The More You Believe in Technology, the More You Should Believe in Value Investing

For the past three years, I’ve stood in two places in AI at once, and gotten opposite results. In 2023, I threw myself into an AI application-layer startup (back then, models were leaping forward every few months, and capital was happy to pay for imagination). Three years later, the underlying capabilities of AI have advanced even faster than we expected, but the company didn’t survive: first our cash flow tightened, little by little, and then the fundraising window shut. Over the same period, as an ordinary investor, I bought NVIDIA and Tesla. Both have delivered substantial excess returns, and I still hold them for the long term.

Both results came from the same belief: AI will change the world. That belief has never wavered. What wavered was an inference I had taken for granted: that if the technology wins, the people betting on it win too. There’s a long road between those two things, yet most investment talk about AI skips it and reduces the question to a single choice. If you believe in AI, you should buy; if you don’t buy, you’ve failed to understand an industrial revolution. I believe in AI more than most people do. But precisely because I’ve watched the technology keep moving forward while capital ebbed and flowed, I take value investing more seriously than ever.

This year, the Munger Academy asked: does value investing still have value? My answer is that in the age of AI it has more value than ever. AI makes value easier to create and harder to keep, and value investing measures the distance between the two. AI is stretching that distance wider than ever before (a lead may last only a few months, the spending required to keep it keeps growing, and the risks often sit outside the financial statements). The wider the gap, and the harder it is to measure, the more you need the discipline of value investing.

Better Looms, Poorer Mills

In his 1994 talk at USC’s business school, Munger said that the great lesson in microeconomics is to discriminate between when technology is going to help you and when it’s going to kill you. Berkshire’s textile mill was his example. People came to Buffett and said a new loom had been invented that would do roughly twice as much work as the old ones. Buffett replied: “Gee, I hope this doesn’t work, because if it does, I’m going to close the mill.”

Behind the joke was a cold ledger. Low-end textiles are a textbook commodity. Anyone can buy the new machine, so once costs fall, they simply turn into price cuts across the industry, and all of the benefit flows into the pockets of the people buying the cloth. In Munger’s words, nothing was going to “stick to our ribs as owners.”

Munger went straight on: if you own the only newspaper in Oshkosh and someone invents a more efficient way of composing the paper, then when you get rid of the old equipment and put in new computers, all of the savings come right through to your bottom line. The same technology is a burden to a textile mill and a gift to a newspaper with no rival. The only difference is the competitive landscape.

Munger pointed out that the people selling the machinery, and the people inside the company pushing to buy it, will all show you projections: how much the new equipment saves each year, how many years until it pays for itself. What they never do is the second step of the analysis: how much of the savings will stay home, and how much will flow through to the customer in lower prices? He said he had never seen a single projection that included that second step. So you keep buying equipment that “pays for itself in three years,” and twenty years later you look back and find you’ve earned a return of about 4 percent a year.

Buffett had an analogy for this: at a parade, everyone thinks they’ll see a little better by standing on tiptoe. Then everyone stands on tiptoe, and no one sees any more than before. Burlington Industries, the largest textile company in America, spent about $3 billion on capital expenditures over twenty-one years, and its shareholders’ purchasing power shrank by about two-thirds. Buffett’s conclusion about textiles was: don’t invest. Berkshire’s later conclusion about Apple was: buy heavily. Opposite conclusions, but the same question: who ends up keeping the benefits of technological progress? People who only learn the conclusions will, sooner or later, apply them in the wrong place during the next technological shift.

AI is this generation’s loom, and everyone can afford it. Intelligence will keep getting cheaper; the question is whose house the savings end up in. My view is that AI will amplify the competitive structures that already exist. A company whose moat doesn’t depend on scarce intelligence can use cheaper intelligence to cut costs and keep the savings, because its rivals can’t force it to pass them on. For a company without a moat, the same savings will quickly turn into price cuts across the industry. Today’s “only newspaper in Oshkosh” goes by names like Coca-Cola, Apple, McDonald’s, and Heineken. Others can make the drinks, phones, and burgers they sell, yet billions of people, when in doubt, reach for them without thinking. That trust takes decades to build, bit by bit, and no amount of technological progress can build it for anyone overnight. The cheaper intelligence gets, and the easier products become to copy, the scarcer that trust becomes.

