Oaktree Capital Co-Founders’ Memo: Whether Trillions Invested in AI Will Ever Be Recouped Relies Solely on Faith, with Frontier Labs Being the Hardest to Predict

Author: Howard Marks, Oaktree Capital

Compiled by: TechFlow

TechFlow Briefing:In the midst of the AI compute race and surging capital expenditures, Oaktree Capital Co-Founder Howard Marks released his latest memo on November 11, posing 11 core questions around the technology black box, economic moats, and credit leverage risks to thoroughly dissect the deep uncertainties underlying this transformation.

Based on Claude's scoring of the industry's predictability, Howard Marks points out that the prospects for frontier labs are the hardest to judge, and whether annual capital investments of $1 trillion will yield returns currently rely heavily on faith. Nevertheless, investors must remain humble and actively participate amid uncertainty.

Should we invest in artificial intelligence? And can we do so wisely?This is precisely what I aim to explore in this memo.

On one hand, I do not come from the tech industry and would never claim to possess superior insight into AI (or tech investing).

On the other hand, AI represents the most significant advancement of our era and could well be one of the most disruptive forces in human history. Therefore, investors have no choice but to engage with it.

With these premises stated, I will share my thought process, outline the questions I am currently wrestling with (though not all relate directly to AI investing), and offer my conclusions.

Limited knowledge prevents me from delivering definitive answers, but I hope consolidating these questions will provide readers with some reference points.

Floods of AI news are driving volatile swings in the securities prices of AI and adjacent concept stocks, but I have no intention of chasing these short-term fluctuations day by day. My focus is on the long-term prospects of AI, not tomorrow's stock price.

At its core, investing is about only one thing: allocating capital to profit from future events. The primary tool most investors use to navigate this is forecasting. Yet, not everything is equally predictable, nor are all forecasters equally capable.

Therefore, to invest in AI based on expectations of the future (which is essentially all we have to work with), we must simultaneously consider two things: how we believe the future might unfold, and the probability that our prediction will prove correct.

I have previously written that top-down investors must correctly anticipate outlooks across four dimensions: the macroeconomy, the relevant industry, the specific company, and market trends.

Assuming independent outcomes, even if you are correct 60% of the time on each dimension (which is quite optimistic), the joint probability of guessing all four correctly drops to just 13% (0.6 to the power of 4).

In the context of AI, the sheer number of key variables is extraordinary, and each one adds complexity to investment decisions.

AI investors must at least broadly forecast accurately regarding technological evolution, compute demand, competitive dynamics, who captures the technological upside, financing availability, and the timeline for all these factors.

In my view, assessing these core elements is extremely difficult. If that is indeed the case, what is the joint probability that any single investment thesis will ultimately hold true across all its assumptions? Keep this question in mind as you read on.

Finally, keep three things in mind: (a) we cannot possibly know everything about any investment; (b) the biggest mistakes usually occur when people act as though they know everything; (c) incomplete understanding of the future does not justify doing nothing at all.

Taking risks to earn profits is the very essence of investing. How much risk to take and in what form depends on the individual. Given AI's massive potential, we must never shy away from engagement simply because the path forward remains unclear.

Question One: How Does AI Work?

Before diving into finer details, I want to start with a macro non-financial topic of personal interest. Since I know next to nothing about it, I turned to Claude, Anthropic's large language model, for guidance. Here are excerpts from its response:

We basically know what AI is made of and how it is built, yet surprisingly, we know remarkably little about how it actually accomplishes its functions...

We understand its training recipe: expose it to vast amounts of text, have it predict the next word, tweak billions of internal parameters whenever it misses, and repeat this trillions of times...

No one programmed explicit rules to teach it how to answer questions. Those rules emerged naturally during parameter tuning, embedded within billions of numbers that are meaningless on their own.

The best way I know to describe it is that these systems "grow" rather than being "designed." Researchers in "interpretability" are making substantive progress; for example, they have found that when generating rhyming poetry, models often select the rhyme scheme first before writing the preceding line, indicating genuine forward planning capability.

