The AI Valuation Problem: When Math Stops Making Sense

There's something odd happening in AI right now. The technology is real—I mean genuinely, transformatively real—but the stock prices make you wonder if people have forgotten how to do arithmetic.

I've been watching this sector carefully, and what strikes me most is how it splits into two distinct realities. There's the fundamental reality, where companies are building actual products that actual customers are paying actual money for. And then there's the valuation reality, where those same companies are priced as if they've already won games that haven't been played yet.

The gap between these realities is starting to matter.

The Profitable Bubble

Here's what makes this different from the dot-com era: these companies are profitable. Nvidia isn't some vaporware startup burning cash on Super Bowl ads. They made $130 billion last year with 75% gross margins. Microsoft's Azure is growing 40% annually. Google Cloud just turned meaningfully profitable.

This is the weird thing about calling it a bubble. In 2000, most internet companies had no revenue model at all. They were literally just hoping to "figure it out later." Today's AI leaders have figured it out. They're printing money.

But—and this is the crucial bit—they're priced as if they'll keep printing money at these rates forever.

The Math Problem

Let's talk about Nvidia, since everyone else is. At a 53x P/E ratio with 41x forward earnings, the market is saying: "Yes, we believe you'll grow 126% this year. And next year. And the year after that."

Except nobody really believes that. Not even the analysts projecting 20-30% growth. Not even Nvidia's management, though they can't say so publicly.

The interesting question is: what growth rate is actually priced in? When you work backwards from the current stock price, you get something like 27% perpetual growth. Not this year—perpetually. As in, forever.

This is where the math starts to feel like magic thinking. Nvidia's entire addressable market would have to grow from $100 billion to $500 billion by 2030, and they'd have to maintain their current market share, and their margins couldn't compress despite AMD, Google, Amazon, and Microsoft all building competing chips.

That's a lot of ands.

The Custom Silicon Problem

Here's something most people miss: Nvidia's biggest customers are also becoming their competitors. Not obviously, not loudly, but systematically.

Google built TPUs. Amazon built Trainium. Microsoft has Azure Boost. Meta is designing custom silicon. These aren't side projects. Google's TPUs now deliver 4-6x better cost efficiency than H100s for inference workloads. Amazon is projecting $60 billion in AI infrastructure spend, and a meaningful chunk of that is going to not buying Nvidia chips.

Why does this matter? Because Nvidia currently has 85%+ market share in AI chips, and that share is the asset. It's what justifies the 75% gross margins. It's what makes the CUDA software ecosystem valuable. It's the moat.

But moats erode. Not in one quarter, not in one year, but in the way water wears down stone—slowly, then all at once.

The moment Nvidia's margins compress from 75% to 65%, analysts will re-rate the stock 30% lower overnight. Not because the business is broken, but because "durable monopoly with 75% margins" is a very different valuation story than "strong player in competitive market with 65% margins."

The Capex Paradox

Then there's the spending problem. Microsoft, Google, Amazon, and Meta will collectively spend $293 billion on AI infrastructure in 2025. That's up 62% from last year.

Now, capex itself isn't bad. Companies should invest in growth. But here's the paradox: as their capex has increased, their returns on that capex have decreased.

Microsoft is now spending 40% of free cash flow on capex, up from 15% in 2020. Google is at 60%, up from 30%. These are companies generating $72 billion and $100+ billion in annual cash flow—they can afford it. But if you're spending 60% of your cash on building infrastructure, you need that infrastructure to generate returns. And increasingly, the utilization numbers are worrying.

Recent reports suggest only 60-70% of deployed H100/H200 capacity is actually in use. That means 30-40% of these expensive GPUs are sitting idle. You can't earn returns on idle capacity.

What happens when CFOs start asking: "Why are we spending $30 billion a year on chips we're not fully utilizing?" That's when the capex cycle turns.

