Why Stock Prices for AI Companies Defy Valuation Logic

The Math Simply Doesn't Work

There is a growing problem at the heart of modern markets: some of the most talked-about companies in the world cannot be rationally valued using any conventional financial framework. Not because analysts lack sophistication — but because the numbers, when run honestly, don't come close to justifying current prices.

Take SpaceX as a concrete example. The company generates approximately $6.8 billion in operating cash flow. Running a discounted cash flow (DCF) analysis — the gold-standard method for estimating what a business is actually worth — using extremely generous assumptions (20% annual growth for a decade, a 20x exit multiple, a 10% target return, and ignoring capital expenditures entirely) produces an intrinsic value of roughly $440 billion. No margin of safety. SpaceX's current private market valuation? Approximately $2.5 trillion. That's more than six times the already-optimistic DCF estimate.

This isn't a minor discrepancy. It's a mathematical gulf. And it raises an uncomfortable question every serious investor needs to wrestle with: when stock prices for AI companies are this disconnected from earnings reality, what exactly are buyers purchasing?

Howard Marks, founder of Oaktree Capital and one of the most respected value investors alive, has a clear answer: they're speculating. Not necessarily foolishly — but speculating nonetheless.

What Howard Marks Actually Said — And Why It Matters

Marks recently argued that investing in AI IPOs like Anthropic or OpenAI is fundamentally different from analytical investing. His core point: to value any company properly, you need two things. First, a forecast of future earnings. Second — and this is the part most investors skip — a calibrated judgment of how probable that forecast actually is.

Most retail investors running DCF models on AI companies nail the first step and completely ignore the second. They'll project 25% annual revenue growth for a decade without seriously interrogating whether that growth rate has any historical precedent or structural support. Marks argues this is precisely where bubbles are born.

His challenge is direct: if someone claims they know what Anthropic's net earnings will be in 2036, he'll bet them they're not within 50% of the actual figure. That's not pessimism — that's epistemic honesty about the limits of forecasting in genuinely novel industries.

This framing matters because it reframes the AI investment debate away from "is AI transformative?" (almost certainly yes) toward "can you price that transformation accurately enough to make a sound investment?" (almost certainly no, at current valuations).

The Historical Precedent Is Not Comforting

Marks draws a direct line from the current AI frenzy to every major technology bubble of the past 150 years: railroads in the 1860s, radio in the 1920s, automobiles, mainframe computers in the 1950s and 60s, and internet stocks in the late 1990s. Each followed the same pattern:

• A genuine, world-changing technology emerged
• Capital flooded in, far exceeding rational allocation
• Infrastructure was massively overbuilt relative to near-term demand
• Prices were paid that could never be justified by eventual earnings
• Many investors who funded the build-out lost significant capital

The internet bubble is the most instructive comparison. Broadband and e-commerce did transform the global economy — just not on the timeline or through the winners that 1999 investors assumed. Cisco, a darling of that era, took over 20 years to recover its dot-com peak price. Amazon survived and thrived, but it fell 90% from peak to trough between 2000 and 2001 before its eventual ascent.

The lesson isn't that transformative technology is a bad bet. It's that being right about the technology and being right about the investment are two completely separate things — and the gap between them is where fortunes are lost.

What makes AI arguably more dangerous than previous cycles is what Marks describes as its near-unlimited and poorly-specifiable upside. With railroads, investors knew trains would move goods coast to coast. With radio, they knew it would carry messages. With AI, even the companies building it cannot fully articulate what it will do, for whom, at what margin, or in what competitive landscape. That ambiguity cuts both ways — and current valuations appear to price in only the upside scenario.

Goldman Sachs estimates that Meta, Microsoft, Amazon, and Alphabet alone will spend $5.3 trillion on capital expenditures between fiscal years 2025 and 2030. In 2026 alone, those four companies are projected to deploy $725 billion in capex — a 77% increase year-on-year. That is an enormous bet on AI demand materialising at scale. History suggests not all of it will.

The Risk Spectrum: Where You Play Determines What You Risk

None of this means investors must sit entirely on the sidelines. Marks lays out a clear three-tier framework for thinking about AI exposure, and it's worth internalising:

Tier 1 — Hyperscalers (Lower relative risk): Companies like Alphabet, Microsoft, Amazon, and Meta. These businesses have established revenue streams, wide economic moats, diversified income, and massive operating cash flows. AI is a significant strategic bet for them — but it isn't existential. If AI spending yields less than hoped, these companies don't collapse. They adapt. For investors who want AI exposure without betting the farm on a single product category, this is the rational entry point.

Tier 2 — Established AI Pure-Plays (Medium risk): Companies like Nvidia, OpenAI, or Anthropic. These businesses are further along than startups — they have revenues, real products, and genuine market positions. But they are far more exposed to AI-specific risk. If the market for large language models consolidates aggressively, or if open-source models commoditise the space, these companies face existential threats that hyperscalers do not. Price paid matters enormously here. At reasonable valuations, the risk-reward could be compelling. At current IPO prices, the margin for error is razor thin.

Tier 3 — Startups and Pre-Revenue AI Companies (Lottery-ticket risk): Ground-floor access to AI startups can theoretically produce extraordinary returns — but the base rate of startup success is brutal. Most will return zero. A small number will produce life-changing gains. This is not investing in any conventional sense; it's venture capital-style speculation. Position sizing accordingly — meaning a small, defined percentage of a portfolio that an investor can afford to lose entirely.