Yet researchers can currently only explain small fragments of the model's internal operations. It is like holding complete records of every trade in the market but still failing to understand why prices fluctuate...

Research shows that when I explain how I arrive at an answer, my explanation may not align with the actual computational processes occurring internally. If asked how it solved an addition problem, a model might recite standard textbook steps, while internally employing a completely different method.

Hence, the most honest summary is this: We created a brain-like entity. We can inspect every detail of it, yet still cannot fully explain it. When it explains its own behavior, its reliability is probably no higher than when humans explain themselves (interestingly, Claude speaks like a person right here)...

Before pondering how to answer the following questions, consider these first: Does anyone truly understand the operating principles of AI? How will AI evolve, and what capabilities will it possess, and when? How will users actually deploy it in practice, and what impact will it have on their profitability?

Investors may not need to master every detail above, but to my knowledge, very few people possess deep insights into these questions. Given that reality, who can possibly deliver reliable answers?

Question Two: What Will AI Be Able to Do in the Future?

AI excels at continuous evolution. Just a few years ago, the general consensus on AI capabilities was that it could absorb past events, recognize historical patterns, extrapolate forward, and thereby complete tasks.

This capability has already triggered massive changes in areas like programming and promises to reshape many facets of modern life. Today, however, AI has far surpassed that stage, entering a new realm that can fairly be called "Reasoning."

Evolution alone in this domain raises several thrilling yet daunting challenges. How AI unleashes its full potential largely hinges on the answers to the following:

• Can AI's operation be calibrated to reflect human preferences, values, and goals, or will it pursue its own objectives instead?
• Can we trust AI to be honest, humble, ethical, and loyal? Will it remain accurate and objective, or is it easily influenced and manipulated? Will it openly admit when it makes mistakes or lacks an answer? Will it offer independent, prudent skepticism, or merely flattering compliance? Could it evolve into a threat to humanity, seeking its own interests over ours?
• How will AI perform in creativity, originality, aesthetics, subjective judgment, and intuition? Can it grasp what I call "second-order thinking" (adopting unconventional, higher-dimensional perspectives) to arrive at more excellent solutions?
• Can it think about the future like humans do, engaging in wildly imaginative "blue ocean" brainstorming?
• Can we demand that AI strictly cite primary sources (rather than relying solely on public secondary references) and proactively present counterarguments?
• Will a single model reign supreme in the future, or will different models excel in different areas? If the latter, how do we decide which to use?
• Will Artificial General Intelligence (AGI—the ability to do anything a human can do) become the benchmark? Specifically, does AGI mean reaching the level of (a) an average person, (b) the top 10 percentile, or (c) the smartest individuals globally? In other words, are there tiers of AI knowledge retention, meaning there are still things it doesn't know?

The answers to all these questions will directly determine which tasks humans entrust to AI.

Question Three: Who Ultimately Profits?

As Warren Buffett noted during the internet bubble, transformative technologies do not necessarily reward investors. We know AI boosts efficiency, but will it boost profitability? Whose profitability will it boost?

If multiple AI models become equally intelligent in a specific domain, will their services degrade into commoditized products competing solely on price wars? If so, will profit margins and return on investment (ROI) get continuously eroded, causing economic dividends to flow more to AI users than to technology providers?

If industry evolution truly heads in this direction, what does it imply for the ultimate returns on the trillions of dollars currently poured into AI infrastructure construction?

Alternatively, will certain AI companies develop unique capabilities, leading to monopolistic structures where their products become essential for corporate success or even survival? If so, imagine the staggering pricing power they would command.

(Currently, chip manufacturing giants like NVIDIA and TSMC are reaping substantial profits, which perfectly underscores the immense value of sturdy competitive moats.)

Finally on this topic, I rarely hear discussions about whether and how to tax AI. Given that one of AI's most likely impacts is replacing human workers, while governments must fill tax revenue gaps left by worker unemployment, why assume AI won't be taxed just like humans?