The Enterprise Reality Gap

The other thing that doesn't quite add up is enterprise adoption. Survey after survey shows 70-80% of companies "using AI." That sounds amazing until you dig into what "using" means.

Most companies are running pilots. A small team trying ChatGPT for writing emails. A demo of an AI coding assistant. Maybe a predictive analytics dashboard that nobody actually makes decisions from.

The gap between piloting and production is vast. Only 5-10% of companies have moved past experimentation to actual, scaled deployment that changes workflows. This matters because pilots generate one-time infrastructure purchases, but scaled production generates recurring revenue.

Right now, the valuations assume everyone's moving to production. But the data suggests people are stalling in pilot purgatory. Not because the technology doesn't work—it does—but because organizations are slow, risk-averse, and uncertain about ROI.

Microsoft's Copilot attach rate is below 10% of eligible users despite aggressive sales efforts. Salesforce's Einstein is attached to 40-50% of users but generates only $2-5 per user monthly. These aren't the metrics of a revolution. They're the metrics of a feature.

The Correction Math

So when does this correct? And by how much?

If I had to bet, I'd say there's a 60% chance of a 15-25% correction within 12 months. Not a crash—a repricing. Here's why:

The correction will come when one of three things happens:

1. Nvidia misses guidance or signals growth is slowing
2. A major hyperscaler (Microsoft, Google, Amazon) cuts AI capex
3. Enterprise AI attachment rates stay low for another 2-3 quarters

Any of these events would force analysts to revise their perpetual growth assumptions downward. And when those assumptions go from "30% forever" to "20% for 3 years, then 10%," the stock price math changes dramatically.

The interesting part is that this doesn't require the business to break. Nvidia could still be an excellent company growing 15-20% annually with 60% margins. That's an amazing business! But it's not a $3 trillion business. It's maybe a $2 trillion business.

That $1 trillion difference is the bubble.

What's Actually Worth?

Here's my rough hierarchy of value right now:

Worth holding: Microsoft, TSMC, Google. These companies have real, diversified businesses with reasonable valuations given their growth and moat strength. Microsoft at 35x earnings for a company growing 17% with Azure accelerating and massive cash generation seems fine. TSMC at 24x earnings with a monopoly on advanced chip manufacturing seems cheap.

Worth avoiding: AMD, Broadcom, anything trading above 80x earnings. AMD at 136x P/E growing 18% is just mathematical nonsense. Broadcom at 89x P/E might be a great business but it's a terrible price.

Worth watching: Everything in the private markets. OpenAI at a $300 billion valuation on $12 billion revenue with no clear path to profitability is the kind of number that makes sense only if you squint really hard and assume everything goes perfectly for a decade. Anthropic at $183 billion on $3 billion revenue is even more aggressive.

These private valuations matter because they set expectations. When these companies eventually go public—or worse, need down rounds—it'll reset everyone's mental models about what AI companies are "worth."

The Long View

Here's what I think is actually happening: AI is a genuinely important technology that will create enormous value over the next decade. But most of that value will accrue to users, not shareholders. The companies building AI will compete margins down to normal levels, just like companies in every other technology wave have done.

The internet created trillions in value. But most internet companies didn't capture that value—they competed it away. Google captured some because they had a genuine monopoly. Amazon captured some because they had better execution than everyone else. But most value went to consumers in the form of free services and businesses in the form of productivity gains.

AI will probably follow the same pattern. The technology is real. The value creation is real. But the stock market prices assume that the companies building the infrastructure will capture most of that value, and history suggests that's rarely true.

The smart play isn't to short everything or avoid AI entirely. It's to be selective. Own the companies with real moats (TSMC's manufacturing monopoly), proven monetization (Microsoft's enterprise customers), and reasonable valuations (not paying 100x earnings for anything).

And maybe keep some cash available. Because when this corrects—and I think it will—there will be some genuinely good companies trading at genuinely good prices.

That's when the real opportunity starts.