It’s clear why this piece may be relevant to long/short investors, but why should long-only investors care? The answer is simple: When the unraveling comes, it may affect a broader swath of equities than many of us may like to admit. Companies such as Alphabet, Microsoft, TSMC, and numerous others that may appear reasonably valued at the moment will not be immune to the negativity that comes with post-bubble disillusionment.

This piece lays out a path for how the bubble may unravel. While focused on the high-flying data storage darlings, the potential sequence of events is instructive to investors in adjacent tech segments as well.

Table: The Current Mania, Quantified

The storage and memory complex is exhibiting mania characteristics that exceed the equivalent storage/memory cohort of 1999-2000 on most measurable axes. SanDisk (SNDK) trades at $1,478, up 3,866% over one year and 523% year-to-date. Western Digital (WDC) trades at $484, up 865% over one year and 181% YTD. Seagate (STX) trades at $812, up 621% over one year and 195% YTD. The trio carries a combined market cap of $568 billion. The peak-valuation comparables of the prior cycle, including EMC at 125-130x P/E, NetApp at ~700x P/E, Rambus at ~200x P/E, and Cisco at 201x P/E, eventually drew down 90 to 97 percent over the following 24 to 30 months.

Today’s stocks trade at trailing P/Es of 26 to 76x and 14 to 17x trailing sales on share prices that have advanced 9 to 40 times in 12 to 15 months. The mechanical analogue is uncomfortably tight; the differences are real but narrower than the bull narrative suggests.

Table: Peak Valuations and Drawdowns, Prior Cycle

This article quantifies the current setup, anchors it to the 1999-2000 mania with hard valuation numbers, builds a structured catalyst framework for the bear case, and sketches an unravel scenario. The bull case has genuine merit: hyperscaler capex visibility ($725 billion committed for 2026 by the Big Four alone, up 77% year-over-year), oligopolistic industry structure (HDD is now a three-player market), supply discipline in NAND, and real prepaid multi-year long-term agreements (LTAs) from creditworthy customers. But at current valuations the margin of safety is negative on any normalized earnings framework, and several of the leading indicators of cycle peak are already flickering.

Read more

The Question

His note, paraphrased: Berkshire is sitting on a record cash pile (about $380 billion at the end of Q1, now closer to $397 billion when you include T-bills), and Berkshire is a net seller of stocks. Is Apple really that overvalued, or is Buffett seeing storm clouds? A panelist with access to Buffett reported that Buffett has expressed concern about a large market drop. At the meeting, Buffett reprised his cathedral-and-casino analogy and said more people are headed to the casino; he called zero-day options “straight-up gambling.”

The member’s worry is historical. Does today rhyme with 1969, when Buffett shut his partnership down, sat for four-plus years, and then watched the market crater roughly 50% over 21 months? Or with 1999, when the Nasdaq jumped ~85% on the year, then collapsed 78% from its March 2000 peak while the S&P fell 56%? Liquidity is sloshing around. A large IPO pipeline may drain it. “I don’t like to make macro calls,” he wrote, “but these things are getting my attention.”

It’s a fair question. Here’s how I’m thinking about it.

Yes, the Market Is Richly Valued. The Indicators Aren’t Subtle.

I’ve believed the market has been quite richly valued for some time, and the headline gauges agree.

• The Buffett indicator (total US stock market capitalization to GDP) sits at roughly 232% as of May 8, 2026, the highest month-end reading on record. Buffett himself once said 75–90% was reasonable; anything over 120% suggests overvaluation. We are nearly double that ceiling. [source]

• The Shiller CAPE is currently in the 39–41x range, depending on the data provider. The long-run average is around 17x. The all-time high was 44.2x at the dot-com peak. [source, source]

• Concentration at the top of the S&P 500 is unprecedented. The top ten companies now command roughly 36–40% of the index. The 30-year average is closer to 25%, with prior cycles bottoming near 17%. [source, source]

From that standpoint, I’d probably be holding a lot of cash right now, were it not for one large complication.

The Currency-Debasement Problem

The stock market is quoted in nominal terms. When you short, or sit in cash, in an era where the dollar is being debased and where CPI is likely understating inflation in the real categories consumers actually face, time is not on your side. The market in nominal terms could be meaningfully higher five or ten years from now simply because the unit of measurement itself is shrinking. That’s a tricky call.

I do believe we’re going to see a lot of volatility from here. A 50% drawdown is something I would not rule out. But if you just hold cash at a low yield, you really have to stay in very short-duration instruments, because you do not want to go out on the curve given what’s happening with inflation. You are then holding cash that’s yielding less than true inflation. Think about your own cost of living. It isn’t going up at 3% a year. In my own case, I see high-single-digit increases across the categories that matter. Across technology and digital goods you can argue we’re getting deflation, and I’d agree, but we cannot live on technology alone. The things we actually have to buy are getting more expensive, and that’s being felt most acutely by consumers on the lower end.

So the question isn’t really “Cash or stocks?” It’s: where in the equity market are you being paid to take risk?

Where I Am Actually Deploying Capital

I am finding value in areas that are inflation-protected and neglected (or even hated) by investors at the moment.

Read more

The emergence of Generative AI fundamentally fractures the economic logic of this digital hegemony. By collapsing the cost of digital creation of code, content, and even intellectual property, AI is commoditizing the very assets the market is valuing most highly.

We stand at the cusp of a major capital regime change. As the digital world begins to drown in a deflationary supply shock of its own making, the three-dimensional world of atoms, energy, metals, and infrastructure is entering an era of scarcity. The “playbook of the last thirty years is broken,” according to Will Thomson of Massif Capital.

2026 may be retrospectively identified as the pivot point where antifragility shifted from the owner of the software company to the owner of the copper mine. What follows is an examination of this impending bifurcation and the case for a strategic reallocation into real assets.

Subscribe now

A Tale of Two Worlds

AI is a tech marvel reshaping human productivity. But while investors bid up the creators of this digital intelligence, a more profound shift is underway. The global economy is approaching a stark divergence between two distinct ecosystems: the “2D World” of screens, software, and bits, and the “3D World” of energy, infrastructure, and atoms.

For the past fifteen years, the 2D World has been the only game in town. It offered zero marginal costs and infinite scale. A piece of software code can be replicated a million times without spending a penny on steel, fuel, or shipping. The economics of “infinite margin” rightfully seduced the financial world, ultimately leading to unprecedented concentration in the S&P 500.

The rise of Generative AI is a double-edged sword for this 2D dominance. By democratizing code and content creation, AI is unleashing a near-infinite supply shock upon the digital economy. It is commoditizing the very things that the market has been trained to value most. When an AI agent can write code better and faster than a human, the value of a “SaaS wrapper” collapses. When an AI can generate text, images, and video instantly, the value of digital media inventory approaches zero.

In contrast, the 3D World is facing a supply cliff. As Will Thomson points out, one cannot “prompt” a new copper mine into existence. One cannot “generate” a gigawatt of baseload power with an algorithm. The 3D world is governed by the laws of physics, geology, and thermodynamics, not Moore’s Law. Overlay that with the political forces regulating the 3D word, and it is clear that the addition of supply in the real world remains a slow, painstaking process.

Exhibit 1: Key Differences — 2D versus 3D

The Prevailing Bias

For many investors, the psychological difficulty of this necessary rotation cannot be overstated. The market has been conditioned by more than a decade of “Tech Exceptionalism.” Since the Global Financial Crisis, the prevailing narrative has been that technology is the only sector capable of growth in a low-growth world. This belief has become so entrenched that it is now treated as a law of nature rather than a cyclical phenomenon.

Similarly, the belief in “US Tech Exceptionalism” has driven US outperformance for years. Investors have crowded into the same trade (long US large-cap growth, short commodities and value), creating a “suction pump effect,” as described by Murray Stahl of Horizon Kinetics. This effect pulls capital out of the broader market and concentrates it into a handful of mega-caps, distorting valuations.

In investing “what worked yesterday” is often the enemy of “what will work tomorrow.” The market is pricing much of the 2D world for perfection, essentially assuming that margins will remain high forever, while pricing much of the 3D world for extinction. This is the classic setup for a capital cycle reversal. The Mag 7 recently comprised nearly 35% of the S&P 500 market cap, a concentration that leaves passive investors exposed to a change in regime.