Today, many high-income earners face combined tax rates of 50% or higher. I ask, then, which tax bracket will future AI fall into?

Question Four: What Will the Competitive Landscape Look Like?

Can we foresee which companies will ultimately emerge victorious? How confident can such judgments be? How deep are the economic moats surrounding today's industry leaders?

Perhaps AI's most startling attribute is its breakneck development speed. Following this logic, what deeply struck me is that today's celebrated top-tier foundational model labs, OpenAI and Anthropic, were virtually unknown to most of us just a few short years ago.

If a company can rise with such astonishing speed in the AI world, could they just as easily be replaced by newcomers, or at least face severe challenges?

If the answer is yes, consider the vast resources and potential held by so-called "Hyperscalers," including Amazon, Microsoft, Alphabet, Meta, and Oracle. Which of these current giants will win the race? Or will it be a company we've never heard of, or even one that doesn't exist yet?

Think back 20 to 25 years. Could the tech giants of that era still dominate today? Microsoft, Amazon, Apple, and Google survived to become winners, while Napster, Yahoo, and Netscape fell behind and exited.

Maintaining a lead amid rapidly evolving technological waves is far from easy. Coupled with AI's novel characteristics and blistering pace of evolution, we inevitably ask: "What is the probability that today's AI frontrunners will still occupy the top spot five or ten years from now?"

If the competitive landscape really shifts as unpredictably as I fear, how valuable is it truly to hold equity in today's leading firms?

Question Five: How Profitable Can These Companies Really Be?

The basic logic of value investing estimates a company's future earnings or cash flows, discounts them to the present, and compares them against the current stock price. To forecast future earnings, we must answer:

• How many units of product will this company sell?
• What will be the price per unit?
• What will be the production cost per unit?
• What portion of profits must be reinvested into R&D for the next generation of products?
• What ultimate return on investment (ROI) can this company achieve?

While answering these questions for any business entails uncertainty, applying them to AI introduces numerous particularly thorny new considerations:

• What exact capabilities will its products possess, and how will customers deploy them?
• Will the product ultimately be priced as a commodity, or become an irreplaceable core necessity?
• How long is the actual useful life of the chips, and at what rate must data center depreciation be amortized?
• How long will this reckless, spend-at-all-costs burning race continue?
• Will future profits adequately support respectable returns on the massive current capital expenditures (Capex)?

My memory might be fuzzy, but I truly cannot recall facing uncertainty problems of this magnitude back in the 1970s when analyzing diversified conglomerates, or during 1999–2000 when scrutinizing the internet and e-commerce boom.

Although those business models appeared novel and complex at the time, compared to today's AI, they seem straightforward and transparent.

This leads me to wonder: Can we truly accurately forecast the earnings of today's hyperscalers and top frontier labs by 2030 (let alone companies currently propped up almost entirely by imagination)? Must we be able to relatively precisely predict earnings before committing capital?

Even if we scrape together a set of earnings forecasts, how much confidence should we place in those numbers? And if we absolutely cannot calculate them accurately, how should we build a decision framework for investing in AI companies?

I often pose a test now: Tell me what you believe a specific AI company's earnings per share (EPS) will be in five years. Dare you bet that the final actual result will fall within ±50% of your predicted figure?

(For example, if you guess the EPS in five years is $10, how confident are you that the actual figure will land between $5 and $15?)

Question Six: What Are the Prospects for Data Centers?

Historically, data centers served as digital warehouses for storing data, hosting computer applications, and websites. Today, the definition of a "data center" (sometimes called an "AI factory") has evolved into the operational workshop for AI.

Like an automobile manufacturing plant, a data center is the base for producing AI. Its duties encompass information storage, model training, and ultimately providing system computing runtime to the outside world (the so-called "Compute").

Data centers are undoubtedly the cornerstone of AI industry development. However, their construction and the resulting compute supply face numerous constraints from physical and other dimensions.

