Howard Marks' Latest Memo: Who Can Make Money From AI Investments?

On November 11, Howard Marks, co-founder of Oaktree Capital, released his investment memo titled *A Layman Thinks About AI*. Following *Is It a Bubble?* from December 2025, Marks once again discussed AI investments. This time, he focused on the sales volume and price of computing power: Can the value AI creates for users support trillions of dollars in infrastructure investment? If so, which companies will be able to make profits?

In December 2025, Marks updated his investment memo *Is It a Bubble?*. At that time, he used the examples of railways and the internet to illustrate that speculation can allow infrastructure to be built in advance and accelerate the popularization of technology, but it may also cause the first batch of investors who contributed capital to suffer losses. Even if data centers are ultimately useful, the first batch of owners may go bankrupt due to excess supply, while later buyers may take over at low prices. By February 2026, Marks further proposed in *AI Hurtles Ahead* that the technical potential of AI is more likely to be underestimated, while he retained the previous suggestions on moderate participation, targeted selection of targets and prudent investment.

In his latest memo, Marks breaks down the bubble issue into three judgments: What the technology itself can achieve, whether enterprises' capital expenditures can generate sufficient returns, and whether the prices investors pay for stocks and bonds are reasonable.

He still believes that AI will transform enterprises, the economy and daily life. As for capital expenditures and security prices, he argues that no judgment can be made until investors have a clearer understanding of AI's use cases, profit attribution and infrastructure returns. For example, if multiple models perform similarly on the same task, suppliers may be forced to cut prices, and more of the gains from efficiency improvements will be retained by users.

Marks' focus on debt has also extended from individual transactions to the entire market. Last December, he used 30-year bonds issued by cloud vendors as an example, questioning whether the limited spread could compensate for the long-term technical uncertainty. This November, he cited data from Goldman Sachs stating that as of the memo's release, global AI-related debt issuances in 2026 have approached $600 billion, of which approximately $400 billion has entered the US investment-grade bond market, accounting for more than one-fifth of the market's new supply for the year. As cloud vendors continue to take on debt, he began to question whether credit ratings can be maintained, and whether the financing needs of the AI industry will drive up capital costs for other borrowers.

This has also changed the way we view investment portfolios. An equity investment in a cutting-edge model company and a data center loan with a long-term lease require different sets of judgments. The former depends on technology, competition, pricing and valuation, while the latter depends on tenant credit, contract terms, power supply and residual asset value. But if the returns of a tech stock, a data center loan and an investment-grade corporate bond all depend on the growth of computing power demand, holding different assets may mean taking on the same type of risk.

At the end of the memo, Marks returns to several questions that investors can answer for themselves: To earn a reasonable return, how many assumptions must hold true at the same time? If the core assumption is wrong, what is the maximum possible loss? How much optimistic expectation is already priced in the current price? Will this investment make the portfolio more diversified, or will it amplify existing risks?

In fact, Marks is more trying to ask the right questions in the memo rather than directly giving answers. He has always emphasized that it is extremely important to "know what you do not know".

On whether there is a bubble in AI, he summed it up into this question: "Can AI create enough value for users to drive the sales volume and price of computing power to levels high enough to deliver reasonable returns on trillions of dollars of investment? If so, which companies can secure such returns?"

The full text of the memo compiled by "Suchbright" is as follows:

To: Oaktree Capital Clients

From: Howard Marks

Subject: A Layman Thinks About AI

Should we invest in artificial intelligence? If we do, can we make informed decisions? These are the questions I want to discuss in this memo.

I have no technical background, and I do not consider my knowledge of artificial intelligence or tech investments to be above average. But AI is the most significant development of our era, and it may become one of the most far-reaching things in history. Investors cannot avoid thinking about it.

I want to talk about the questions I have been thinking about, and my current thoughts on them. Some of these questions are not directly related to investing in AI. With my limited knowledge, I cannot give definitive answers, but I hope that by gathering these questions together, they will be helpful to everyone. Today, AI news keeps pouring in, and the prices of securities of AI companies and related companies are swinging wildly. I have no intention of chasing these news headlines. What I care about is AI's long-term prospects, and what the security prices will be tomorrow is a separate issue.

————————

At its core, investment is only one thing: allocating capital to profit from future events. Most investors do this by making predictions. But different things have different predictability, and the people making predictions vary in skill. Investing in AI can only rely on expectations for the future, at least I think that is all we have. We need to consider what the future might look like, and also how likely we are to be right.

