker efficiency, rather than creating new, large-scale transformations. In the short term, the number of tasks that AI can automate is limited. Many tasks that require real-world interaction, such as transportation, manufacturing, mining, etc., cannot be significantly improved by AI in the short term. The main impact of AI will be on pure cognitive tasks, but the number and scale of these tasks are not large. To quantify this, I first studied a comprehensive study by Eloundou et al., who found that generative AI and other AI technologies can change over 20% of production tasks, but this is a long-term forecast. Another study estimates that only about 23% of tasks can be efficiently automated by AI in the next 10 years, which means that only about 4.6% of tasks will be affected by AI. The average labor cost savings are about 27% In the next 10 years, AI is expected to increase productivity by about 0.53% to 0.66%, leading to a GDP growth of around 0.9%. Allison Nathan: Recent studies estimate that using AI can save 10% to 60% of costs, but you believe it can only save around 30%. Why is that? Daron Acemoglu: There are three detailed studies on the cost-saving effects of AI. One of them (Peng et al.) estimated savings of up to 56%, but it focused on simple tasks like using AI to help programmers write HTML. These tasks are easy for AI to accomplish, but more complex tasks are not as straightforward. Therefore, I ignored this study and only considered the other two more realistic estimates. Allison Nathan: Historically, technological advancements have often improved and reduced costs. Will artificial intelligence technology follow a similar trend? Daron Acemoglu: It is certainly possible. However, I do not believe that simply increasing data and computing power can rapidly enhance AI capabilities. Many people think that more data and computation will make AI better, but what does doubling AI capability specifically mean? For example, in customer service or complex text summarization, there is no clear indicator that AI output will be twice as good. Additionally, the quality of data is crucial, and it is currently unclear where to obtain more high-quality data. Finally, current AI technology itself may have limitations. Human cognition involves multiple complex processes, and current AI is far from reaching the level of intelligence seen in HAL 9000 from "2001: A Space Odyssey." Allison Nathan: Even if you are conservative about the impact of AI in the next 5 to 10 years, are there downside risks? Daron Acemoglu: There are indeed risks. Technological breakthroughs are always possible, but even with breakthroughs, it takes time to see results. If AI performs poorly in improving complex tasks, then even my conservative estimate may be too high. Large companies may quickly adopt AI tools, but the speed of adoption by small companies may be slower. Allison Nathan: Looking ahead, how likely do you think it is for AI to achieve superintelligence? Daron Acemoglu: I doubt whether AI can achieve superintelligence in the longer term. AI may completely change scientific processes in 20-30 years, but humans will still be in control. True superintelligent AI would be able to complete all tasks without human intervention, but I think even 30 years from now, this scenario is unlikely to occur. Allison Nathan: Your colleague David Autor and co-authors have indicated that technological innovation often drives the creation of new professions, with 60% of workers today engaged in occupations that did not exist 80 years ago Daron Acemoglu： Undoubtedly, technological innovation has had a significant impact on almost every aspect of our lives. However, this impact is not a natural law. It depends on the types of technologies we invent and how we use them. Therefore, I once again hope that we can use artificial intelligence technology to create new tasks, products, business professions, and capabilities. In my example of how artificial intelligence tools can fundamentally change scientific discovery, artificial intelligence models will be trained to help scientists conceive and test new materials, allowing humans to become more specialized through training and provide better input for artificial intelligence models. This evolution will ultimately bring better possibilities for human discovery. But this is by no means inevitable. Allison Nathan: Technological innovation often creates new professions. Today, 60% of workers are engaged in professions that did not exist 80 years ago. In the long run, will the impact of artificial intelligence technology be greater than you expected? Daron Acemoglu: The impact of technological innovation on life is enormous, but it does not happen automatically. It depends on the types of technologies we invent and how we use them. If we use AI technology to create new tasks, products, and professions, such as AI assisting scientists in researching and testing new materials, then the long-term impact of AI may be greater, but this requires conscious effort. Allison Nathan: With a large amount of money being invested in artificial intelligence technology today, will some or even most of it end up being wasted? Daron Acemoglu: This is an interesting question. Basic economic analysis suggests that an investment boom should