Are AI Stocks Really Overpriced? You Might Be Using the Wrong Math (or None at All)

AI and semiconductor stocks have been on a run. Two kinds of critics show up when that happens.

The first group reaches for the P/E ratio. They compare it to the S&P 500’s long-run historical average of around 16, a number derived from data going back to the early 1900s and blended across two world wars, the Great Depression, and the postwar boom. “We’re being disciplined,” they explain. What they skip is that the S&P 500 hasn’t actually traded near 16 in any sustained way for decades. Over the last 25 years it has averaged closer to 22 to 25 in normal years and currently sits around 32. The 16 baseline isn’t modern market context. It’s a century-long average that makes today’s multiples look more extreme than they are.

The second group doesn’t bother with metrics at all. They look at market caps doubling and tripling, stocks up 200% in a year, and announce it’s obviously a bubble. Not because of any calculation. Because the sheer scale of the moves feels crazy. Prices that big, moving that fast, have to be irrational. Right?

Both groups are wrong. One is using the right formula with the wrong inputs. The other isn’t using a formula at all. They’re reacting to the size of the numbers and calling it analysis.

When you run the traditional valuation models the way the textbooks say to run them, accounting for growth rather than ignoring it, some of the stocks everyone is calling overpriced come out looking like screaming bargains.

The whole market could crater tomorrow from a rate shock, a geopolitical event, or a bad quarter from somewhere nobody saw coming. None of that has anything to do with any company’s earnings trajectory. If we could actually see the future, we’d be on a yacht somewhere, not writing blog posts on the internet. This is about whether the valuation logic holds up. The macro is a separate and genuinely unknowable question.

The Rearview Mirror Trap

When analysts or news anchors mention a company’s “P/E ratio,” they almost always mean the trailing P/E: price divided by earnings over the last twelve months. Backward-looking by definition. It tells you what the company earned last year, priced against what the stock costs today.

For a slow-growth utility or consumer staple, where next year will probably look a lot like last year, trailing P/E is a reasonable proxy. The past is a defensible guide to the future when the future is just more of the same.

For a company in the middle of a steep earnings ramp, trailing P/E is measuring the wrong thing. You’re pricing the future using numbers from before the inflection happened. That’s not conservative. That’s a category error.

Micron (MU) is the clearest example of what this looks like in practice. Micron makes HBM, high-bandwidth memory, the specific chip that physically sits inside every AI accelerator. Without it, the AI compute stack doesn’t run at scale. There are exactly three companies on earth that produce it: Micron, SK Hynix, and Samsung. Every major AI data center in the world needs HBM, supply is constrained, and demand is being driven by one of the largest infrastructure buildouts in tech history. Micron isn’t a company catching AI vibes. It’s one of three suppliers at the chokepoint of the entire AI buildout.

As of May 2026, the trailing P/E is 45.87. On that number alone, plenty of people call it expensive and move on.

Now run the forward numbers. Analyst consensus for Micron’s fiscal year 2026 (ending August 2026) is EPS of $58.73, up from $8.29 last year. That’s 608% earnings growth. At a share price of $971, the forward P/E for this fiscal year is about 16.5.

The stock trading at 46x trailing earnings is trading at 16.5x this year’s earnings. That is almost exactly the long-run S&P 500 average the traditionalists keep citing as fair value.

Go out one more year. FY2027 consensus EPS is $105.28. The forward P/E drops to roughly 9.2. By that measure, Micron is trading at less than half the historical average people call normal, while growing earnings at a pace most companies never approach in their entire existence.

The trailing P/E looked scary because it was measuring a company just coming out of a low-earnings trough. The expensive-looking multiple was a snapshot of last year, not a prediction about next year.

Accounting for Growth Is the Point

Any valuation method that ignores how fast a company is growing isn’t a complete valuation. That’s not a controversial position. It’s arithmetic. Every major valuation framework, from Discounted Cash Flow to relative multiples, requires an assumption about future growth. The only debate is how to quantify it.

One clean shorthand is the PEG ratio (Price/Earnings-to-Growth): take the P/E, divide it by the annual earnings growth rate. The standard benchmark, used across finance for decades, is that a PEG under 1.0 means the stock is undervalued relative to its growth.

Peter Lynch, who ran the Fidelity Magellan Fund from 1977 to 1990 and delivered 29.2% average annual returns over that stretch, popularized the PEG ratio in his 1989 book One Up on Wall Street. He didn’t invent the underlying principle. He made the shorthand famous.

Run the same math on two hypothetical stocks:

Now apply it to Micron using the forward P/E and next-year growth rate:

A PEG of 0.21, using the exact P/E multiple traditionalists consider fair value, against real analyst consensus numbers. The math says undervalued. The headlines say expensive.

The critics calling this a bubble are doing what the framework says not to do: looking at the multiple in isolation, skipping the denominator, and presenting the mistake as discipline.

The Textbook Agrees

The same logic runs through the Discounted Cash Flow model, the bedrock of corporate finance. Project the company’s future cash flows, discount them back to present value, and you get intrinsic worth. The terminal value, the present value of everything beyond the explicit forecast period, uses the Gordon Growth Model:

TV = CF x (1 + g) / (r - g)

The most sensitive variable in the formula is g, the growth rate. As g increases, the present value of the firm scales nonlinearly. A company credibly doubling its cash flow every year will produce a fair-value share price that demands a very large current multiple. Not because the model is being generous. Because the math is working correctly.

The critics plug in a 6% terminal growth rate for a company growing at 80%, announce the stock is overpriced, and conclude the market has lost its mind. What actually happened is they manually overrode the most important variable in the formula because the number felt uncomfortable, and then blamed the equation.

The textbooks aren’t broken. The inputs are wrong.

The Money Is Actually Following the Earnings

In a real bubble, money flows into everything in the sector indiscriminately. That’s not what’s happening.

Microsoft, one of the biggest AI players on the planet, has largely stagnated. Nvidia had a historic run, then plateaued as its earnings growth rate moderated. The capital has rotated toward the next tier: AMD, Intel, and TSMC, companies whose earnings are now accelerating as AI infrastructure buildout drives chip demand. The money is tracking earnings trajectories, not hype cycles. When a company’s growth rate flattens, the market moves on to whoever is growing fastest.

That’s the market working as designed. It’s the opposite of a casino dumping money into every company that has “AI” in its pitch deck.

Not Every High Multiple Is Earned

None of this means every high-multiple stock is justified. There’s a real difference between a company with minimal revenue valued on narrative and social media hype, and a company selling every unit it can manufacture at high gross margins with committed forward orders. The first might be a bubble. The second is a fast business, and traditional valuation models will confirm that if you actually run them with the real growth numbers.

If you’re calling it overpriced but haven’t checked the growth rate, you didn’t do a valuation. You just reacted to a big number.