Bill Ackman has carried Munger’s second step into a modern setting. An AI coding company can rewrite a big bank’s decades-old legacy systems in modern code, doing in days what used to take months, and the bank’s operating costs drop meaningfully as a result. Can the bank keep those savings, or will it hand them to its customers? Ackman’s view is that it depends on the nature of the business and on pricing power, and money is itself a commodity: money is what banks sell. So in the age of AI, the good investments won’t necessarily be in companies that sell intelligence. They may be in companies that buy intelligence and get to keep what it saves them.

Raisins Can’t Save a Bad Business

Run the second-step analysis on our own company, and we were on the textile mill’s side. Our product packaged the capabilities of large models into a tool for one specific use case. In 2023 that packaging had value (the models were still unreliable, and it took a lot of engineering to turn them into something users would pay for), and that work was our reason to charge. But every time the models improved, part of our work became redundant. People who build AI applications have a saying: “the arena resets every three months.” Each model release wipes out most of whatever lead you had built. Users got, for less or even for free, capabilities they used to pay for, and the benefits of technological progress passed straight through us to them. Customer acquisition stayed expensive, the big platforms controlled the traffic, and plenty of people tried the product while few paid for it long term. We used to think our features and accumulated technology were our advantage. Only later did we realize those are exactly what AI is best at copying. Our business itself was being repriced by AI.

At Berkshire’s 2000 annual meeting, talking about internet stocks, Munger said: “If you mix raisins with turds, they’re still turds.” (Real progress, once it’s mixed up with speculation, still ends badly.) Our company was that kind of bowl. AI was the raisin, but if a business has neither a moat nor pricing power, no technology, however world-changing, can save it.

This mechanism shows up everywhere among companies born of technological revolutions. One layer up the stack, frontier model companies look more and more like the business Buffett hated most. At Sun Valley in 1999, Buffett did the math: from the dawn of aviation through 1992, the U.S. airline industry’s cumulative profits were zero. He also said that the worst sort of business is one that grows rapidly, requires significant capital to engender the growth, and then earns little or no money. “Think airlines.” Seats are hard to differentiate. Competition hands the fuel savings from new planes over to passengers through fares, and you have to keep buying new planes just to keep up (the global airline industry’s net margin this year is expected to be only about 2 percent). Model companies are in a similar position. Their products are hard to differentiate, and open-source models typically trail the frontier by only three or four months. Switching providers isn’t hard for customers either: Anthropic’s draft prospectus shows that nearly a quarter of its 2025 revenue came from two customers, and many large customers have not signed long-term contracts.

Anthropic CEO Dario Amodei once sketched some illustrative math. A model trained in 2023 for $100 million brings in $200 million of revenue in 2024. But in 2024 the company spends $1 billion training the next one, which brings in $2 billion in 2025, while the company is spending $10 billion to train the one after that. Viewed model by model, training and deploying each one is a high-margin business, yet the company as a whole loses more and more (in 2025, Anthropic had revenue of about $4.6 billion and an operating loss of $8.06 billion). The business model of a model company is that the money each generation earns has to be poured into a more expensive next generation, and as long as rivals keep chasing, the vicious cycle can’t stop. It’s the same as airlines replacing planes, except that a plane can fly for more than twenty years, while a model generation stays ahead for only a few months. That’s why I’ve always had reservations about the economics of frontier labs.

Losing money today doesn’t mean you can’t make money later. There are two kinds of losses. They look similar on the income statement, but they are fundamentally different. One kind is like GEICO’s. Buffett estimated that, thanks to referrals from existing customers, GEICO could maintain its policy count with only a fraction of its advertising budget; most of the rest was spent deliberately, for growth. What it bought was households that renew year after year. Stop the spending, and the customers are still there. The other kind is like that textile mill’s: the spending only keeps you from falling behind, and once it stops, customers and revenue walk out the door.

Buffett’s “owner earnings” subtract the capital expenditures a business requires to fully maintain its long-term competitive position, and for model companies, that word “requires” is exactly what’s hardest to estimate. I test it with a thought experiment: if a model company stopped training for a year, would its customers stay? Would its revenue hold? If so, today’s spending looks more like GEICO’s customer acquisition. If not, it looks more like Burlington’s equipment upgrades. In today’s landscape, the answer is no. If someday model companies’ gross margins keep thickening, existing customers reliably throw off cash, and training spend shrinks as a share of revenue year after year, I’ll revise my view. This year has already brought one signal pointing the other way: some new models are priced higher than the generation before. That could be the beginning of pricing power, or it could just be costs being passed on. It remains to be seen.