occur because today's artificial intelligence technology is mainly used for automation, which means algorithms and capital are replacing human labor, which should trigger investment. This explains why my estimate of GDP growth is almost twice my estimate of productivity growth. However, reality shows that some spending will eventually be wasted because some projects will fail, some companies will be too optimistic about the extent of efficiency improvements and cost savings they can achieve or their ability to integrate artificial intelligence into their organizations. On the other hand, some spending will sow the seeds for the next more promising stage of this technology. The devil is in the details. So, I do not have strong prior knowledge of how much of the current investment boom will be wasted and beneficial. But I expect both to happen Allison Nathan: Are the other costs of artificial intelligence technology not receiving enough attention? Daron Acemoglu: Yes. GDP is not everything. Technologies that can provide good information can also provide bad information and be abused for evil purposes. I am not too concerned about deepfakes at the moment, but in terms of how bad actors can abuse generative AI, they are just the tip of the iceberg. Investing one trillion dollars in deepfakes would increase GDP by one trillion dollars, but I think most people would not be happy about it or benefit from it. Allison Nathan: From all the things we have discussed, is the current enthusiasm for artificial intelligence technology excessive? Daron Acemoglu: Every human invention is worth celebrating, and generative artificial intelligence is a true human invention. But excessive optimism and hype can lead to premature use of technology that is not yet ready. Today, the risk of using artificial intelligence to drive automation seems particularly high. Premature over-automation could create bottlenecks and other problems for businesses as they no longer have the flexibility and troubleshooting capabilities provided by human capital. Moreover, as I mentioned, using such widespread and powerful technology to provide information and visual or written feedback to humans in ways that we do not fully understand and cannot fully regulate could be very dangerous. While I do not believe superintelligence and evil artificial intelligence pose significant threats, I often think about how people will view current risks 50 years from now. In 2074, our descendants may blame us for acting too slowly in 2024, sacrificing growth, a risk that seems much lower than the risk of acting too fast and destroying institutions, democracy, and other things in the process. Therefore, the costs of the mistakes we make are more asymmetrically negative. That's why it is important to resist hype and take a cautious approach, which may include better regulatory tools as artificial intelligence technology continues to evolve. Interview with Jim Covello, Global Head of Stock Research at Goldman Sachs, who believes that to get sufficient returns from expensive artificial intelligence technology, AI must solve very complex problems, which it currently cannot do and may never be able to do. Allison Nathan: Your enthusiasm for current generative artificial intelligence is not as high as others. Why is that? Jim Covello: My main concern is that the cost of developing and operating artificial intelligence technology is high, which means AI applications must address extremely complex and important problems for businesses to achieve a proper return on investment (ROI). We estimate that in just the next few years, building AI infrastructure will cost over $1 trillion, including expenditures on data centers, utilities, and applications. So, the key question is: what $1 trillion problem will AI solve? Replacing low-wage jobs with extremely expensive technology is basically the opposite of the previous technological transformations I have closely observed in the tech industry over the past thirty years Many people try to compare today's artificial intelligence with the early days of the Internet. Even at its inception, the Internet was a low-cost technological solution that enabled e-commerce to replace expensive existing solutions. Amazon could sell books at a lower cost than Barnes & Noble because it didn't have to maintain expensive physical stores. Fast forward thirty years, Web 2.0 still provides cheaper solutions that are disrupting more expensive ones, such as Uber replacing luxury car services. While the question of whether artificial intelligence technology can deliver on the promises that excite many today is certainly controversial, one less controversial point is that artificial intelligence technology is very expensive. To justify these costs, the technology must be able to solve complex problems, which is not its design purpose. Allison Nathan: Even though artificial intelligence technology is expensive today, won't technology costs often decrease significantly as the technology develops? Jim Covello: Technology is usually expensive at the beginning and then becomes cheaper, a view that is a correction of history. As we just discussed, e-commerce was cheaper from day one, not ten years later. But even setting aside this misconception, the tech industry is too complacent