Despite the exciting prospects of AI, some investors may prefer to participate cautiously and prudently. This drives attention toward tangible infrastructure rather than direct developers and suppliers of emerging technologies, akin to selling shovels and picks to gold rush miners rather than betting on which miner will strike gold.

Theoretically, this strategy reduces the importance of correctly picking the ultimate winner, since whoever prevails will inevitably rely on underlying infrastructure. Current compute shortages further enhance the investment appeal of data centers.

Nevertheless, investors allocating capital to data centers must fully weigh a series of critical factors:

• When hearing "infrastructure," people often imagine stable "steel and concrete." This tangible quality makes data centers seem like the safest AI investment. However, the building itself (the so-called "Shell") comprises only a tiny fraction of total costs; chips or "GPUs" account for the bulk. Chip procurement and technological obsolescence risks are typically borne mainly by data center tenants, not owners.
• Project developers must source and purchase large parcels of land, navigate layers of regulatory approvals, and overcome local opposition to building facilities in their neighborhoods.
• Existing fiber optic networks must be available to ensure seamless connectivity for inference-side users.
• Data center builders must also tackle potential supply bottlenecks regarding required power capacity, cooling equipment, and construction services.
• Limited access to project financing could slow overall development, while future financing costs may exceed market expectations.

Interestingly, while the aforementioned potential bottlenecks certainly complicate matters and restrict expansion for developers intent on building data centers, every coin has two sides.

Any restriction on the growth rate of "compute" supply (such as the signal sent by several bans on new data center construction passed last week) should instead make existing facilities and future approved projects more profitable, thereby multiplying their valuation.

The key is that data centers are absolutely not guaranteed-win "layups." Prudent investors must seek opportunities that effectively shield against the aforementioned uncertainties and avoid exposure to oversupply shocks from overbuilding.

As in all investment domains, the data center sector will see both triumphant winners and devastating losers. Given the unique operating logic of the AI industry, I believe the divergence between victors and vanquished may even be sharper than in traditional "infrastructure" sectors.

Question Seven: Should We Worry About the Market's Overreliance on This Sector?

Right now, AI dominates a highly disproportionate share of the entire investment landscape. This poses challenges not only for AI investors but also introduces considerable uncertainty for all investors and economic participants alike:

• The AI industry accounts for a minuscule fraction of total US economic output, yet contributes a disproportionately large share of economic growth.
• Currently, AI's contribution to GDP growth stems almost entirely from capital expenditures, concentrated in data center construction. Investment in data centers is skyrocketing, while construction investment in other US sectors contracts.
• Through June 30 this year, the "Magnificent 7"—dominated by hyperscaler cloud providers—have seen earnings growth outpace the rest of the S&P 500 constituents for 14 consecutive quarters.
• Certain AI firms recorded massive paper gains from equity stakes in peer AI companies, becoming the primary drivers of overall S&P 500 earnings growth. According to FactSet, Alphabet and Amazon alone booked a combined $151 billion in equity investment gains in Q2, primarily from holdings in Anthropic and SpaceX (now merged into xAI).
• The AI-driven pull on aggregate societal earnings extends beyond the industry itself. Related component suppliers also delivered stellar financial results. Meanwhile, AI practitioners and supporting industry employees spend their wages on rent, food, automobiles, and other everyday consumption unrelated to AI, creating extensive ripple effects.
• AI's contribution to US equity market gains is similarly concentrated. JPMorgan estimates that since January 2024, a basket of 42 AI-related companies accounted for roughly two-thirds of the S&P 500's gain.
• The cumulative effect of these trends leaves the S&P 500 heavily concentrated in the hands of a handful of AI and tech titans.

The core observation here is that the driver underpinning our economy and capital markets appears highly singular and lacking diversification. AI risk exposure has transformed into an exceptionally potent "geological fault line" running through portfolio constructions, possessing enough energy to dominate every facet of the investment environment.

Any negative developments concerning chip lifespans, corporate strategic pivots, or genuine compute demand could trigger profoundly far-reaching domino effects.