Exhibit 2: The Evolution of S&P 500 Market Cap Concentration

AI as an Inflationary Force

The central insight here is counter-intuitive. Many investors view AI as a purely deflationary force because it lowers the cost of labor and intelligence. While this is true for the 2D economy, AI is an inflationary force for the 3D economy.

The irony of the AI age is that it is voraciously physical. The cloud is made of steel, and the data center runs on electrons. An AI query consumes significantly more energy than a traditional search query. The build-out of AI infrastructure (data centers, grid upgrades, cooling systems) requires massive amounts of copper, aluminum, and baseload power.

AI acts as a siphon: it drains value from the digital layer (by increasing the supply of digital goods) and injects value into the physical layer (by increasing demand for physical resources). We are standing at the precipice of a Great Rotation where capital is destined to flee the deflationary, oversupplied 2D World and flood into the inflationary, scarce 3D World. Choosing to get ahead of the Great Rotation is not a speculative act. Rather, it is defensive.

Some may argue that AI will be deflationary in the physical world as well because it will make commodity extraction and production more efficient. Such an outcome is possible, even likely, but it seems out of reach for at least a decade or longer. It is certainly much more remote than AI collapsing pricing in the 2D World.

The AI Eraser

The investment thesis for software has been bolstered by “high switching costs” and “proprietary code.” Generative AI acts as an eraser for these advantages. Consider the SaaS business model. Valuation multiples of 10x, 20x, or even 50x sales have been justified by the idea that once a customer is acquired, they will pay a subscription with 90+% gross margins, essentially forever. This model relies on the difficulty of coding and/or migrating to a competing product.

AI agents are close to being capable of replicating the functionality of many SaaS products, with agent functionality improving rapidly. If an enterprise can use a generic Large Language Model (or a more optimized derivative) to perform the same analytics, customer service, or data processing tasks for which they previously paid a specialized SaaS vendor, the pricing power of that vendor evaporates. We are already seeing signs of a “pricing collapse” in the software sector, where features that were once paid add-ons are now expected to be free utilities. This goes well beyond the typical deflationary forces in the sector.

The implications for a product such as Blinkist, which provides book summaries, may be undeniable, but those same implications also apply — with a modicum of imagination — to a much broader array of software and content.

Exhibit 3: Recent Examples of 2D Pricing Pressure

Exploding Supply and Margin Compression

In the 2D world, supply is quickly becoming near-infinite.

• Content: The cost to produce marketing copy is becoming negligible.

• Code: The cost to produce an application is trending toward zero.

• Analysis: The cost to synthesize complex data is plummeting.

When supply becomes abundant, price collapses. This is the deflationary tsunami facing the 2D world. The moats that investors thought existed are drying up because AI can replicate them quickly and cheaply.

Even in Vertical SaaS, a segment many investors have considered immune, the threat appears acute. The consensus has been that companies providing specialized software for niche industries like construction, legal services, or hospitality are immune to the pressures facing generalist tech. Vertical SaaS products are considered “hard-wired” into customers’ fragmented legacy systems, creating high switching costs. In addition, Vertical SaaS sales teams understand the specific nuances of a customer’s industry, acting not only as vendors but consultants.

Yet, in a world of infinite 2D supply, those are not sustainable moats. AI changes the physics of data integration, as an AI agent does not need a brittle, hard-coded API to talk to a legacy database; it can interpret the data schema, read the documentation, and write the integration code. Or, more radically, it can simply ingest the unstructured data from the legacy system and restructure it on the fly. When an AI agent can serve as a fluid, instant bridge between any two systems, the value of the Vertical SaaS incumbent’s proprietary “plumbing” falls toward zero.

Similarly, the solution-based sales moat is crumbling. In the AI era, the customer no longer needs a vendor to dictate best practices. A construction manager or a law firm partner can prompt an AI model to “design a workflow for managing Tier-2 supplier contracts” or “optimize the patient intake schedule.” The information asymmetry that allowed Vertical SaaS sales teams to charge a premium for “solutions” is disappearing. Even if we are rightfully skeptical of claims that today’s AI can do all of the foregoing, time is decidedly not on the skeptics’ side. What multiple should we put on a business whose moat might be gone in five to ten years?

A Combustible Cocktail: Tech Exceptionalism X Passive Investing

Murray Stahl has poignantly identified the structural risks from an investor’s perspective. The IT sector, inclusive of Amazon, Meta, and Alphabet, recently comprised ~46% of the S&P 500 market value (see page 3 of Horizon’s Q3 letter). If AI makes “intelligence” a commodity, IT sector margins will inevitably compress. The 2D world is entering a war of attrition where the only winner is the consumer, not the shareholder (and even the point about the consumer is debatable due to the quality-of-life implications of the addictive nature of screens and the non-negligible existential threat of AI).

Exhibit 4. Market Cap of Largest US Stock Relative to 75th-Percentile Stock (x) (left axis); Weight of Top 10 Stocks in S&P 500 (%) (right axis)

The mechanism of tech overvaluation is the rise of passive investing. Fund manager Daniel Gladiš discussed eloquently the perils of indexing in a recent interview.

A timely example of the value-destructing reality of passive investing is the impending addition of Carvana to the S&P 500 Index. The shares jumped nearly 10% on the announcement, having already risen 113% year-to-date and witnessed relentless insider selling throughout 2025, in part at materially lower prices. Surely, this is an opportune moment for pension funds and endowments to create value for their beneficiaries by acquiring Carvana shares.

The passive investing trend has created a “suction pump,” according to Murray Stahl, where capital flows blindly into the largest-capitalized companies regardless of valuation. This creates a feedback loop: the higher the stock price goes, the more weight it gets in the index, and the more capital it attracts. This has divorced the price of 2D assets from their fundamental economic reality.

This mechanism also works in reverse. If the underlying fundamentals of the 2D world deteriorate, the “suction pump” will reverse, leading to a liquidity crisis in the very stocks everyone owns. There is “nowhere to hide” in the index when the dominant sector corrects.

Exhibit 5. S&P 500 Sector Weightings

Let’s not forget that the 2D world is constrained by a hard physical limit: human attention. There are only 24 hours in a day. As AI explodes the volume of digital content, the value of an incremental unit of digital engagement approaches zero. We are drowning in “digital noise.” This degrades the value of digital advertising, the lifeblood of the 2D economy. Sure, the creators and companies able to capture scant attention will win, but how do we find them? Nearly every highly valued public company seems like a future winner, but the underlying math does not favor such a conclusion.

Scarcity Premium: The Renaissance of 3D

While the 2D world drowns in abundance, 3D remains constrained. The physical domain is defined by friction, and that friction is now the source of value.

The 3D world has been in a bear market for most of the last decade. Capital has fled the sector since the commodity peak in 2011, resulting in massive underinvestment in productive capacity. As Will Thomson notes, we are moving from an era of globalization and abundant resources to an era of “unwinding” interdependency and resource nationalism.

The “Intelligent Investing 3D” thesis is bolstered by capital cycle theory, popularized by Marathon Asset Management. The theory states that high returns attract capital, which leads to oversupply and lower returns. Conversely, low returns repel capital, leading to undersupply and higher returns.

The 2D world appears to be at the peak of the capital cycle, while the 3D world is at the trough. For a decade, mining and energy companies have been punished by ESG mandates and shareholder demands for dividends over growth. They stopped exploring and stopped building. Bank credit dried up, as pointed out by Mohnish Pabrai in reference to the coal sector.

Meanwhile, oil wells and copper mines deplete every year. If you do not reinvest, supply shrinks. As a result, we have a coiled spring of potential returns. The world has consumed the surplus capacity of the 3D world, and now, just as demand is inflecting upward (driven by AI), supply is hitting a wall.

Goehring & Rozencwajg make a nuanced distinction in the commodity landscape between cyclical weakness and structural strength. In their recent commentary, they split the commodity complex into two camps: (A) Uranium, gold, and platinum group metals have remained “stubbornly immune” to economic malaise because their supply deficits are so acute that they override cyclical demand fears; and (B) oil and base metals, which have reacted to fears of recession and tariff-induced trade wars.