The Internet Won, the Investors Lost

In 2023 and 2024, the question the market asked most was whether AI would pay off at all. Model companies’ revenue answered half of it. OpenAI’s annualized revenue went from $2 billion in 2023 to a reported nearly $70 billion as of this September, which demolished the claim that “nobody will pay for AI.” The other half remains unanswered. Someone paying doesn’t mean the people being paid have earned back what they put in (according to the Financial Times, OpenAI expects cumulative negative free cash flow of about $278 billion from 2026 through 2030), let alone that all the capital across the industry chain will be recovered. Creating value and capturing value are two different things, and the history of fiber optics proves it.

In the late 1990s, a popular claim in the U.S. telecom industry was that internet traffic doubled every 100 days. Researchers later estimated that from 1997 on, traffic roughly doubled every year. Over five years, those two rates differ by a factor of nearly ten thousand, and people at the time were building to the first one. On the supply side, new technology multiplied the capacity of a single strand of fiber dozens of times over in just a few years. Demand really was growing fast, but supply grew faster, and bandwidth prices collapsed. Global Crossing spent about $15 billion building a transoceanic network, filed for bankruptcy protection in 2002, and a little over a year later, more than 60 percent of its equity sold for just $250 million.

In that same bubble, the companies whose position most resembles today’s AI upstream were the ones selling equipment. To move their gear, the telecom equipment giants Lucent and Nortel extended large amounts of vendor financing to cash-strapped new carriers, so customers were effectively buying their equipment with money they had lent. During the boom, these deals showed up all at once as the vendors’ revenue, receivables, and investments in customers. When the customers fell, they all turned into losses together. Lucent’s stock fell from a peak of $84 at the end of 1999 to about $0.55 in 2002. Nortel later went bankrupt.

The fiber didn’t disappear. After bankruptcies and restructurings, it changed hands at prices far below what it had cost to build, and the internet’s later growth didn’t wait for the first builders to earn their money back. The internet won, the first fiber investors lost, and the people who picked up the assets cheaply made money. A study from the Richmond Fed pointed out that this overbuilding didn’t require everyone to lose their minds: every company assumed it would win a sizable share, and every business plan made sense on its own, but everyone’s shares added up to more than 100 percent.

Our company fell to a similar mismatch, just on a much smaller scale. The technology changed every few months, and each change rewrote the value of our product. Commercialization, meanwhile, was slower than anyone expected; you can’t rush the day users are willing to pay steadily. But salaries, server bills, and investors’ demands for growth arrived right on schedule. Commercialize a year late, and the question is no longer whether AI has value, but whose cash lasts through that year. Ours didn’t. The mismatch in timing explains how we fell. But why we couldn’t withstand it comes back to Munger’s bowl of raisins. A business with a moat and pricing power can live on its own earnings when commercialization slows. A business without one, when capital pulls back, can only press its palms together and pray the funding doesn’t stop.

The Bill Isn’t in the Index

The same mismatch is now playing out across the AI industry chain, only with stakes several orders of magnitude larger. The upstream players (chips, memory, and data centers) were the first to make money. Model companies may not have great business models, but their revenue is real. Ultimately, though, the chain depends on someone downstream paying the bill, and the biggest payers are the companies spending heavily to buy model intelligence. Whether they make money is the question that matters most. The question of returns hasn’t been solved, only postponed and pushed onto downstream businesses. Who will make money? The answer is still Munger’s second step. Companies with moats get to keep the savings at home. Companies without moats buy AI the way a textile mill buys looms: skip it and you fall behind; buy it and the savings go to your customers anyway. Until that question is answered, however real the upstream orders and profits are, they rest on an assumption that hasn’t fully paid off yet (the equipment makers’ revenue in the fiber era was real too). If downstream companies don’t see enough return on their AI spending, the money they can squeeze out of their balance sheets and pass up the chain will shrink. And today’s upstream valuations clearly already price in a great deal of future demand for compute.