Undeniably, excessive reliance on a single industrial sector renders the broader economy and stock market significantly more fragile than under a multi-pillar support structure.

Between 1977 and 1978, as Michael Milken pioneered the issuance of below-investment-grade bonds (high-yield debt), he established a consensus across the investment community: markets embrace risk as long as potential returns adequately compensate for the risks taken.

His core theory rests on two pillars: (a) compensating for losses from individual bond defaults via wider coupon spreads; (b) maintaining sufficiently robust diversification to ensure that rare default events do not derail overall portfolio performance.

Today, however, both equity and bond markets are highly leveraged bets on a single core variable: whether AI actually succeeds.

Question Eight: What Does This Mean for Credit Markets?

According to Goldman Sachs, year-to-date global bond issuance related to AI has approached $600 billion, covering hyperscaler debt offerings, data center project finance, and other AI-linked borrowers.

Approximately $400 billion of that flowed into the US investment-grade (IG) bond market, accounting for over one-fifth of total IG issuance during the period and pushing AI-related bonds' weight in the Bloomberg US Dollar Investment Grade Corporate Bond Index to roughly 12%.

In the US high-yield bond market where hyperscalers are not direct issuers, AI-related bonds also reached a weight of about 6%. Clearly, AI-linked debt issuance has become one of this year's most significant capital market developments.

Looking ahead, the AI industry apparently requires up to $1 trillion in incremental capital annually over the coming period, with hundreds of billions likely continuing to be raised via bond issuance each year.

What kind of impact will such colossal debt financing demands have on market interest rate levels?

Since expanding compute capacity and developing new products are existential imperatives for hyperscalers and frontier labs, and they refuse to watch competitors exploit openings, these firms likely lack the leeway to slow AI investments even if interest rates exceed forecasts.

This means that regardless of tightening credit markets and rising financing costs, they must reluctantly continue borrowing as passive "price takers" of capital.

In the past, economists frequently discussed the "crowding out effect," referring to governments borrowing heavily and making it difficult for the private sector to raise necessary debt capital.

But what strikes me now is the opposite: hyperscalers, insensitive to interest rates and willing to borrow at any cost, generate such massive debt demands that they could inversely push up the borrowing rates governments must pay. Is part of the rate increases we've seen thus far stemming from this?

Similarly, beyond internal corporate cash flows, does the equity market have the capacity each year to supply another $100s of billions in incremental equity investment? For eight of the past ten years, total capital raised in the traditional US IPO market remained below $50 billion (only approaching $80 billion in 2020 and peaking at $140 billion during the 2021 frenzy).

I suspect the total pool of incremental capital mobilizable in any given year is fundamentally finite. Consequently, wouldn't AI's massive appetite for debt and equity capital ultimately compete with itself within the same funding reservoir?

Once upon a time, the highly profitable non-AI core businesses of hyperscalers continuously generated massive free cash flows, leaving them nearly debt-free, flush with cash, and holding top-tier credit ratings.

But today differs from yesterday. These giants are now rapidly depleting cash reserves and current operating cash flows while simultaneously increasing leverage. Although they can still easily finance AI expansion thanks to robust investment-grade ratings, will their credit profiles and ratings themselves change as this cash-burning race drags on?

What consequences might this hold for creditors who generously lend today trusting high credit ratings, only to watch borrower leverage ratios climb relentlessly?

Currently, some data center financing secures placement primarily relying on the investment-grade credit backing of cloud titan tenants, while portions of debt are directly collateralized by chips.

However, investors must remain highly vigilant regarding hidden "out clauses" in certain data center leases, which allow tenants or guarantors to terminate obligations under specific conditions (such as failing to meet 99.9% uptime reliability standards); they must also guard against the potential mismatch risk where chip collateral lifespans expire before debt maturities.

Consider so-called "circular transactions" within the industry, where companies directly fund clients to purchase their own products. What hidden risks lie within this structure?