This split may present an entry opportunity. The “Camp B” assets are being priced for a recession or worse, while the gloominess may be overshadowed by secular demand from the AI buildout. The “Next Inflationary Surge,” according to G&R, will lift all boats in the 3D sector, but the entry point today is most attractive in the unloved cyclicals.

It is also instructive to look at the “Return on Invested Energy.” As geology degrades, it takes more energy to get resources out of the ground. We are moving from the “easy oil” and “high-grade copper” era to a period where resource extraction is more energy-intensive and capital-intensive. This raises the floor price for all commodities. The 3D world is not just scarce; it is becoming structurally more expensive to access, granting immense pricing power to the incumbents who own the high-quality assets. Exploration and production, aided by AI, may be an offset in the future, but such a future appears quite remote at this point.

The Great Convergence: Why 2D Needs 3D

The most critical misunderstanding in the market today is the belief that the 2D economy is independent of the 3D economy. In reality, the 2D world is parasitic on the 3D world. It seeks to consume rapaciously the energy and materials of the physical world. As we move from “traditional computing” to “AI computing,” the physical toll of the digital world is skyrocketing.

• Energy intensity: Training a single large AI model consumes gigawatt-hours of electricity.

• Inference costs: Every time a user asks a frontier model a question, a server farm spins up, consuming electrons and generating heat.

• Data centers: These are not virtual entities. They are massive industrial warehouses packed with silicon, copper, and steel, requiring water for cooling and backup diesel generators for reliability.

The bottleneck for the AI revolution is not Nvidia chips; it is the copper wire and the power plant. According to data synthesized from Massif Capital and Goldman Sachs, an AI hyperscale data center is estimated to be four times more copper-intensive than a traditional data center. A facility may require between 3,000 to 5,000 tonnes of copper for power distribution, grounding, and connection.

Goldman Sachs projects that AI-driven demand is pushing the copper market into an imminent structural deficit. Fast-forward ten years and the global supply gap could swell to six million tonnes, a shortfall that is mathematically impossible to fill with currently commissioned mine projects.

Exhibit 6: The Long Journey of Bringing New Copper Supply to Market

We are seeing a shift from demand constraints (where we worried about the economy slowing down) to supply constraints (where the economy literally cannot grow because we lack the physical inputs). In this environment, pricing power shifts to the provider of the bottleneck. If Microsoft needs copper to keep its $100+ billion AI investment running, it will pay a higher price. Demand is mostly inelastic.

The same logic applies to energy. The AI buildout is colliding with a grid that is already strained by the “green transition” and the electrification of vehicles. We are asking the grid to do two impossible things at once: retire dispatchable fossil fuel power (coal/gas) and add massive new load from AI data centers.

This leads to the “Greenflation” and “Tech-flation” thesis championed by Will Thomson. The demand for reliable, baseload power (nuclear, natural gas, hydro) is set to explode. The 2D world’s need for 99.999% uptime cannot be met by intermittent wind and solar alone. This will drive a renaissance in “old economy” energy sources that can provide the stability AI requires.

As noted by Massif Capital, we are already seeing the “bright spot” in European industrials related to power grids and energy infrastructure. This is the leading edge of the Great Convergence: the digital world paying a premium to the physical world.

Historic Valuation Disconnect

The divergence in valuation between the 2D and 3D worlds offers the most asymmetric opportunity set seen since the dot-com bubble burst.

Investors are paying ~20x sales for Snowflake, a business that consumes capital rather than returning it, and whose core value proposition is under threat from commoditized AI analytics. Even mature giants like Salesforce trade at multiples that imply their moats are sustainable. Conversely, the market is pricing Freeport-McMoRan as if copper is a stagnant industry. Freeport owns the Grasberg mine, a geological anomaly that cannot be replaced.

Exhibit 7: A Few Examples of the Disconnect

Bob Robotti offers a crucial correction to the standard “value rotation” narrative. He argues that we are poised to writness not merely the “Revenge of the Old Economy,” which implies a temporary, cyclical bounce. Rather, it is the “Metamorphosis of the Old Economy”. Bob shared his views on the resurgence of North American cyclicals and commodity-based businesses at Latticework 2025.

Industrial companies have spent the last decade consolidating, cutting costs, and cleaning up their balance sheets. They are no longer the bloated giants of 2011; they are lean, cash-generating machines. They have transitioned from being “disadvantaged” by globalization (which favored cheap labor and outsourcing) to being “advantaged” by near-shoring (which favors domestic capacity and reliability). They are trading near liquidation values despite being the essential engines of the future economy.

Resilience and Antifragility

In a deglobalizing world, the physical location of assets matters. We are entering an era of “friend-shoring” and resource nationalism. Access to physical resources is no longer just a commodity trade; it is a matter of national security.

Will Thomson notes that the next decade will be dominated by the “use of trade as a tool of government statecraft”. Owning a copper mine in Arizona is fundamentally different from owning one in a hostile jurisdiction. 3D assets within secure borders are “antifragile”; they gain value during geopolitical chaos because they represent security of supply.

By comparison, 2D businesses are vulnerable to platform shifts (e.g., Apple changing privacy settings), algorithm changes (Google Search updates), and cyber warfare. A software company can lose much of its value overnight if a competitor releases a better AI agent. A copper mine cannot be “disrupted” in the same way. The rock is still in the ground, and the world still needs it.

Then there’s inflation. Murray Stahl provides the critical nuance here: even if official CPI looks tame, “localized inflation“ in scarce assets is rampant. Stahl uses the example of the Sabine Royalty Trust (SBR), an oil royalty company (a hard asset proxy) that simply collects a percentage of the top line. As oil prices rise, the royalty check rises without a commensurate increase in costs.

3D assets are natural hedges against the debasement of fiat currency. If the dollar loses value, the copper mine (priced in dollars) gains nominal value. 2D assets, which rely on long-duration cash flows discounted back to the present, are in a bind if long-term interest rates rise due to inflation while pricing power vanishes due to AI.

The 2026 Pivot

The central thesis of this essay has been playing out already to some extent. However, we are still in the very early innings.

It would not surprise me if 2026 turns out to be the first year of “The Great Rotation.” The initial hype of AI appears like to settle. The market may recognize that the 2D world is fighting a deflationary war of attrition, while the 3D world is enjoying pricing power born of structural scarcity. The narrative may shift from “software eats the world” to “software needs the world.”

For investors and allocators, waiting to pivot is the more speculative choice. Taking advantage of the still-wide valuation disparity to move capital from 2D to 3D assets at a high “conversion rate” strikes me as not only the defensive but also the forward-looking choice.

Let’s move from concept to execution with The Great Rotation Playbook. Start with the actionable ideas laid out in Part 1: The Energy Backbone.

Featured Events

• Ideaweek 2026 (FULLY BOOKED), St. Moritz (Jan. 26-29, 2026)

• Best Ideas Omaha 2026, Omaha, Nebraska (May 1, 2026)

• The Zurich Project 2026, Zurich, Switzerland (Jun. 2-4, 2026)

• Latticework 2026, Chicago, Illinois (Nov. 10-11, 2026)

The content of this website is not an offer to sell or the solicitation of an offer to buy any security. The content is distributed for informational purposes only and should not be construed as investment advice or a recommendation to sell or buy any security or other investment, or undertake any investment strategy. There are no warranties, expressed or implied, as to the accuracy, completeness, or results obtained from any information set forth on this website. BeyondProxy’s officers, directors, employees, and/or contributing authors may have positions in and may, from time to time, make purchases or sales of the securities or other investments discussed or evaluated herein.

This AI-driven capex frenzy is eerily similar to the telecom bubble of the late 1990s and early 2000s. Extravagant spending on fiber optics and network infrastructure promised growth but delivered catastrophic oversupply and collapsing prices. Today’s hyperscalers may be repeating history’s costly mistakes.

Capital cycle theory suggests that this massive investment wave, rather than securing dominance, could become a value-destructive exercise. With hundreds of billions flooding into AI infrastructure, competition is poised to intensify rather than diminish. High costs threaten to compress margins, as driving high utilization of fixed-cost infrastructure becomes a priority.

While free cash flows are already contracting, investors still view the spending generously as “growth capex” in the service of future market dominance. Such willingness may wane if a more sobering picture of the competitive landscape and prospective returns on capital begins to emerge.