Risk also hides in the ways industry players finance one another. In August, NVIDIA disclosed that it is providing a residual value guarantee on a large data center campus leased to an OpenAI affiliate, with cumulative payment obligations capped at about $105 billion. If the tenant defaults, NVIDIA has to cover the gap between the asset’s guaranteed minimum value and whatever it fetches when it is re-leased or sold. That money hasn’t been lent out, and it isn’t an expected loss. But a guarantee like this is most likely to be called on precisely when the assets are worth the least and the tenant most needs cash.

The campus owner, SB Energy, belongs to SoftBank, whose cumulative investment in OpenAI has reached $64.6 billion; its most recent tranche was funded with dollar bonds carrying coupons as high as 9.75 percent. These deals are all real, and the equipment really was delivered. The problem is that the same risk is showing up on several companies’ books at once, under different names (revenue, equity investment, guarantees), while ultimately depending on a single source of payment: how much end customers are willing to keep paying for AI. That is exactly where Lucent came undone.

Some of the risk doesn’t show up in stock indexes at all. In 2000, the bubble was plain to see in the index (Cisco traded at more than 200 times earnings). Today the common refrain is “this time is different,” because most of the Nasdaq’s biggest companies make real money and the QQQ trades at only a little over 30 times earnings. That argument misses something: this cycle’s most aggressive pricing is happening in private markets. Many of the neolabs founded by researchers who left frontier labs have no revenue yet but are already valued at billions or tens of billions of dollars. Embodied AI companies in both China and the U.S. doubling their valuations within a few months is not unusual. These prices all bet on the same thing: being the last one standing in a race that may change the world. But every one of those valuations assumes it will be that one, and once again, the shares add up to more than 100 percent.

Ackman attributes this pricing to the fear of missing out on the future. Pershing Square recently started investing in early-stage companies too. One company they met wasn’t even planning to raise money; two days later, an investor preempted with $50 million at a $400 million valuation, and two weeks after that it raised another $50 million at a $1 billion valuation. Too much money is crowding into the same deals, scrambling to bid.

These valuations aren’t in the index’s P/E, but public markets are far from insulated from them. Through equity stakes, cloud contracts, and guarantees, public companies have already taken some private-market risk onto their own books, and IPOs are another channel. Unitree Robotics, one of the few companies in this group that is already profitable, listed on Shanghai’s STAR Market in August; at the open, its market value briefly topped RMB 440 billion, more than 700 times the previous year’s profit. As more neolabs, embodied AI companies, and model companies go public, today’s private-market valuations will become tomorrow’s index P/E. Using today’s index P/E to prove that this time is different measures only the tip of the iceberg.

Analogies have limits. GPUs have a much shorter economic life than fiber, today’s big tech companies have far stronger cash flows than the telecom upstarts of that era, and I’m not predicting when a correction will come. What a value investor can do is stress-test from the bottom up. If end customers’ payments arrive a year later than the build-out plans assume, who runs short of cash first? Who has to honor their commitments no matter what? Who can sell their assets on? There’s no single answer to these questions. The only way is solid research, one company at a time.

Good Companies Can Be Bad Investments

In March 2000, Cisco’s market value briefly passed Microsoft’s, making it the most valuable company in the world, at more than 200 times earnings. For the twenty-plus years that followed, it remained a good business: net income grew to about five times what it was then. But its P/E fell to a little over 30, and the rise in one and the fall in the other cancelled out. Anyone who bought at the peak had to wait until December 2025 to get back to their purchase price. And the 2000 profits were themselves a peak: in the very next fiscal year, Cisco took a $2.25 billion charge for excess inventory and lost money for the year.

What matters more than whether a P/E is high or low is whether the earnings in the denominator are normal or a peak. Cisco’s denominator in 2000 was a peak. When I bought NVIDIA in early 2024, the situation was the reverse. Its P/E on the past year’s earnings was somewhere around 60 or 70, but it was at an inflection point, and its earnings at the time were far below where they would soon be. Earnings then grew at a blistering pace, the stock rose to three to five times its level at the time, and the P/E actually fell to around 30. A low P/E can deceive you just as easily, especially with cyclical stocks. When the memory cycle peaked in 2018, Micron’s P/E was only four or five, and while it was still reporting near-peak earnings, its stock had already fallen more than half from its high. In this round of memory price increases, Micron’s P/E is back in the single digits. Excess profits earned from tight supply usually don’t last long.