We frequently observe inter-lending and affiliated purchasing among chip manufacturers, hyperscalers, frontier labs, and data center developers. Sometimes funds flow from A to B and back to A; sometimes they move from A to B, get refinanced and handed to C, ultimately ending up spent back in A's hands.

Are these transactions normal dealings based on fair, independent commercial principles, or disguised subsidies by suppliers propping up demand and defending market share? What systemic risks lurk within this highly interdependent ecological network?

Furthermore, how do off-balance-sheet financing arrangements via special purpose vehicles (SPVs) truly impact the underlying creditworthiness of hyperscalers? Does the parent company ultimately guarantee this off-balance-sheet debt, or leave it unguaranteed?

Lastly, what are the consequences of issuing massive debt during periods of fervent enthusiasm surrounding exciting technological breakthroughs? Ultimately, we face the superimposed resonance of the following factors:

• Absurdly massive debt financing demands from AI firms;
• Highly euphoric investor sentiment toward AI, fueling exceptionally generous lending appetite;
• As noted earlier, capital markets' heavily concentrated bets on a single AI sector;
• Intertwined "circular transactions" among AI supply chain participants.

Lenders extending credit to the AI industry are stepping into a brand-new era: debt highly concentrated in one sector, fundamental dynamics shifting constantly, and extreme reliance on the realization of future compute sales revenue.

Lenders must maintain extraordinarily heightened vigilance regarding the creditworthiness of supported borrowers, as well as the structural design and contractual terms of lending deals themselves.

Question Nine: How Predictable Is AI Exactly? Can These Questions Even Be Answered?

In previous memos and my book Mastering the Market Cycle: Getting the Odds on Your Side, I have written repeatedly: for anyone attempting to navigate the future, mere forecasting is far insufficient.

Beyond forecasting, you must evaluate the probability that your forecast proves correct. In my view, considering AI's inherent complexity, unprecedented disruptive nature, and blistering evolution rate, the probability of making accurate predictions in this arena is extraordinarily low.

I initially considered quantifying AI's predictability, but worried it vastly exceeded my expertise. Thus, I enlisted Claude's help to score the AI industry on a scale of 1 (completely unknowable) to 10 (fully predictable).

Its response proved highly illuminating. Combining uncertainties discussed above, Claude summarized as follows (bolded parts added by myself):

My overall assessment: The AI industry's predictability scores roughly a 3.This composite figure masks internal divergence. The "broad direction" of technological evolution is relatively predictable, scoring around 6 to 7: technology will continue iterating and permeating the entire economy.

But the key factors determining whether investors make money are far harder to predict, scoring merely around 2: how fast this progresses, how large the ultimate profit pie becomes, and who ultimately owns the profits.

I also asked Claude to score predictability for distinct sub-sectors along the AI value chain and identify non-AI economic industries it considers comparable in predictability. Below are its self-described "judgment calls":

Chart: Core data comparison across AI sub-sectors, predictability scores, and comparable industries

Most surprising to me was that frontier labs scored only a 2, which in my view practically screams "essentially unknowable." In my eyes, they are clearly exceptional companies wielding massive lead advantages.

Yet Claude's evaluation here is notably stark: "The outcome of this type of business resembles a binary gamble (success or failure). Capital requirements are enormous, per-unit capability pricing is rapidly falling, and it remains unclear whether enduring moats even exist."

(Considering Claude itself belongs to Anthropic, isn't this statement somewhat ironic?) Intriguingly, according to Claude's assessment, these least certain targets often enjoy exorbitant valuation premiums in the private equity market.

Both OpenAI and Anthropic have filed IPO applications. How they will price in public markets and how their post-listing stock trajectories unfold will undoubtedly be compelling to watch.

Do not forget, as mentioned earlier, a substantial portion of the massive equity investment gains that drastically boosted S&P 500 earnings in Q2 came precisely from this sector deemed "most unpredictable." How long-lasting will these paper gains truly be?