Ironically, the ultimate winners of the AI arms race may not be the hyperscalers themselves, but nimble, asset-light companies riding atop their infrastructure—much like Netflix capitalized on telecom and cable’s broadband investments. These smaller, specialized firms could capture disproportionate value by creating innovative AI services without incurring the heavy capital burden plaguing the larger players.

Finally, the less glamorous, asset-heavy sectors providing critical commodities and industrial products—for power generation, electrical wiring, cooling, and more—stand to reap substantial benefits from the hyperscalers’ capital deployment. Yet, sector leaders remain available at below-market multiples.

Investors should consider the risk-reward: as the hyperscalers aggressively wager their enviable business models on an uncertain future, a “new generation” of more agile players and an “old generation” of commodity-based businesses may end up capturing a lion’s share of the economic value created.

Subscribe now

The Great AI Capex Surge

Not since the dot-com buildout of the late 1990s have we seen capital spending on tech infrastructure at the scale now unfolding. Over the last two years, the largest tech companies have dramatically loosened their purse strings to fund an AI computing arms race. Data centers packed with AI chips don’t come cheap, and the numbers are staggering.

In 2024 alone, the five biggest hyperscalers poured an estimated $197 billion into AI-related infrastructure. That marks an inflection from prior years, up more than 50% from roughly $126 billion in 2023. And the trajectory points steeply upward: J.P. Morgan Asset Management projects these companies’ combined capital outlays will reach $234 billion in 2025 and $249 billion in 2026. The hyperscalers are now investing in one year what they used to spend over a decade not long ago.

Capex by AI Hyperscalers: Alphabet, Amazon, Meta, Microsoft, Oracle ($ billions)

Training large language models and running AI services at scale require vast computing and storage. A few notables from recent disclosures:

• Amazon plans to spend $100+ billion on capital projects in 2025, up from about $75 billion in 2024 and roughly $50 billion in 2023. Andy Jassy defends the splurge as seizing a “once-in-a-lifetime” opportunity.

• Microsoft is projected to spend about $80 billion on AI data center infrastructure (excluding its OpenAI partnership) in 2025, roughly double its recent capex. Management has indicated a multi-year investment cycle to build out Azure’s AI computing capacity. According to Satya Nadella, Microsoft is “all-in” on AI.

• Alphabet reportedly plans around $75 billion in capex for 2025 , above analysts’ prior expectations of ~$58 billion. Sundar Pichai defends the spending as critical to “build infrastructure for future applications.” However, Google’s high-profile AI products have yet to prove they can monetize at scale.

• Meta’s capex is projected at $60–65 billion in 2025, an increase from roughly $40 billion in 2024. At $60+ billion, Meta’s capex would be roughly one-third of its 2024 revenue, suggesting how capital-intensive the business is becoming.

• Oracle has also jumped on the AI infrastructure bandwagon, partnering with NVIDIA and others to offer cloud GPU capacity. Heavy capex in 2025 and beyond follows capex in the high single-digit billions in 2024.

• SoftBank has emerged as a wild card. In January, OpenAI, SoftBank, and Oracle a new joint venture dubbed “Project Stargate,” aiming to invest up to $500 billion over four years to build AI data centers across the U.S. The initial phase is $100 billion, with SoftBank and OpenAI each committing $19 billion as lead partners. Speaking at an event in late 2024, Masayoshi Son envisioned an artificial superintelligence by 2035 that would require “400 GW of data center power” and 200 million AI chips — at a cumulative capex cost of $9 trillion. Yes, trillion with a “t”. He argued that $9 trillion could be “very reasonable” and even remarked that $9 trillion might be “too small” an investment in the grand scheme. Such statements reflect a mindset that no sum is too great in pursuit of AI dominance.

• Apple, Alibaba, Tencent, and other major players are not even considered here, but they will undoubtedly play a role in — and add to — the massive spending.

Taken together, the hyperscalers’ AI capex binge is without precedent. By 2025, just four companies are on track to spend over $320 billion combined on capex. That’s larger than the GDP of countries like Chile or Finland. It also represents a large portion of these firms’ revenues — on average over 20% of sales in 2025, roughly double the percentage from a decade ago.

Big Tech is quickly moving from asset-light to asset-heavy. AI is seen as vital to remain competitive, and not investing aggressively could mean ceding ground to rivals or upstarts. The hard questions: What is the timeline for payback? How much of this spend is truly “growth capex” that will generate high returns, versus just keeping up with the Joneses? The history of tech and infrastructure booms is littered with examples of overinvestment.

Asset-Light to Capital-Heavy

For years, the hyperscalers were paragons of the asset-light, high-margin business model. They demonstrated how scaling software and networks could generate enormous revenues with relatively limited physical assets.

The “AI infrastructure” business is fundamentally more capital-intensive than the pure software or advertising businesses that drove past growth. As hyperscalers invest tens or even hundreds of billions in steel and silicon, they start to exhibit traits of heavy industries or utilities.

The most immediate effect of the AI build-out is the pressure on free cash flow. The hyperscalers are still growing revenue healthily, but capex is growing much faster. According to Morgan Stanley, the hyperscalers’ combined FCF actually fell year-over-year in the second half of 2024, and they project FCF to drop 15–20% in 2025 despite revenue growth. Analysts across the board have been revising down FCF forecasts for 2024–26 by on the order of 20+% to reflect the higher spend.

Meta in 2024 generated a record ~$52 billion in FCF. However, with capex set to jump from ~$40 billion in 2024 to potentially $60+ billion in 2025, even if Meta’s operating cash flow grows, FCF could be nearly halved. Similarly, Amazon’s FCF, which was already thin relative to earnings because of heavy fulfillment and AWS investments, could be consumed entirely by AI-related capex.

The income statement will reflect this over time through higher depreciation. Cloud providers already depreciate servers over short lifespans. As capex front-loads costs on the balance sheet, operating margins will feel the drag of accelerated depreciation. Of course, company managements and sell-side analysts will then steer investor attention away from reported earnings toward cash earnings, just as they currently suggest investors look at cash earnings but exclude growth capex. And throughout, investors should pay scant attention to stock-based compensation.

Financial reporting gymnastics aside, the financial profile of hyperscalers may start looking less like the high-flying 2010s and more like that of mature, capital-intensive businesses where cash flows are reinvested as fast as they come in. This could lead to a pivotal change in the investment narrative, especially if we enter a period of greater market volatility and lower investor tolerance for pie-in-the-sky thinking.

From “Scalable” to “Saturated”?

The market historically afforded hyperscalers premium valuation multiples because of their scalable models and wide moats. A critical question: does the move to capital intensiveness undermine those moats?

Morgan Stanley highlights that hundreds of billions have been — or are about to be — plowed into AI by hyperscalers and VCs, yet recent developments like open-source AI models challenge the idea that only a few big firms will reap the rewards. DeepSeek reportedly achieved GPT-4 level performance with far fewer parameters and less hardware, calling into question how much data center muscle will ultimately be needed for advanced AI. If AI innovation allows more to be done with less, the hyperscalers may find they overspent on capacity that isn’t utilized at pricing levels assumed. In such scenarios, the competitive advantage one hoped to secure by massive spend dissipates — a classic capital cycle pitfall.

If investors come to view these companies as quasi-utilities or heavy industry players, we could see a de-rating. For instance, utility companies or telecoms with large asset bases but moderate growth often trade at single-digit to low-teens P/E ratios. At the peak of the 2000 tech bubble, telecom equipment makers like Cisco were valued like high-growth stars; a few years later, they traded more like cyclical industrials.

Another aspect is capital returns. Asset-light tech companies generate copious excess cash, much of which has been returned via buybacks. But if cash is tied up in capex, fewer buybacks occur, potentially reducing EPS growth.

Big Tech may effectively “nationalize” a chunk of its business model — pouring shareholder money into infrastructure in a way akin to state utilities, but without state-like guaranteed returns.

The Telecom Boom and Bust

“Build it and they will come.” That was the mantra of telecom entrepreneurs in the late 1990s, much as today’s tech giants seem to believe “build it and AI will come.” The telecom frenzy, which saw companies lay tens of millions of miles of fiber optic cable and construct vast networks, holds lessons for the AI infrastructure race. Back then, as now, the lure was exponential demand and the fear of missing out on a paradigm shift. And back then, as may happen now, reality fell far short of hype, leaving a glut of capacity, crushed prices, and investor carnage.