in thinking that the cost of artificial intelligence will decrease significantly over time. Moore's Law drove the development of smaller, faster, and cheaper chips, propelling the history of technological innovation, but it turns out that this law is correct because Intel's competitors (such as AMD) forced Intel and other companies to lower costs and innovate continuously to stay competitive. Today, NVIDIA is the only company capable of producing the GPUs needed for AI. Some believe that the semiconductor industry or mega-corporations (Google, Amazon, and Microsoft) themselves will become competitors to NVIDIA, which is possible. But compared to the current situation, this would be a huge leap because for the past 10 years, chip companies have been trying to overthrow NVIDIA's dominance in the GPU field but have all failed. Technology is difficult to replicate, to the point where no competitor can do so, allowing companies to maintain a monopoly and pricing power. For example, ASML, the advanced semiconductor material lithography technology company, is still the only company in the world capable of producing cutting-edge lithography tools, so their machine costs have increased from tens of millions of dollars two decades ago to billions of dollars in some cases today. NVIDIA may not follow this exact pattern, and the scale is different in dollars, but the market is too complacent about the certainty of cost reduction. The starting point of costs is also high, and even with cost reductions, they must decrease significantly to make AI automation tasks affordable. People point out that since the late 1990s when servers were introduced, server costs have dropped significantly in a few years, but the number of expensive chips needed to drive the transformation of artificial intelligence today is insignificant compared to the number of $64,000 Sun Microsystems servers needed to drive the transformation of Internet technology in the late 1990s, even without including the costs of replacing the power grid and other costs needed to support this transformation, these costs themselves are very expensive Allison Nathan: Are you just concerned about the cost of artificial intelligence technology, or do you also have doubts about its ultimate transformative potential? Jim Covello: I am skeptical about both. Many people seem to think that artificial intelligence will be the most important technological invention in their lifetime, but I disagree with this view because the internet, mobile phones, and laptops have fundamentally changed our daily lives, enabling us to do things that were previously impossible, such as making phone calls, calculations, and shopping anytime, anywhere. Currently, artificial intelligence shows the greatest potential in improving existing processes (such as coding) efficiency, although the estimated values of these efficiency improvements have decreased, and the cost of using this technology to solve tasks is much higher than existing methods. For example, we found that artificial intelligence can update historical data in company models faster than manual updates, but the cost is six times higher than manual updates. More broadly, people often greatly overestimate the capabilities of today's technology. Based on our experience, even basic summarization tasks often produce unrecognizable and meaningless results. It is not just a matter of making adjustments here and there; despite the high cost, this technology is far from reaching the level required to complete these basic tasks. I find it hard to believe that this technology can achieve significant enhancement or replace the cognitive reasoning required for human-machine interaction. Humans add the most value to complex tasks by identifying and understanding outliers and subtle differences, and it is difficult to imagine a model trained on historical data being able to do this. Allison Nathan: But isn't it difficult to predict the transformative potential of these technologies in the early stages? So why are you so sure that artificial intelligence will not eventually prove to have the same or even greater transformative power? Jim Covello: It is a mistake to think that the transformative potential of the internet and smartphones was not understood in the early stages. When smartphones were first introduced, I was a semiconductor analyst, and in the early 21st century, I attended hundreds of presentations on the future of smartphones and their features, most of which met industry expectations. An example is integrating GPS into smartphones, although it was not yet ready to take over the bulky GPS systems commonly found in rental cars at the time. There were roadmaps for what other technologies could eventually do when they were first introduced. There is no similar roadmap today. Supporters of artificial intelligence seem to just believe that use cases will skyrocket as the technology develops. But 18 months after the advent of generative artificial intelligence, no truly transformative (let alone cost-effective) applications have been found. Allison Nathan: Even if returns and rewards can never offset costs, considering competitive pressures, do companies have other choices besides pursuing an artificial intelligence strategy? Jim Covello: Given the hype