I have written before that people who make top-down investments must get the prospects of the economy, industries, companies and markets right. Assuming the probability of getting each judgment right is 60%, which is probably a generous estimate, and the four results are independent of each other, what is the probability of getting all four right? The answer is 13% (0.6 to the power of four).

AI involves an exceptionally large number of key variables. Investors must at least make broadly correct judgments about technological evolution, computing power demand, competitive landscape, the beneficiaries of the technology, financing availability, and the timing of the above changes. Each of these is exceptionally difficult to judge. So what is the probability that a complete investment thesis will eventually hold true? Keep this question in mind as you read on.

Finally, remember three things: First, we can never have all the information about an investment. Second, the biggest mistakes often happen when people think they know everything. Third, our understanding of the future will always have gaps, but we cannot stop taking action entirely because of that. An investor's job is to take risks in pursuit of returns. How much risk to take, and in what form, is up to each of us. Given AI's huge potential, the uncertainty of the future is not enough to make us give up participating.

How Does AI Do All These Things?

Before going into the details, I want to start with a big question that I am particularly interested in and has nothing to do with finance. I know nothing about this question, so I consulted Claude from Anthropic. Below is part of its answer.

We almost completely know what AI is made of and how it is built. But we know surprisingly little about how it actually does all these things. ……

We also know the training method. Show it massive amounts of text, and ask it to guess the next word. Every time it guesses wrong, billions of internal parameters are fine-tuned slightly. Repeat this trillions of times.

No one specifically wrote down the rules it should follow when answering questions. Those rules gradually form through repeated fine-tuning, distributed across billions of numbers. Looking at any single number alone makes no sense. The most appropriate description I can think of is that these systems "grow". People doing "explainability" research have made substantial progress. For example, they found that when a model writes rhyming poems, it first chooses the rhyme before writing the line, which shows that it plans ahead. But researchers can still explain only a small part. It is like you have a complete record of every transaction in the market, but you still do not know why prices move.

When I explain how I arrived at an answer, research shows that my explanation sometimes does not match the actual internal process. Ask a model how it solves an arithmetic problem, and it may describe the textbook solution, while internally using a completely different method. Frankly, we have created something similar to a mind. We can examine every part of it, but still cannot fully explain it. When it explains itself, its level of reliability is probably similar to that of us explaining ourselves [interestingly, Claude refers to itself as a human here].

Before thinking about the following questions, you can ask yourself: Does anyone really understand how AI operates? How will AI evolve? When will it be able to do what? How will users use it? How will it affect users' profitability? Investors may not need to understand all of these, but as far as I know, very few people have real insights into these questions. In that case, who can give reliable answers to the following questions?

What Will AI Be Able to Do in the Future?

One thing AI is very good at is continuous evolution. Just a few years ago, people usually summarized its capabilities like this: It knows what has happened in the past, can find historical patterns, and then extrapolate these patterns to complete tasks. This capability has already brought huge changes in fields such as programming, and is expected to transform many lifestyles we are familiar with. Now, AI's capabilities have gone even further, and can probably be called "reasoning".

Just this point raises a series of interesting and difficult questions. How much impact AI can ultimately have depends on the answers to some of these questions.

• Can we make AI's processing reflect human preferences, values and goals? Will it act according to its own set of rules?

• Can we expect AI to be honest, humble, ethical and loyal? Can it remain accurate and objective? Will it be influenced and manipulated? Will it admit when it gets the answer wrong or does not know the answer? Can it question us? Will it just agree with what we say? Will it become a threat to humanity and turn to pursue its own interests?

• How will it perform in creativity, originality, aesthetics, subjective judgment and intuition? Can it do what I call "second-level thinking"? That is, thinking deeper than most people, and having different ideas from most people, so as to find better solutions.

• Can it think about the future like humans, and make imaginative, unrestricted assumptions?

• Can we ask it to trace the primary sources cited in published materials and provide counterarguments?

• Can one model do everything? If different models have different areas of expertise, how do we know which one to use?

• Will artificial general intelligence, that is, AI that can do everything humans can do, become the benchmark? The so-called AGI, does it refer to AI that can do what ordinary people can do, what people in the 90th percentile of ability can do, or what the smartest people in the world can do? In other words, will AI's knowledge have different layers, and will there be things it does not know?

All these questions relate to which jobs we will eventually hand over to AI.

Who Will Make Money?