To be clear, I am a firm believer in the utility and strong growth of AI. The challenge as an investor is to figure out where my capital will be treated best. It is quite easy to get blinded by the promise of returns. Examples abound: automobiles (revolutionary as compared to horse-drawn carriages), airlines (revolutionary as compared to other forms of long-distance travel), mobile phone service (revolutionary as compared to old-school telephony), and other hugely beneficial innovations.

Following the U.S. Telecom Act of 1996, dozens of carriers and network builders sprang up. Incumbents and startups raced to install fiber-optic networks, convinced that an explosive growth in internet usage would require vastly more bandwidth. Companies like Level 3, Qwest, Global Crossing, 360networks, Williams, and WorldCom raised and spent hundreds of billions of dollars in aggregate. WorldCom’s CEO claimed in 1996-97 that Internet traffic was doubling every 100 days — an exaggeration — but it fueled a frenzy of investment as firms feared falling behind.

Global Crossing — Stock Price, 1998-2002

A poster child was Global Crossing, founded in 1997 with a bold plan to lay undersea fiber linking continents. In just eighteen months, Global Crossing went public and saw its valuation skyrocket to $47 billion by 1999. It epitomized the era’s hubris. By 2001, the tides turned. Demand did grow, but not nearly to the level the capacity builders expected. By late 2001, Global Crossing was hemorrhaging cash — it lost $3.4 billion in Q4 2001 on $793 million of revenue. The overbuild was so severe that much of its fiber lay dark. In early 2002, Global Crossing filed for bankruptcy — one of the largest bankruptcies at the time, wiping out equity holders.

The entire sector imploded. By the mid-2000s, around 85% of the fiber optic cable laid in the late 90s remained unused. Companies had installed far more capacity than the market needed. Even four years after the dot-com bust, the vast majority of strands were unlit. Bandwidth prices plummeted ~90% in the early 2000s. This is basic economics: when supply overwhelms demand, prices crash. The value of network capacity fell to a fraction of what companies had projected, ruining their business cases. Contracts that were expected to yield rich returns became nearly worthless.

A “trillion-dollar telecom boom” turned into a trillion-dollar bust. From the market peak, about $2 trillion of market cap was erased in telecom stocks, contributing heavily to the broader market downturn. That $2 trillion figure would correspond to a much larger figure in today’s market.

Telecom Boom Darlings — Before and After

Two dozen telecom companies went bankrupt in 2001-2002 alone. WorldCom, led by Bernie Ebbers, had grown via acquisitions and network builds, only to be felled by both accounting fraud and a wildly overbuilt network carrying price-depressed traffic. The industry accumulated over $1 trillion in debt by 2002, much of which had to be written off. As FCC Chairman Michael Powell noted then, a huge portion of the debt incurred to fund these networks would “never be repaid”.

The Price of Overspending: Telecom Stock Performance vs. Mobile Adoption

This was a classic case of overestimation of demand and underestimation of competition, leading to gross capital misallocation. Each player assumed if they built a superior network, they’d capture the lion’s share of surging internet traffic. But everyone assumed that, so everyone built, leading to massive oversupply. In capital cycle terms, high anticipated returns attracted too much capital, which destroyed those very returns.

The parallels to today’s AI infrastructure race are striking — and a bit eerie:

• Exponential demand forecasts: Then it was internet traffic; now it’s AI model demand. In the 90s, many genuinely believed every home would soon need massive bandwidth, and that businesses would require unlimited data — true in the long run, but the timing was off by a decade or more. Could AI demand likewise take longer to materialize than the rosy projections suggest?

• Cheap capital facilitating overbuild: The late 90s featured easy capital — IPOs, bond offerings, and loans were plentiful for any company with a fiber story. This led to less discipline in investment decisions — exactly what capital cycle theory warns. SoftBank’s ability to marshal $100 billion for “Stargate” and the hyperscalers’ war chests are analogous to the glut of funding telcos had in 1999.

• New entrants and competition: In telecom, deregulation opened the field to many new competitors. In AI cloud, while the hyperscalers are few, we do see others — e.g., Oracle partnering with NVIDIA, startups like Cerebras building specialized AI supercomputers, not to mention China’s tech giants. We might witness a price war in AI cloud services reminiscent of bandwidth price collapses. Already, some cloud providers are differentiating on price per GPU-hour or offering discounts to lock in AI customers.

• Speculative excess and reality check: Telecom’s motto was “if we build the network, uses for it will come.” Killer applications did ultimately arrive, but not fast enough to save the over-investors. One warning sign: Goldman Sachs points out that while Big Tech has ramped capex to fuel generative AI, they “have yet to show sustainable business models” for these investments.

• Hubris and overconfidence: The late-90s telco executives were brimming with confidence — they were building the future’s highways, and nothing seemed too expensive or too large-scale. Mark Zuckerberg declaring a “year of efficiency” in early 2023 has quickly shifted to effectively a year of inefficiency as he doubled capex for AI. That cognitive dissonance — needing to invest heavily to not be left behind, even if it hurts financial efficiency — is reminiscent of prior attitudes.

The survivors of the telecom bust did lay a foundation for future growth. The fact that 85% of fiber was dark in 2005 meant that companies like Netflix and Google in 2010 could stream video because bandwidth was abundant and dirt cheap. The infrastructure did get used — just under new ownership at pennies on the dollar. Infrastructure investments often pave the way for someone else to profit.

From 2000-2002, telecom stocks lost on average ~80-90% of their value — a wipeout rarely seen outside depressions. I do not anticipate a similar cataclysm for Big Tech, but even a partial parallel is worrisome and not reflected in the consensus.

Infrastructure vs. Over-the-Top: Who Captures the Value?

A pattern repeated throughout modern economic history is that those who provide fundamental infrastructure often struggle to capture as much value as those who build services on top of that infrastructure. Railroads vs. those who shipped goods on them, telegraph wires vs. financiers arbitraging information, telephone networks vs. call services — the list goes on.

In the 2000s, the term “over-the-top” arose to describe services that ride on telecom/cable networks without owning them. The OTT providers have often created disproportionate shareholder value relative to the infrastructure providers. I see a potential analog in the AI landscape: the hyperscalers are building the “rails” and “pipes” of AI, but the biggest value creation might come from clever, asset-light companies that use those rails to deliver AI-driven products and services.

Netflix vs. the Cable/Fiber Companies

In the late 90s and 2000s, telecom and cable firms invested massively in broadband networks. By mid-2010s, high-speed internet had become ubiquitous in developed markets, thanks largely to those investments. However, the investors in many of those networks had meager returns. Telecom companies like Verizon, AT&T, Comcast, and their global peers often traded at low multiples, burdened by debt and heavy ongoing capex for maintenance and upgrades.

Netflix, founded in 1997, pivoted to streaming video around 2007. The company did not build any broadband network; it simply leveraged the internet connections its customers already had. By investing in content and a user-friendly platform, Netflix rode on top of the cable/fiber infrastructure. Netflix is now worth more than any individual telecom or cable company. The OTT player captured the value while the network owners were often relegated to lower-margin “dumb pipe” roles.

European telecom CEOs argued that companies like Netflix, YouTube, and Facebook were “free-riding” on their networks — using bandwidth that cost billions to deploy, without paying network operators beyond basic transit fees. Calls for “fair share” payments emerged. But regulators largely didn’t force OTTs to pay extra; instead, telcos had to be content with selling bandwidth as a commodity. The investment risk was borne by the network builders; the rewards reaped by application builders.

YouTube became a massive video platform without Google laying a single cable to homes — it relied on broadband networks built by others. WhatsApp and Skype essentially displaced a huge portion of telcos’ high-margin SMS and international call revenue by using the internet as a free transmission medium. The telcos provided the connectivity, but third parties claimed the profit pools.

The pattern extends beyond telecom. In computing, the “Wintel” era saw Microsoft and Intel capture most of the profits of the PC revolution, not the manufacturers of hard drives or memory chips. In cloud computing, ironically, Amazon AWS was initially an OTT service riding on data centers and connectivity provided by others — though Amazon quickly vertically integrated by building its own data centers.