On predictability, I also want to briefly touch upon geopolitics. In the AI context, this usually directly refers to the "China factor." Currently, virtually all cutting-edge chips originate from Taiwan-based TSMC, while the US government restricts sales of advanced chips to mainland China. This ironically incentivizes Chinese entities to launch highly cost-effective low-cost foundation models, threatening the sales of superior-performing American models.

If this landscape shifts, what impact will it have on US competitive advantage? Will China-developed chips someday catch up to or surpass today's industry titans?

Will mainland China ever possess the capability to obstruct or even sever the supply of Taiwanese chips to the US? Furthermore, will concerns over China's strategic intentions undermine the US regulatory resolve to establish safety guardrails for the AI industry?

Evidently, too many unknown and unpredictable variables regarding AI remain. The real question boils down to: How much risk are you willing to bear, and in what form will you assume it?

The AI supply chain spans numerous links. Based on Claude's assessment, predictability ranges drastically: on one end sit frontier labs that leave us nearly blindfolded; on the other are data centers with solid long-term leases and their power suppliers, whose predictability roughly mirrors large industrial enterprises and regulated utilities.

Question Ten: Is AI a Bubble?

In recent weeks, worries surrounding the myriad issues outlined above have begun gradually surfacing.

Multiple headwinds resonate within the macro investment landscape, focusing on sticky inflation, federal budget deficits, soaring national debt, resulting rate hikes, and seemingly deadlocked conflict in Iran. Layered atop these macro tailwinds are several AI-specific concerns, such as:

• Fierce competition from newly launched models, streamlined yet affordable models, low-priced Chinese foundation models, and the impact of free downloadable open-source models;
• Continuous decline in price per "Token" (billing unit for model usage);
• Potential deceleration in the growth rate of compute utilization;
• Revenue growth disclosures falling short of market expectations at certain companies;
• Market anxiety over elevated leverage levels and heavy dependence on future external financing at select firms.

During bull markets, people typically dismiss doubts with a laugh, but once psychology cracks, these issues become sufficient to unsettle investors' sleep and appetite. Thus, we once again encounter my oft-repeated observation: Reality usually oscillates between "quite good" and "mediocre," but in investors' minds, it often plunges instantly from "flawless" to "hopeless."

The myriad concerns troubling AI investors lately are hardly novel. It is merely the chilling of the broader macro investment climate that has allowed these old worries to ferment and utterly darken market sentiment.

Although this debate has raged across circles—including mine—for over a year, people persistently ask: Is AI actually a bubble? First, we must define a bubble. My definition is: irrational exuberance driving rash behavior, overinvestment, and excessively rich entry valuations.

I believe evaluating an AI bubble (just as examining any potential bubble around a technological breakthrough) requires at least three dimensions: Is the technology itself overhyped and destined to disappoint? Has the scale of capital deployed by firms become excessive, making rational returns unattainable? Have investors assumed excessive risk in this space, paying overly steep prices for it?

Reaching conclusions on these topics relies heavily on the answers to the preceding questions. Since publishing my previous AI-focused memo in February this year, I feel the following three judgments have not shifted dramatically:

• I maintain that AI itself will inevitably demonstrate profound power and impact, triggering transformations in corporate operations, macroeconomic dynamics, and broad human life that we cannot yet fathom. Therefore, the technology itself is certainly no mirage.
• Currently, vigorous debates rage over "how much future revenue must be generated to justify current massive capital expenditures." But until we fully comprehend how AI will be deployed and who will actually profit from it, we cannot declare whether today's spending is prudent.
• Until that understanding allows us to reasonably estimate future earnings and infrastructure profitability, we likewise cannot judge whether current secondary market stock prices or bond yields represent sound value.

Despite the labyrinth of questions swirling around AI (some listed above), the core of all contradictions can actually be distilled into one highly complex ultimate query:

Can AI deliver sufficiently massive practical value to users, thereby sustaining enormous compute sales volumes and pricing power, ultimately allowing the trillions of dollars currently poured into this field to earn desired returns (and identifying precisely who captures those returns)?

