Howard Marks: AI is the least predictable phenomenon for investors. What is his advice?

Howard Marks, a renowned value investor and co-founder of Oaktree Capital Management, believes that assessing the investment prospects of AI is difficult due to numerous uncertainties, but he does not discourage investment in the sector. In his new memorandum, he outlined how to make investment decisions regarding AI.

In the document, Marks poses several dozen questions for AI investors. Some of them are quite practical, while others are rather philosophical.

For example, he begins with a fairly general question: “How exactly does artificial intelligence work?” and turns to Claude from Anthropic for an answer (the model agrees that very few people actually understand how AI does what it does).

Marks himself suggests we consider whether anyone truly understands how AI will develop and what it will be capable of, how it will be used in the future, and how it will affect companies' profitability.

“Perhaps investors don’t need to know the answers to all these questions. But, as far as I understand, very few people have a truly deep understanding of these processes. If that’s the case, can anyone at all reliably answer the questions that follow?” Marks asks.

According to the investor's definition, a bubble occurs when excessive optimism leads to ill-considered decisions, excessive investment, and inflated asset prices.

And right now, Marks doesn’t have a clear answer to the question of whether a bubble has formed around AI: Until there is sufficient knowledge to make informed predictions about corporate profits and the returns on infrastructure projects, it is impossible to judge whether stock prices and the proposed bond yields are justified. The main question is whether AI will be able to create enough value for users to ensure an acceptable return on investments worth trillions.

When it comes to threats to the economy and the markets, he is more confident. He recalls the two main principles of Michael Milken, who was at the forefront of the high-yield bond market: an investor can take on increased risk if they receive sufficient compensation for it and diversify their investments. However, according to Marks, the current concentration of the stock and bond markets in the AI sector makes them more vulnerable to potential problems in the industry.

AI’s main contribution to economic growth comes from large-scale investments, primarily in the construction of data centers. These expenditures are rising rapidly, while other segments of the construction industry are experiencing a decline. In addition, equipment suppliers are reaping additional profits, and spending by AI company employees is stimulating demand for goods and services.

In the stock market, the reliance on AI is even more pronounced. According to JPMorgan, just 42 technology-related companies have accounted for about two-thirds of the S&P 500’s growth since January 2024. For 14 consecutive quarters through June 30, 2026, the “Magnificent Seven” posted faster earnings growth than the rest of the index. Part of the tech giants’ profits stems from the rising value of their stakes in other companies. According to FactSet, in the second quarter, Alphabet and Amazon reported a combined $151 billion in gains from such investments, primarily in Anthropic and SpaceX.

Just a few years ago, the main advantage of AI lay in analyzing accumulated knowledge, identifying patterns, and performing tasks based on historical data, but today it is already capable of reasoning. Its future potential depends on whether it can take human values and goals into account, remain honest, objective, and ethical, and resist manipulation, argues Howard Marks.

In his view, the question of whether a single universal model will emerge or whether different systems will specialize in specific tasks also remains open. And finally, will there be different levels of AI intelligence and areas in which it lacks knowledge?

In discussing the second question, Marks quotes Warren Buffett, who said during the dot-com boom that revolutionary technologies do not necessarily reward investors. Marks believes this remains relevant today: AI can improve business efficiency, but it is unclear whether its use will lead to profit growth—or who will reap those profits.

If competition among AI models leads to lower prices and reduced profitability for developers, users will benefit—and this will impact the return on the trillions invested in AI infrastructure.

But if certain companies in the AI sector gain a monopoly, they will be the ones able to dictate prices and reap high profits, just as Nvidia and TSMC do today, Marks notes.

Another issue of concern to investors is the future taxation of AI. Today, for people with the highest incomes, the total tax burden can exceed 50%. Howard Marks wonders whether AI will pay taxes and which tax category it will fall under.

Value investors evaluate a company based on its future profits. To forecast these profits, it is necessary to estimate the company’s sales volume, prices, production costs, and the portion of profits that will need to be invested in new product development.

However, Marks writes, in the case of AI, new questions arise in addition to these: what capabilities will AI products have, how will customers use them, and will these products be sold as ordinary goods or as unique solutions? Furthermore, it is important to understand whether today’s capital expenditures will yield an acceptable return on investment, and how to account for the useful life of chips and the depreciation of data centers.

Marks notes that he cannot recall such widespread uncertainty even during the dot-comboom of the late 1990s or in the 1970s, when, as an analyst, he evaluated corporations that were acquiring companies from various, often unrelated, industries.

The current uncertainty stems in part from the fact that it is unclear how accurately investors can forecast the profits of cloud companies—such as hyperscalers and leading AI developers—for example, in 2030, and whether such estimates can be relied upon when making investment decisions.

Marks proposes a test for investors: predict an AI company’s earnings per share five years from now and assess whether they are willing to bet that the actual result will differ from the forecast by no more than 50%. For example, if the forecast is $10 per share, the earnings should fall within the range of $5 to $15.

As for competition, he notes that just a few years ago, almost no one had heard of OpenAI and Anthropic, yet today they are market leaders. But since these companies have surged ahead so quickly, Marks writes, doesn’t that mean they could be displaced just as quickly?

He points out that the previous technological revolution has already shown how difficult it is to maintain a leadership position: Microsoft and Amazon have held onto their positions, but Napster, Yahoo, and Netscape have fallen by the wayside.

This is perhaps one of the most pressing issues in the market right now. Marks believes that investing in data centers is a less risky way to profit from AI, although he notes that even here there is no guarantee of success for such investments.

However, it is important to note that when building data centers, the building itself accounts for only a small portion of the project’s cost: the main expenses are for chips, and the costs associated with them—as well as the risk of them becoming obsolete—are typically borne by the data center users. In addition, project implementation depends on the availability of electricity, cooling systems, land, and fiber-optic connectivity. Regulatory restrictions and resistance from local residents can hinder project development. These are already significant unknowns. Furthermore, the cost of raising capital for future projects may turn out to be unexpectedly high.

All of this complicates matters for companies launching data centers, but on the other hand, any slowdown in the growth of computing capacity supply could increase the profitability of such projects and the value of existing facilities.

Marks calls AI the most significant and least predictable phenomenon for investors in at least the last generation. The correct response to this, he writes, is “measured humility” when choosing a strategy, determining position sizes, and taking on risk, and secondly, a willingness to act despite the uncertainty.

For example, an investor purchasing shares in a leading AI lab must accurately assess the pace of technological progress, the company’s competitive position, its ability to influence prices, future profitability, and the fair value of the business. A lender financing the construction of a data center under a long-term contract must consider the tenant’s creditworthiness, the terms of the agreement, the availability of electricity, and the residual value of the facility.

For convenience, Marks has compiled a list of questions that a potential investor should try to answer when deciding whether to invest in AI.

***

In the “Guru Portfolios” section on Oninvest, you can track the composition and changes in the portfolios of the world’s largest investors and funds. The service allows you to analyze the largest holdings, new ideas, and changes in asset allocations based on 13F filings, as well as compare portfolio performance over time.

This article was AI-translated and verified by a human editor