AWS Capex vs. AI Infrastructure Capex

Some may argue the example of AWS serves as proof that a high-capex business can create massive value. However, Amazon’s capital spending on AWS fundamentally differs from its and other hyperscalers’ massive investments in AI infrastructure.

AWS had an inherently clearer and lower-risk business case from inception because Amazon itself was a guaranteed customer. AWS was launched as a capital-light experiment leveraging Amazon’s existing retail infrastructure. Early capex was modest, amounting to tens of millions per quarter, and Amazon gradually expanded AWS only as real customer demand materialized. By the time AWS publicly launched services like S3 in 2006, thousands of developers were already signed up, providing immediate revenue and justifying further investments.

In sharp contrast, today’s AI infrastructure involves enormous upfront capital commitments. Collectively, Amazon, Microsoft, Google, and Meta are projected to spend over $300 billion in a single year on AI infrastructure. This level of investment vastly surpasses AWS’s gradual, demand-driven approach, representing a speculative build-out based largely on anticipated rather than proven revenue streams.

Moreover, AWS scaled in a relatively benign competitive environment. In its early days, AWS had limited competition, achieving substantial market share before rivals like Microsoft Azure and Google Cloud gained traction. This allowed AWS to build a dominant position and maintain pricing power. The AI infrastructure market is intensely competitive and global from the outset.

Asset-Light Innovators in the AI Era

Who could be the “Netflix of AI” — the agile, capital-light companies that leverage the hyperscalers’ AI infrastructure to deliver high-value solutions?

A few possibilities:

• AI-powered software startups, which do not train their own giant models from scratch; instead, they use APIs from OpenAI, Azure, AWS, or fine-tune open-source models on rented cloud infrastructure. They pay usage fees; the multiple on those costs could be high if they deliver unique value to end-users.

• OpenAI itself began as an asset-light layer on top of Microsoft’s Azure. When ChatGPT took off in 2023, it was running on Microsoft’s data centers. OpenAI’s value skyrocketed, despite OpenAI not owning the vast majority of the hardware. It has since involved more complicated partnerships.

• Consumer AI apps and platforms may scale to hundreds of millions of users without owning hardware. Think of an AI-driven social network, or a personal AI assistant app that becomes ubiquitous.

• Integration and consulting offer another asset-light opportunity. Companies that become experts at weaving AI services into business workflows could add lots of value. Their main assets are human capital and perhaps some proprietary data.

• Open-source ecosystem: Hugging Face, a platform for sharing AI models, has become a critical hub for AI developers. By facilitating model distribution and community building, Hugging Face’s value lies in its network, not hardware.

In essence, the OTT opportunity is any business that leverages the plentiful AI computing power being built, without sinking capital into it, to deliver something uniquely useful or creative. I would not be surprised if some of the most valuable AI companies end up having relatively low fixed-asset bases. They might monetize data, algorithms, or networks of users rather than compute infrastructure.

Software Incumbents: More to Lose Than to Gain?

Software-as-a-Service giants like Salesforce, Adobe, ServiceNow, and SAP have enjoyed high growth, robust pricing power, and premium valuations, benefiting from hyperscaler infrastructure. However, the AI boom poses serious risks to their dominance. AI represents more than a new feature; it’s a paradigm shift capable of fundamentally altering how software is consumed.

AI lowers barriers to entry, enabling startups to rapidly develop powerful, AI-driven software at reduced costs. The result is an influx of competitors targeting every niche within SaaS. Hundreds of AI-powered alternatives emerge weekly, flooding markets previously dominated by large incumbents. This surge erodes incumbents’ pricing power, as businesses gain leverage to negotiate lower prices or switch providers.

AI may also commoditize SaaS platforms through AI-powered agents. As they mature, autonomous agents will likely become the primary user interface, relegating SaaS platforms to backend utilities. This shift drastically reduces the value of distinct SaaS brands, user interfaces, and specialized features.

History illustrates how rapidly incumbents can collapse when competition surges dramatically. Groupon, for instance, initially enjoyed explosive growth, yet quickly faced thousands of copycat competitors due to easily replicable technology and low barriers to entry. This intense competition swiftly undermined Groupon’s uniqueness and profitability, causing severe erosion in market share and valuation.

Satya Nadella underscored this existential risk to SaaS in a recent interview by asserting, “the notion that business applications exist, that’s probably where they’ll all collapse in the agent era.” Nadella sees AI agents supplanting SaaS workflows, shifting critical business logic into a dynamic AI layer. He foresees value shifting from standalone SaaS platforms to integrated AI-driven interfaces and automation layers, fundamentally altering the competitive landscape.

See the first four-and-a-half minutes:

Will Hyperscalers Move Up the Stack?

The hyperscalers are trying to play on both levels — providing infrastructure and building AI services. In a sense, they are attempting to be both the railroad and the train, to capture value at multiple layers.

This is wise strategically, but it’s challenging culturally. Focus is a finite resource. The likes of Netflix succeeded because the telcos weren’t great at creating streaming services. Similarly, startups may always have the agility and domain-specific focus to out-innovate a generalized platform on particular use cases.

Capital Cycle Theory: Will Hyperscalers’ AI Investments Destroy Value?

Applying the theory laid out in the seminal book Capital Returns to the hyperscaler AI frenzy yields some sobering conclusions. Over the past decade, Big Tech enjoyed extraordinary profits, which signaled an open invitation: “invest here, the returns are great.” And invest they did — initially in areas like new ventures and R&D — but by 2022-2023, a clear target emerged: Generative AI.

“High profitability loosens capital discipline,” as Marathon Asset Management’s capital cycle framework states — companies with flush profits are inclined to boost capital spending. That’s exactly what happened: cloud profits were reinvested into yet more capacity. And just as the theory predicts, competitors are following suit — no one wants to be left behind, and no CEO wants to tell the Board “we missed the AI boat”. CEO egos and tech FOMO reinforce the cycle: even if internally there are doubts, externally everyone exudes confidence and doubles down.

One marker of potential overinvestment is when capacity growth outpaces demand growth. Are we seeing that? Early signs suggest maybe. By late 2024 cloud GPU shortages seemed to be easing. If all hyperscalers have significantly more GPU clusters by 2025, they will be eager to rent them out — possibly pushing each other on price or sales incentives to win customers. We also see a proliferation of AI chips, which could flood the market with compute options. Capital cycle theory warns that the competitive position of firms deteriorates when everyone has expanded supply.

If indeed this is an overbuild, we will see the next phase as returns on capital decline. Already, hyperscalers’ FCF-to-revenue and ROIC are trending down. The crucial question is whether future profits will rise enough to justify the capex. Capital cycle analysis is inherently skeptical of such happy endings when a lot of capital is chasing a theme. It posits that unless there is some barrier to entry or restraint, the increase in supply will outstrip what demand can profitably absorb. A few scenarios:

• AI compute becomes a commodity service, with thinning margins.

• Some fraction of this expensive capacity sits idle, or is utilized but at such low pricing that ROI is subpar. For example, if open-source efficient models reduce the need for massive GPU hours, customers might use smaller clusters or edge devices to do tasks, stranding some hyperscaler capacity.

• The extreme outcome: Write-downs or restructuring — recall how Intel had to write off manufacturing capacity in past downturns. In cloud, a direct writedown would be unusual since servers can be repurposed, but we might see a slowdown and layoffs in the data center construction, or cancelled projects.

The capital cycle perspective argues that the presence of well-capitalized competitors investing simultaneously virtually assures that capital will be misallocated. Marathon’s approach would have investors look for the opposite: industries where capacity is leaving or consolidating, not entering. Right now, AI infrastructure is clearly in an expansion phase, not consolidation.

Let’s consider barriers and mitigating factors:

• Is there an oligopoly that can rationalize capacity? In some industries, a few players coordinate to avoid glut — think OPEC in oil, or semiconductor foundries scaling more carefully. In cloud AI, while it’s effectively an oligopoly of three to five in each region, they are in hyper competition, not cooperation.

• Technological leapfrogs: A risk of heavy investment is tech obsolescence. Companies might be stuck with last-gen infrastructure if a new breakthrough emerges. If someone figures out an AI algorithm that is far more efficient, suddenly you don’t need as many GPUs or large clusters. DeepSeek’s R1 model is an open-source model that purportedly challenges closed models with less compute. If such approaches proliferate, the scale of compute might become less of an advantage. In capital cycle terms, that would mean the assumption that more capital = more moat fails; instead it becomes sunk cost with no moat.