and fear of missing out in this field, large tech companies have no choice but to participate in the AI arms race, so the massive investment in AI development will continue This is not the first time that a hype cycle in technology has led to investments in technologies that ultimately did not succeed; virtual reality, metaverse, and blockchain are typical examples of technologies that have received significant funding but are currently rarely (if at all) applied in the real world. Companies outside the tech industry are also facing immense pressure from investors to implement AI strategies, even though these strategies have not yet borne fruit. Some investors have accepted that these strategies may take time to yield returns, but others do not agree. For example, Salesforce invested heavily in AI, but recently its stock price experienced its largest single-day drop since the mid-2000s, as its second-quarter performance showed that despite the investments, revenue hardly grew. Allison Nathan: How likely do you think it is that AI technology will eventually increase the revenue of non-tech companies? Even if there is no revenue growth, can cost savings still pave the way for diversification expansion? Jim Covello: I think the likelihood of revenue growth related to AI is very low because I believe this technology is not yet smart enough, nor is it likely to make employees smarter. Even one of the most reasonable use cases for AI, improving search functions, is more likely to help employees find information faster rather than find better information. If the benefits of AI still mainly revolve around improving efficiency, this may not lead to diversification expansion, as cost savings will be arbitraged away. If a company can use robots to improve efficiency, then its competitors can too. Therefore, a company will not be able to charge higher fees or increase profits. Allison Nathan: What does all this mean for AI investors in the short term, especially considering that the "pick and shovel" companies that are most easily affected by AI infrastructure development have made progress so far? Jim Covello: Despite my skepticism, significant spending on AI infrastructure will continue, and investors should continue to invest in the beneficiaries of this spending, ranked in order: NVIDIA, utility companies and other companies expanding their grids to support AI technology, and mega-cap enterprises, which are also investing heavily but will also gain incremental revenue from AI construction. These companies have indeed risen significantly, but history shows that if a company's expensive fundamentals remain unchanged initially, relying solely on expensive valuations cannot prevent its stock price from rising further. I have never seen a stock fall simply because it is expensive—deteriorating fundamentals almost always play the leading role, only then will valuation come into play. Allison Nathan: If your skepticism is ultimately proven correct, the fundamental story of AI will collapse. What would that look like? Jim Covello: Overdeveloping things that are useless or not yet ready often leads to bad results. The Nasdaq index fell by about 70% between the peak of the Internet bubble and the founding of Uber. The bursting of today's artificial intelligence bubble may not be as problematic as the bursting of the Internet bubble, because many companies spending money today are more well-capitalized than those at the time. However, if the use cases and adoption rates of artificial intelligence technology ultimately fall below current widespread expectations, it is hard to imagine that this will not be a problem for many companies spending money on this technology today. Nevertheless, one of the most important lessons I have learned over the past thirty years is that bubbles may take a long time to burst. That's why I recommend continuing to invest in artificial intelligence infrastructure providers. If my skepticism is proven wrong, these companies will continue to benefit. But even if I am right, at least they have already generated substantial revenue from this theme, which may enable them to better adapt and grow. Allison Nathan: So, what signs should investors look for to judge when the bubble is about to burst? Jim Covello: How long investors can still be satisfied with the slogan "if you build it, they will come" remains an open question. The longer the time without significant artificial intelligence applications, the more challenging the artificial intelligence story becomes. I guess if important use cases do not start becoming more apparent in the next 12-18 months, investor enthusiasm may start to wane. But the more important area of focus is corporate profitability. Sustained corporate profitability will allow for continued experimentation with projects that have negative returns on investment. As long as corporate profits remain strong, these experiments will continue. Therefore, I expect companies not to cut back on spending for artificial intelligence infrastructure and strategies until we enter a more challenging phase of the economic cycle, which we do not expect in the short term. That being said, if corporate profitability starts to decline, spending on these experiments is likely to be one of the first projects to be cut