The key insight from this framework: it's not binary. Investors don't have to choose between full AI exposure and complete avoidance. They can calibrate their position on the spectrum based on risk tolerance, time horizon, and portfolio size — then size each position proportionally.

The Boring Businesses Buffett Would Rather Own

Warren Buffett famously sat out the late 1990s tech bubble while being publicly mocked for missing out. He then watched much of the market fall 50–80% while his portfolio held its value. His reasoning was simple: he couldn't model the earnings trajectory of internet companies with enough confidence to justify the prices being asked. He could model Coca-Cola, See's Candies, and Wells Fargo.

Marks echoes this thinking directly. The sectors he points to as more defensible right now are telling: energy, food, timber, home building, metals and mining, paper, chemicals, and transportation infrastructure. The common thread? These industries have lower intellectual content, which makes them structurally less likely to be disrupted by an AI system that is, at its core, an intellectual productivity tool.

Think about what AI cannot easily replace: a barrel of oil extracted from the ground, a cubic metre of timber, a tonne of copper ore. These are physical outputs with relatively predictable demand curves. Modelling a lumber company's cash flows over the next decade is not easy — but it is orders of magnitude more tractable than modelling OpenAI's.

This is also where genuine pricing inefficiencies still exist. Large-cap tech dominates financial media, analyst coverage, and retail investor attention. That means the boring sectors — waste management, industrial manufacturing, specialty chemicals, community banking — often trade at more reasonable multiples, with less competition for information advantage. This is the core of what Mohnish Pabrai calls "boring but brilliant" investing.

What Rational Long-Term Investors Should Actually Do

The synthesis of Marks' arguments produces a practical framework rather than a counsel of despair. Here's how to think about it:

• Acknowledge the speculation for what it is. If you're buying Anthropic at IPO or loading up on a pre-revenue AI startup, you're speculating — not investing analytically. That doesn't make it wrong, but it should change how you size the position.
• Run the DCF yourself — and then ask how confident you are. A forecast without a probability assessment attached to it is half a thought. If your growth assumptions require AI to eat entire industries within five years, ask honestly: what are the odds that happens on schedule?
• Size AI pure-plays as a fraction of a diversified portfolio. Even committed growth investors rarely allocate more than 5–10% of a portfolio to genuinely speculative positions. Above that, you're not managing risk — you're ignoring it.
• Look at the hyperscalers with fresh eyes. Companies like Alphabet and Microsoft are not cheap in absolute terms, but they are far more defensible than pure-play AI bets. They have pricing power, recurring revenue, and the balance sheets to absorb years of heavy AI investment without existential risk.
• Consider rebalancing toward sectors with predictable earnings. Energy infrastructure, consumer staples, materials, and industrials are not exciting. They are, however, where Warren Buffett and Howard Marks both suggest the better risk-adjusted opportunities may currently lie.

The stock market is not just a voting machine for enthusiasm about the future — it's eventually a weighing machine that demands earnings justify prices. In every previous technology cycle, that reckoning arrived. The investors who survived it best were the ones who knew exactly what they owned, why they owned it, and what they were willing to lose if they were wrong.

---

This article is for informational purposes only and does not constitute financial advice. Always consult a qualified financial professional before making investment decisions.

Frequently Asked Questions

Why can't analysts value AI companies like Anthropic or OpenAI accurately?

Traditional valuation methods — particularly discounted cash flow analysis — require two inputs: a forecast of future earnings and a confident probability that the forecast is accurate. AI companies fail on the second requirement. Their revenue models, competitive dynamics, regulatory environments, and addressable markets are all highly uncertain over a 5–10 year horizon. Even the companies themselves cannot specify their end-state business models with confidence. This doesn't mean they won't be valuable — it means no one can price that value reliably today.

Is the current AI investment environment really comparable to the dot-com bubble?

The structural similarities are significant: a genuinely transformative technology, massive capital inflows, winner-take-all competitive framing, and valuations that require optimistic long-term assumptions to justify. Howard Marks explicitly draws this comparison, noting that every previous technology cycle — railroads, radio, internet — was accompanied by a bubble where too much capital flowed in, too much infrastructure was built, and many investors lost money. The distinguishing feature of AI is that its potential upside is arguably even harder to specify than past technologies, which cuts both ways on uncertainty.

What does Howard Marks recommend for investors who want AI exposure without excessive risk?

Marks describes a three-tier risk spectrum. At the lower-risk end sit hyperscalers — large technology companies with diversified revenue streams and strong moats that are integrating AI without being entirely dependent on it. In the middle sit established AI companies with real revenues and products but higher concentration risk. At the highest-risk end are early-stage startups with unproven business models. Marks' framework suggests matching position size to risk tier — larger allocations to hyperscalers, much smaller allocations (if any) to startups where the outcome profile resembles a lottery ticket.

Are there investment opportunities outside of AI and technology right now?

Both Howard Marks and Warren Buffett point toward sectors with predictable earnings and lower disruption risk as offering more defensible value. These include energy, food production, timber, home building, metals and mining, chemicals, and transportation infrastructure. The common logic: AI is fundamentally an intellectual productivity tool, so industries based on physical output — extracting, growing, or building things — are structurally harder to disrupt. These sectors also tend to receive less media coverage and analyst attention, which can create more attractive entry prices for patient investors.