• Regulation or external shock: A non-economic factor could also derail returns. Suppose regulators in the EU or US impose restrictions on AI deployment or require expensive compliance that reduces profitability. Or perhaps an AI safety issue leads companies to pause certain projects.

The capital cycle may turn as follows: Companies start showing poor returns, stock prices lag, investors get frustrated. Eventually, management teams may pull back on capex. Some weaker players might exit or scale down. Consolidation or sharing could occur — maybe companies collaborate on AI infrastructure to avoid redundancy. As capital investment slows, the oversupply gradually clears with growing demand, and the survivors eventually see better returns again.

We’re not at the bust phase yet — we’re arguably at the peak of enthusiasm heading toward the crest of overcapacity. But investors should be watching for signs of peaking. For example, management rhetoric. Earlier in 2023, tech CEOs were extremely bullish on AI. By early 2025, some tones are shifting to temperance, emphasizing efficiency of models.

A key concept from Capital Returns is that it often pays to invest against the capital cycle. That might mean, for instance, looking at other sectors now unloved because so much capital is going to AI — those other areas could face less competition and yield better returns. Or even within tech, maybe the second-order beneficiaries might be better investments than the hyperscalers during this heavy-build phase.

In AI, the combination of herd behavior, hubris, and cheap capital from prior success is textbook. The trillion-dollar question: will this cycle break the way of telecom or can Big Tech manage a softer landing? Big Tech has much more diversification and financial resilience than the one-trick telecom ponies did. Even if AI capex overshoots, these firms have other profit streams to cushion the blow. That might prevent outright disaster, but it could weigh on valuations.

Feeding the AI Beast: Asset-Heavy Beneficiaries of AI Growth

Asset-heavy, commodity-based businesses meeting AI’s needs are companies that provide the raw inputs and physical necessities for AI data centers — from electricity to cooling systems to critical materials like copper. As AI infrastructure demand soars, these suppliers could experience a secular boost in volume and pricing power.

Electricity

Deloitte estimates data centers will account for 2% of global electricity consumption in 2025 and could double to ~1,065 TWh by 2030 — more electricity than some G20 countries consume annually — largely driven by power-intensive AI growth. If I was a betting man, I would take the over on Deloitte’s 2030 estimate — by a lot.

Electricity is the oxygen of AI, and those selling it will see increased demand. If supply struggles to keep up, electricity prices could rise, benefiting generators’ margins. AI data centers could strain grids, which could be good for power producers’ businesses if managed well.

For electric utilities, particularly in regions with major data center clusters, AI is a source of major demand growth. Utilities that can supply this power reliably stand to benefit from high load factors and long-term purchase agreements with data center operators. Moreover, AI centers often require 24/7 stable power with redundancy. This can translate into investments in not just generation but also grid infrastructure.

One nuance: hyperscalers also invest in renewables and sometimes become their own power providers. But they still typically pay utilities or independent power producers. The regions that become AI hubs might see their local utilities’ revenues surge. For example, in Quincy, Washington or parts of Utah and Oregon, large cloud data centers have become the dominant customers for local power companies.

There’s even talk of innovative solutions like small modular reactors dedicated to powering data centers. If that happens, companies in nuclear tech or other generation tech could find a new customer base in hyperscalers.

Cooling and HVAC

Running tens of thousands of high-performance chips in close quarters produces enormous heat. Traditional air cooling is hitting limits in AI data centers. Advanced cooling solutions are in high demand — from chilled water systems to immersion cooling where servers are dunked in special fluids.

Cooling is not trivial: cooling systems can consume up to 40% of a data center’s electricity themselves. So improving cooling efficiency directly lowers operating costs. Companies that provide industrial chillers, heat exchangers, liquid cooling equipment, fans, and air handlers are seeing a boom. For instance, Vertiv, a company specializing in data center power and cooling, has reported strong orders tied to data center expansion. Others include Schneider Electric, Siemens, Johnson Controls, and niche firms like Rittal or Daikin that have data-center-specific cooling products.

Liquid cooling is gaining traction because it can handle high-density racks better than air. It can also reduce or eliminate the need for chillers in some designs. Companies innovating in this space could become key suppliers. Even large chip makers like Intel have initiatives on new cooling approaches.

Water suppliers also indirectly benefit — many data centers use evaporative cooling which consumes water. As noted, a 30MW data center can use >50 million gallons of water a year for cooling. That has started to raise sustainability concerns, but it’s an opportunity for water treatment and supply companies, or for firms offering waterless cooling tech. Horizon’s Murray Stahl has made the case for an investment idea that may benefit from rising demand for water cooling — San Juan Basin Royalty Trust.

Beyond GPUs: Semiconductor and Hardware Beneficiaries

Leaving aside the obvious — and highly valued — beneficiaries of AI chip demand, it’s clear that AI data centers need power equipment, electrical components, and networking gear. Even seemingly mundane things like server racks, power distribution units, and backup generators are in higher demand. For example, Cummins and Generac produce backup generators for data centers to ensure uptime during grid outages.

Memory manufacturers also benefit: AI workloads are memory-intensive, driving demand for high-end DRAM and SSD storage. They are asset-heavy and have been ramping production of specialized high-bandwidth memory for AI GPUs.

Building data centers requires a lot of construction work and often land in specific areas. Data center REITs own data centers and lease to tenants — though hyperscalers often build their own, some still lease colocation space for certain needs. If the hyperscalers ever slow building in-house, they might rent more, which could help these REITs. Construction firms that specialize in tech infrastructure have robust business building out these facilities.

Copper and Raw Materials

AI data centers use vast amounts of copper in power cabling, connectivity, and even within chip interconnects. Copper is critical for electrical conductivity in these high-power environments. An estimate from Man Institute notes that US data center expansion for AI could add 0.5-1.5% to global copper demand. Again, I would take the over on that, by quite a lot. The copper market is often in tight balance, so even a low single-digit swing can mean the difference between surplus and deficit. A sustained low single-digit increase in demand without commensurate new supply could create a copper shortage and drive up prices.

Copper miners and producers could indirectly ride the AI wave. Copper prices have already been buoyed by the clean energy transition; AI data centers add another source of secular demand.

Similarly, other materials: rare earth elements, specialty plastics and composites, and silicon photonics components. While these commodity suppliers are asset-heavy and often low-margin, a demand surge can improve their utilization and prices, boosting profits. Capital cycle theory would say when everyone else is overinvesting, you might want to own the input providers who haven’t yet overbuilt mines or factories — they might enjoy a period of pricing power.

To illustrate how real these demands are, consider NVIDIA’s switch from fiber optics to copper for short connections. NVIDIA has signaled a move to use more copper cables within AI data centers, due to needing extremely high bandwidth between GPUs. This might seem small, but it may signal a major increase in copper usage. Multiply such effects across power distribution and wiring and the impact is large.

Other Infrastructure

• Data center REITs are effectively the “landlords of AI.” Equinix, Digital Realty, GDS, and similar companies have a global footprint.

• Grid and transformers: As power demand rises, manufacturers of high-capacity transformers benefit — e.g., Siemens, GE Grid, Hitachi Energy.

• HVAC companies for precision environments: Data centers need not just cooling but humidity control, fire suppression. Companies like Halma may benefit.

• Security firms: Physical security of data centers represents a growing need.

• Logistics and installation: Moving and installing thousands of heavy servers and cooling units helps logistics providers.

Building and operating AI facilities is a supply chain spanning mining, manufacturing, construction, and utilities. Some of these “arms dealers” of the AI boom may be attractive investments, especially if they have a moat or tight supply conditions. A utility serving an area of high data center growth could see years of stable, growing cash flows due to long-term power contracts with hyperscalers.

One could also think of beneficiaries in the energy sector beyond electricity — e.g., natural gas suppliers if gas power plants are main source for new electricity, or even backup diesel fuel suppliers for generators. These are more tangential but show how widespread the ripple effects can be.

As investors, we may want to construct a portfolio that underweights the big spenders and overweights the enablers and utilizers of AI. In doing so, we align ourselves with the idea of the “capital cycle” — avoid where capital is flooding in and favor those selling scarce inputs or crafting valuable outputs.