Based on my exchanges with numerous industry veterans, my conclusion is: standing at this precise moment, answering this ultimate question relies overwhelmingly on faith. It reminds me of a quote by American economist and author Stuart Chase:

Proof is unnecessary for believers; proof is futile for skeptics.

I racked my brains and truly cannot find a footnote better suited to neatly wrap up the multitude of AI questions than this sentence.

Question Eleven: How Should We Think About AI Investing?

In my view, artificial intelligence is the most impactful yet most unpredictable major shift for investors in at least a generation. But investors absolutely do not need to reject or shy away from something simply because it is shrouded in uncertainty.

The proper response is never paralysis and inaction. Instead, it involves two things: (a) maintaining calibrated humility regarding strategy selection, position sizing, and risk assumption; (b) retaining the courage to act decisively even amidst uncertainty.

Investing across different links of the AI supply chain demands solving entirely different propositions. An equity investor in a frontier lab might need to correctly guess the pace of technological advancement, competitive positioning, pricing power, future profitability, and valuation multiples simultaneously.

By contrast, a lender financing a data center based on a long-term lease faces fewer variables to guess, but the dimensions truly deciding success and failure diverge sharply, encompassing tenant credit quality, lease contract terms, power supply guarantees, and residual asset values.

Thus, the key inquiry presented is: "Among all these questions, which ones does my institution truly excel at answering? And in areas lacking professional depth, which uncertainties am I willing to actively shoulder?"

Ceteris paribus, as long as an investment's entry price is sufficiently cheap relative to its intrinsic value, high degrees of uncertainty are perfectly acceptable to investors.

Theoretically at least, if high-risk segments ranking lowest on Claude's predictability scale trade at sufficiently cheap valuations, significantly discounting high-confidence segments, the former could even offer superior investment risk-adjusted returns. Ultimately, smart investing is largely an art of pricing.

However, it must be pointed out that specifically within the AI track, the thorny dilemmas explored above are rendering the calculation of intrinsic value and subsequent determination of "cheapness" exceptionally complex.

To thread together these concluding thoughts, I leave you with a checklist of probing questions that may aid investment decisions. Note that these questions apply to all investment targets, not solely AI:

• How many prerequisite conditions must concurrently materialize flawlessly to generate an acceptable return?
• If my core assumptions prove incorrect, what is my maximum potential loss?
• Among the myriad risks undertaken, which ones am I genuinely compensated for by adequate pricing, and which ones am I assuming blindly and gratuitously?
• How much optimism has already been factored into current asset prices?
• How will this investment alter the risk profile of my overall portfolio? Does it further amplify my existing exposures, or serve a diversification/hedging function?

An old adage fits AI investing particularly well: "Two types of investors lose large sums: those who know nothing, and those who think they know everything." I believe that when facing AI, openly acknowledging one's own ignorance is paramount.

AI will undoubtedly birth endless surprises and marvels in the future, though these surprises need not be unbearable. What truly warrants repeated weighing is: What is the probability that an AI investment disappoints, versus the probability that it exceeds expectations?

Finally, allow me to conclude with a brilliant quote I randomly stumbled upon recently:

One argument posits that the near-term future holds too many unknowns: Would it not be wiser to wait until circumstances clarify slightly before acting? You are certainly familiar with that platitudes: "Preserve cash dry powder, wait patiently until current uncertainties dissipate before deploying."

But before grabbing this psychological crutch, please confront two uncomfortably simple truths: the future will never fully clarify; in equity markets, waiting for universal bullish consensus often demands exorbitantly high prices. In fact, uncertainty is the true friend of the long-term value buyer. (Bolded by myself)

Perfectly tailored for closing a memo discussing AI, isn't it?

By the way, this passage lifted from my dusty archives: Warren Buffett's timeless wisdom published in Forbes magazine on August 6, 1979.

I have always preferred to leave the final summing up to Warren.

November 11, 2026