The same question applies to the NVIDIA I own: how much of today’s profit is normal, and how much comes from customers building ahead of demand and from temporary shortages? Tesla trades at more than 300 times earnings, a price that makes no sense without Robotaxi. I could be wrong about how things turn out for Tesla, too. Either way, the P/E itself can’t detect an inflection point. In an era of rapid technological change, the P/E is an unreliable ruler: it made NVIDIA look too expensive in early 2024, and it makes Micron look too cheap today. Value investing is often misunderstood as buying low-P/E stocks, but Buffett said long ago that growth and value are joined at the hip, and that it’s far better to buy a wonderful company at a fair price than a fair company at a wonderful price.

Munger liked to say, “Invert, always invert.” Valuation can be inverted too: start by asking what today’s price requires the company to achieve. Buy at a $1 trillion market cap and want a 10 percent annual return over ten years, and the company needs to be worth about $2.59 trillion a decade from now. At 25 times earnings, that means earning about $104 billion a year. A price is itself a forecast. If high growth, high margins, low reinvestment, low dilution, and a durable competitive advantage all have to hold at the same time for the price to make sense, that price will be extremely sensitive to any deviation.

Of these conditions, growth, margins, reinvestment, and dilution at least have historical data to go on. Only “durable competitive advantage” requires you to judge what the competitive landscape will look like twenty or thirty years out. A company’s value is the present value of the cash it generates over its lifetime, so investing for the long term means being reasonably confident about what the company will look like in ten, twenty, and thirty years (confident enough to hold it even if the stock market closed for ten years). And AI greatly raises the risk of disruption.

Once a moat narrows, the valuation collapses before the earnings do. In his 1991 letter, Buffett acknowledged that newspapers, television, and magazines were sliding from “franchises” into ordinary “businesses,” because readers and advertisers had more choices. He ran the numbers: the same $1 million a year in after-tax earnings was worth $25 million as a franchise that could keep growing, but only $10 million as an ordinary business. Later, the internet filled in that moat completely. The Daily Journal, which Munger chaired for decades, is itself a newspaper company. At its 2020 shareholder meeting, he said technology was destroying America’s daily newspapers, small papers like his own included. That same year, Berkshire sold its newspapers to Lee Enterprises. AI will only make this harder: some of the companies that look most dominant will disappear along the way. Each of us is bound to misjudge some company.

What you can’t see clearly is hard to price. The market errs in two directions here. Sometimes it bids on an unclear future as if it were certain. Other times, for fear of disruption, it marks down companies with genuine moats along with everyone else. The second kind of error is the value investor’s opportunity.

Better to Miss Out Than to Be Wiped Out

The harder valuation gets, the more you need to hold to the discipline of value investing, and Berkshire showed how during the last technology bubble. In December 1999, Barron’s ran a cover asking, “What’s Wrong, Warren?” Tech stocks soared that year, while Berkshire’s shares fell about 20 percent. In that year’s shareholder letter, Buffett explained why he didn’t buy tech stocks: he and Munger agreed that these companies would transform society, but they couldn’t tell which ones possessed a truly durable competitive advantage, and that was a problem, he wrote, “which we can’t solve by studying up.” A few months later, the dot-com bubble burst.

This story is often told as a triumph of value investing, but that’s only half of it: Berkshire also missed Google and Amazon. At the 2017 annual meeting, Buffett said GEICO had been paying Google $10 to $11 per click. They had seen firsthand how good that business was, and still didn’t buy. They had blown it, he admitted. Patience has a cost, but missing out isn’t the end of the road. In 2016, Berkshire began buying Apple. By then Apple was already one of the most valuable companies in the world, yet it traded at less than 13 times earnings, and its ecosystem and customer loyalty had been proven again and again. It became one of the most profitable investments in Berkshire’s history. In 2025, Berkshire bought Alphabet, eight years after Buffett admitted to missing Google. Buying a little later isn’t necessarily worse.