Share

For insights into capital cycle theory, please enjoy my recent segment on the subject.

John Mihaljevic is the managing editor of Latticework and chairman of MOI Global, the research-driven membership organization of intelligent investors. John studied under former Yale CIO David Swensen and served as research associate to Nobel laureate James Tobin. He is the author of The Manual of Ideas and a winner of VIC’s prize for best idea. He holds a BA in Economics, summa cum laude, from Yale and is a CFA charterholder.

That is what this post is about. It is not investment advice. Do your own work.

It does feel overwhelming to write about the impact of AI on stock returns, at least in a way that is actionable by investors. In some respects stock prices are already speaking loudly. The price performance of Nvidia, Palantir, and Tesla says volumes about the value creation potential of AI. These companies, and numerous others, are the obvious beneficiaries. Few would argue to the contrary on the fundamentals; any quibbles tend to focus on the likely winners’ stratospheric market quotations:

I will not try to add value here by opining on the market appraisals of Nvidia, Palantir, or Tesla. I’ll let the market decide what multiples it wants to put on these likely AI winners. That said, there are some ways to think about AI winners’ risk-reward tradeoffs. Otherwise, no difference would seemingly exist between investing $1,000 in Nvidia at 1x or 10x the recent market quotation. I explain my thinking in a broader essay entitled, “Why Inversion is Indispensable for Long-Term Success”.

The Bearish Side of the Coin

Today I focus on the fact that we haven’t really seen Mr. Market price in the flipside of AI — the creative destruction that may be about to wreck the business models of a range of public companies. Call it a feature of the long-running bull market: Investors are much more inclined to look for the winners than the losers. Not only that, but even when staring a likely loser in the face, investors seem willing to accept the loser’s narrative about AI being more of an opportunity than a threat.

In the very short term, the losers may be able to throw some AI bells and whistles on top of their legacy products and charge customers an extra fee, thereby boosting short-term results. Investors would be shrewd not to get lulled into complacency or the belief that a loser may actually be a winner. Instead, the short-term bump may be an opportunity to monetize winnings or even reverse the direction of a trade.

Sometimes investors get too deep into the weeds of a company, a situation, or a theme. Industry insiders are often the last to see the writing on the wall. Think Angelo Mozilo or bank analysts in 2007 (“real estate prices only go up”), or Blockbuster executives in the early days of Netflix (“picking a movie at the store is part of the experience”). Not seeing the forest for the trees applies here. Under such circumstances, we may need the perspective of a knowledgable outsider to push us to step back and consider the big picture.

Enter Adam D’Angelo, founder of Quora and AI startup Poe. Adam has a rich pedigree, having joined Facebook shortly after its launch and served as its chief technology officer. He was an investor in Instagram before its acquisition by Facebook. As a board member of OpenAI, Adam voted to remove Sam Altman as CEO. When Sam returned, the other three board members involved in the ouster resigned, but Adam retained his position.

Here’s his take from a couple of weeks ago:

Adam followed it up with this take:

One of the commenters was more explicit:

If you weren’t triggered before, you might be now: “There’s no way AI-generated movies or series can rival the best Hollywood releases. Humans will always prefer movies with real human actors instead of AI-generated characters. Netflix is a platform, so any quality AI productions will simply come onto the platform. Netflix has a massive scale advantage and can invest in AI ahead of its competitors. AI will expand the market, producing many winners, including Netflix.”

The list goes on, and the arguments put forth in favor of Netflix make sense. Yet, lurking beneath the surface is the uncomfortable truth that more supply — especially if it is near-unlimited, near-free supply — has to manifest somehow. Stark example: air — a highly valuable commodity, yet typically free due to near-unlimited supply.

A more mundane example: Groupon, the first mover and scale leader in online deals. Once supply exploded due to new entrants copying Groupon’s formula, the company had to accept an impairment to its business model, and the stock price followed.

Other examples abound in the business world. In hindsight, the argument is easy to comprehend. Looking ahead, though, our imagination usually fails us, especially if we have developed a positive association with an idea.

As an investor in Netflix, how could you not have a positive association with the company? It has been a massive source of good vibes, an example of the adage to let your winners run. You might be thinking, “I am not going to disregard this adage simply because an essay purports to know what might happen.” I don’t blame you.

In the case of Netflix, the most likely manifestation of the coming AI content explosion might be an impairment of the pricing power the company enjoys at the moment. The service can credibly claim that it adds more value than it costs, so demand is inelastic to price increases. Netflix has proven this point multiple times, and the stock price has acted accordingly. What happens if demand becomes more elastic to price in light of a sea of content alternatives, many of which may not only be low cost but also personalized and compelling?

In addition to Netflix, I would put the following companies in the category of incumbents who are likely to be hurt rather than helped by AI:

• Other content owners, such as Paramount and Warner Bros Discovery

• Recorded music owners, such as Universal Music Group

• Gaming giants, such as Electronic Arts and Take-Two Interactive

• Software incumbents, such as Salesforce, Adobe, and even HubSpot

I realize that the above examples are highly controversial, perhaps most of all Universal Music Group. Don’t get me wrong, music will continue to have value, but profitability should migrate toward live events and Spotify, in my humble view.

I won’t argue the specific examples above in detail, as such arguments tend to be futile in foresight. The market narrative and market-implied expectations are simply too optimistic and too powerful in light of the positive historical stock price performance of the companies listed above, with the stark exception of Paramount and Warner Bros Discovery. You might say that what the Internet did to PARA/WBD, AI may do to the other companies on the list. (I would not include PARA/WBD in a bearish trade — the valuations of both companies already reflect deep pessimism.)

Feel free to brush aside my opinion as that of a no-nothing when it comes to high-flying tech areas such as CRM and SaaS more broadly. But how do you brush aside Satya Nadella? In an appearance on the BG2 podcast a couple of weeks ago, the Microsoft CEO provided an outlook for B2B software that shocked many investors.

Check out the first four-and-a-half minutes:

Similarly, in a recent appearance on Bloomberg Technology, Klarna CEO Sebastian Siemiatkowski explained how AI is enabling his company not only to do away with expensive software — cutting out Salesforce and Workday — but also to freeze hiring, leading to a ~20% reduction in headcount annually.

The example of Klarna serves as a warning to investors in incumbents that stand to lose customers — or are forced to make major concessions — in the coming age of AI.

The Bullish Side of the Coin

Perhaps more importantly, Siemiatkowski provides a glimpse into the newfound ability of many companies to dramatically downsize their cost structure. This applies in particular to businesses that are heavy users of technology or, better yet, have large programming staffs.

The bullish setup: Imagine a company whose stock price has declined precipitously as the former highflyer has struggled to turn a profit. Investors have left the company for dead, resulting in a low multiple of enterprise value to revenue. A company that may have enjoyed a valuation of 20x sales or higher a few years ago may now sell for well under 1x sales.

A turnaround in profitability could produce multibagger upside, as sales remain mostly flat while the sales multiple expands due to strong profit growth. Any resumption in sales growth would add fuel to the fire, but the latter is not really needed, as the company already has a large base of sales relative to its recent enterprise value. The company needs margin expansion, which can be achieved through cost rationalization aided by AI.

A key requirement for any tech company I would consider seriously as part of this bullish setup is large insider ownership. I want management to be heavily incentivized to cut costs and grow shareholder value. Otherwise we could be stuck with a situation in which a CEO remains content to run the business near breakeven. I want a CEO who would be reaching into his own pocket if he desired to keep funding an unnecessarily bloated cost structure. Few owner-operator CEOs will do this, if the ability for dramatic cost cutting is staring them in the face thanks to AI.

Finally, I favor companies with strong balance sheets, i.e., a net cash position. This extends the runway and provides strategic flexbility. It also removes the risk of aggressive debt holders opportunistically forcing the company into a dilutive equity offering (or worse).

Behind the paywall, I list three companies that meet my bullish criteria. I believe these are winners of AI over the near- to medium-term. They have highly skewed risk-reward tradeoffs due to high insider ownership, strong balance sheets, large existing revenue bases relative to EV, and a massive margin expansion opportunity due to AI.

I am long all three companies. Not investment advice. Do your own work.

Drumroll please.

Read more