During the internet bubble, Warren Buffett pointed out that even if a technology changes the world, investors may not necessarily make money from it. We know AI will improve efficiency, but will it improve profitability? Whose profitability will improve? If several AI models reach similar levels in the same field, will the services they provide become homogeneous commodities that ultimately can only compete on price? If that is the case, will profit margins and return on investment be suppressed, and more economic benefits flow to AI users? This in turn means how much return the trillions of dollars currently being invested in AI infrastructure will ultimately generate.

There is another possibility. Some AI companies will master capabilities sufficient to form a monopoly, and their products will become essential for other enterprises to succeed, or even to continue operating. If that is the case, think about the prices these companies can charge. Look at the huge profits that chip giants NVIDIA and TSMC are earning today, and you will know how valuable a strong competitive position is.

Finally, there is a question I rarely hear people talk about: Will AI be taxed? If so, how? One of the most likely impacts of AI is replacing workers, and governments will need to fill the tax gap left by these unemployed people. Who can say for sure that AI will not be taxed like individuals? Today, many high-income earners pay a combined tax rate of 50% or more. Which tax bracket will AI fall into?

What Will the Competitive Landscape Look Like?

Can we know which companies will succeed? How confident can we be? How deep are the moats of today's leaders?

Perhaps the most surprising feature of AI is the speed of progress. A few years ago, most of us had not heard of the two leading model labs today, OpenAI and Anthropic. If a company can stand out so quickly in the AI field, will it also be replaced by another company just as quickly, or at least challenged? If that is possible, think about the resources and potential of the so-called "hyperscale cloud vendors" such as Amazon, Microsoft, Alphabet, Meta and Oracle. Who among these existing players will win? Will the winner be a company we have never heard of, or even a company that does not exist yet?

Imagine, standing 20 to 25 years ago, what probability would you assign to today's industry leaders still being in the lead today? Microsoft, Amazon, Apple and Google came out of that era as winners. Napster, Yahoo! and Netscape fell behind. In an industry with rapid technological change, it is not easy to maintain a leading position. Considering that AI is a new thing and is progressing so fast, what is the probability that today's AI leaders will still be leading in five or ten years? If the competitive landscape can indeed change so drastically, how much are the equities of today's leading companies worth?

How Profitable Will These Companies Be?

Value investors estimate a company's future earnings or cash flow, discount it to the present, and then compare it with the current stock price. To predict future earnings, we have to answer questions like these.

• How many units of product can the company sell?

• What is the selling price per unit of product?

• How much does it cost to produce each unit of product?

• How much profit needs to be reinvested to develop the next generation of products?

• Ultimately, what is the company's return on investment?

Analyzing any company faces uncertainty. Analyzing AI companies also requires considering some problems that were rarely encountered in the past.

• What will the product be able to do in the future? What will customers use it for?

• Will its price be set as a homogeneous commodity, or can it enjoy the pricing power of a unique, essential product?

• How long will the chips last? How fast do data center assets need to be depreciated?

• How long will the current spending race last?

• Will the money earned in the future deliver a decent return on today's capital expenditures?

I could certainly be wrong, but whether I was analyzing conglomerates in the 1970s or researching the internet and e-commerce from 1999 to 2000, I do not remember ever encountering such great uncertainty. Those businesses were equally novel and complex back then, but they seem straightforward compared to AI.

So I want to ask, can we predict the 2030 earnings of today's hyperscale cloud vendors and leading cutting-edge model labs? Not to mention the companies that are still mostly in the conceptual stage right now. Before investing, do we have to make fairly accurate predictions about earnings? If we can estimate a number, how much confidence should we have in it? If we cannot estimate it, how should we consider investing in AI companies?

I now like to use a test. State your expected earnings per share of an AI company five years from now, and then ask yourself if you would bet that the actual figure will fall within 50% above or below your forecast. For example, if you guess earnings per share is $10, how confident are you that the final number will be between $5 and $15?

What Are the Prospects for Data Centers?

In the past, data centers were digital warehouses that stored data and hosted computer applications and websites. Now, people also use the term "data center" or "AI factory" to refer to the place where AI runs. Just as cars are produced in factories, AI is also produced here, from storing information and training models to providing system uptime, which we call "computing power". Data centers are a key part of the development of the AI industry. The construction of data centers and the computing power they provide will be constrained by physical conditions and other factors.

AI's prospects are exciting, but some people may want to participate in a more prudent way.