Duan Yongping, the Chinese entrepreneur and investor, often cites Buffett’s analogy: when a 300-pound man walks in, you don’t need a scale to know he’s fat, and a great company should be just as easy to recognize at a glance. Among today’s model companies, there’s still no Yao Ming, no one that everybody recognizes the moment it walks through the door. There will be winners in the end, but it’s too early to bet on who. AI will certainly create enormous wealth. But if your capital runs out before the winners emerge, will you still have any money left when Yao Ming finally walks in?

In “Is It a Bubble?”, Howard Marks wrote that transformative technologies always attract excessive enthusiasm and investment, build more infrastructure than is needed, and leave asset prices that prove, in hindsight, to have been too high; if this AI cycle doesn’t follow that historical pattern, it will be the first time. He walked through how things might end for data centers: the building boom creates a glut, some owners may go bankrupt, and a new generation of owners takes over the facilities at rock-bottom prices and makes money once the industry stabilizes. Fiber went down that road, and so did the railroads. But before you “buy the dip,” you need to know what you’re buying: the original company’s stock may already be worth nothing, and the real beneficiaries are those who buy the assets after the restructuring. And to pick up those assets, you need to still have money at that point.

I never try to time the macro. I don’t predict whether this will happen this time, or when, and I don’t short the index. I simply agree with the base rate Marks describes: in technological revolutions, cheap assets turn up more often than most people think. There’s no guarantee they’ll come. If they don’t, buying a great company at a fair price is still a good choice, as long as the price really is fair. And the valuations of frontier model companies, AI valuations in private markets, and the premium the market puts on anything labeled “AI” are, today, far from fair.

Munger once said that the most underrated part of Berkshire’s success is that they were almost never forced by circumstances to make a decision, so they always had options. Ackman designed Pershing Square the same way: its vehicles are all permanent capital. In a panic, investors can sell their shares, but the money stays in. That’s why they’ve been able to act in moments like the COVID crash and the financial crisis.

There are two kinds of waiting. One is waiting on price: I broadly understand the business, but today’s price doesn’t offer enough return. That wait can end because the price falls, or because earnings grow faster than the price. Pershing Square follows a list of companies over long periods. It has long since researched many of them thoroughly; the price is simply too high to produce the return they want, so they set them aside. Early this year, when the market worried that AI would disrupt the software industry, nearly every stock with any connection to software was repriced, and Ackman says they used exactly that opportunity to buy Microsoft (at a forward P/E in the low twenties when they bought). The other kind is waiting on understanding: I can’t yet judge how long a company’s advantage will last or whether its spending will pay off, and in that case a cheaper price doesn’t help. What a value investor waits for is a combination of price and evidence that works in their favor, and a decline is only one of the ways that combination comes about.

Building a startup changed how I understand the margin of safety. It’s the room left by price, cash, fixed obligations, and time combined, far more than a discount applied at the bottom of a valuation model. However cheap a company looks, if it can’t survive until demand materializes, or has to issue lots of new shares at a very low price, the existing shareholders get hurt all the same. And no valuation, however high, can pay next month’s bills. Long-termism also has to allow for admitting mistakes. Time doesn’t repair lost pricing power or diluted equity. Conversely, if the model companies I’m skeptical of eventually prove their cash returns, I should admit I was wrong and bid again.

Closing the Mill

Buffett closed Berkshire’s textile mills in 1985, but the more important decision had been made eighteen years earlier. In early 1967, cash generated by the textile business was used to buy National Indemnity, taking Berkshire into insurance. After that, Berkshire never again put significant new capital into the mills, nor did it buy fancier looms to keep up with its peers. Much of the compounding of the decades that followed was built on that retreat.

Buffett takes no pride in this chapter. He wrote that he deserved to be faulted for not getting out sooner; in those years, he had believed what he wanted to believe. If even Buffett could hang on for years too long in a business whose benefits were destined to flow to its customers, you can see how easy it is to skip the second step.

Value investors in the age of AI face the same task. You don’t need to bet on which loom is the most advanced. You need to figure out where ever-cheaper intelligence finally settles, and whether it can be kept there. My own two results, gains on NVIDIA and Tesla on one side and a lost company on the other, prove nothing about whether AI is real. They only reflect where I stood, what price I paid, and whether I had enough time.

Technology will win. I’ve never doubted that. What value investing has to answer is Munger’s second step: once technology wins, who keeps the benefits? And what price did you pay to get there?