When AI Winners May Become Cyclical Value Traps

I think the AI market are beginning to show characteristics that are often associated with late-cycle speculative excess.

Not because artificial intelligence is fake. Not because the companies involved are all bad. Not because there is no real demand. That is not the point. The point is valuation, cyclicality, capital intensity, and the assumption that today’s extraordinary investment boom can continue for much longer than economic history would normally suggest.

That is where the market may be becoming increasingly fragile.

We are now seeing many of the classic signs associated with stock market bubbles. Vertical stock charts. Extremely high valuation multiples. A powerful story that many investors want to believe. Retail excitement. Institutional fear of missing out. Index concentration. Companies raising very large amounts of capital while the financing window is open. The narrative has become so strong that price seems to matter less than it should.

The SpaceX IPO is a useful symbol of this phase.

SpaceX may be an incredible company. It may dominate launch markets, satellites, communications, defense contracts, and whatever future space infrastructure eventually becomes. That still does not make a valuation of around 100 times revenue easy to justify. That is not a normal valuation. It is not merely an aggressive valuation. It is an exceptionally demanding one.

A price-to-sales ratio of roughly 100 means investors are paying about 100 years of current revenue for the company. Revenue, not profit. Revenue before operating costs, before capital expenditure, before taxes, before depreciation, before anything. For a capital-intensive company, that is a very difficult valuation to rationalize.

This kind of number is so extreme that it becomes hard to interpret intuitively. It can be calculated, but it does not feel anchored to ordinary business economics. It suggests that investors are pricing in a very large and very successful future, not the current business as it exists today.

And this is not isolated. Palantir, Tesla, semiconductor equipment companies, AI infrastructure names, GPU suppliers, data center beneficiaries, cybersecurity names, “agentic AI” stocks — the whole AI-adjacent complex has been bid up as if a very favorable future is already close to guaranteed.

The market appears to be assigning multiple layers of premium valuation to many AI winners at the same time: a monopoly multiple, a software multiple, an infrastructure multiple, a defense multiple, and a narrative premium.

That is where analysis can start turning into story.

My Thesis

My thesis is simple.

The AI and semiconductor winners may become value traps even after their valuations collapse.

The obvious bubble phase is relatively easy to see. A company trades at 50, 60, 80, or 100 times sales. The market is clearly pricing in a highly optimistic future. But the more dangerous phase may come after the first crash.

First, the valuation multiples normalize. A semiconductor stock may first fall from an unusually high valuation multiple to what looks like a much more reasonable level. Investors and analysts may start saying: “Now it is cheap. This is a high-quality company. The bubble is gone.”

But that may be the exact moment when the real value trap begins.

Because the “E” in the P/E ratio may still be peak-cycle earnings.

If current earnings are inflated by an extraordinary AI capital expenditure boom, then a low P/E ratio is not necessarily low. It may be apparent cheapness based on cyclical peak earnings. This is the same classic trap that appears in cyclical industries again and again.

Oil companies often look cheapest when oil is at the top.
Construction companies often look cheapest at the end of a building boom.
Semiconductor companies can look cheapest when the customer investment cycle is running red hot.

A stock can fall 70%, look cheap on current earnings, and still be expensive on normalized earnings.

The Three-Phase Model

I see the potential AI semiconductor unwind in three phases.

Phase 1: The Bubble Phase

This may be where parts of the market are now, and this phase could still continue for some time.

Valuation multiples are stretched to very high levels. Investors are not only pricing current businesses. They are pricing a very optimistic future: sustained AI demand, continued data center construction, GPU shortages, pricing power, semiconductor equipment spending, and margin expansion. At the same time, several classic bubble characteristics are visible: strong price momentum, a dominant market narrative, crowded investor attention, and a rising number of IPOs and equity offerings as companies take advantage of expensive equity.

The market appears to be treating the current AI investment boom less like a cycle and more like a permanent new baseline.

That is often how bubbles develop. A dangerous assumption in markets is some version of: “This time the cycle is not a cycle.”

Phase 2: The Value Trap Phase

This may come after the first major fall.

The stock price collapses. The P/S ratio compresses. The P/E ratio looks reasonable again. Analysts and investors start calling these companies bargains.

But the earnings may still be peak earnings.

The revenue base may still reflect emergency-level AI spending. Margins may still reflect supply shortages, pricing power, high utilization, and customers over-ordering capacity because they are afraid of being left behind.

So the stock may look cheap mainly because the denominator is inflated.

This is where many investors can get trapped. They think they are buying a wonderful business at 10 times earnings. In reality, they may be buying a cyclical peak at 25, 30, or 40 times normalized earnings.

Phase 3: The Fundamentals Reset

This is the phase where the value trap may actually realize.

Revenue growth slows or reverses. Orders slow. Volumes fall. Customers digest inventory. Pricing pressure appears. Margins compress. Utilization drops. Capex budgets get cut. The cycle turns.

Then investors may discover that the supposedly cheap P/E was misleading.

The stock was not necessarily cheap. The earnings were too high.

This is the exact mechanism I am worried about in AI semiconductors and AI infrastructure. The first crash may only remove the valuation bubble. The second leg may come when the market realizes that the profit base itself was cyclical.

The Strongest Evidence: Semiconductors Have Always Been Cyclical

The strongest evidence for this thesis is not just a vague macro feeling. It is the historical cyclicality of the semiconductor industry.

This industry has never been a smooth straight line. It has had capacity cycles, inventory cycles, pricing cycles, customer capex cycles, margin cycles, and overbuilding phases.

Every cycle has its own story. This one has AI. Previous cycles had PCs, smartphones, cloud, crypto, 5G, memory shortages, supply chain panic, and other “structural” narratives. Some of those stories were real. The cyclicality was still real too.

That is the important part.

A real technological trend does not eliminate the cycle. It can make the cycle bigger.

AI demand is real. But that does not mean AI capital expenditure can grow at an extreme pace indefinitely. Data centers, GPUs, networking equipment, memory, power infrastructure, cooling, and semiconductor equipment can all be overbuilt. The fact that the end market is exciting does not remove basic economics.

When many companies build capacity at the same time, eventually there is a risk of too much capacity.

Then come the usual consequences:

• overcapacity,
• pricing pressure,
• inventory corrections,
• customer capex cuts,
• weaker order books,
• lower utilization,
• margin compression,
• earnings downgrades.

This is not an exotic theory. This is how cyclical capital-intensive industries often work.

The GPU Depreciation Problem

There is also another risk that I think may be underappreciated: depreciation and asset value.

AI infrastructure is not magic. It is physical hardware. GPUs are expensive, they age, and their economic value can fall quickly when newer generations arrive or when demand and pricing weaken.

If companies buy enormous amounts of GPUs and assume long useful lives, reported depreciation may look manageable at first. Earnings can look better. Returns can look acceptable. The investment case can look rational.

But if GPU values fall faster than expected, or if older hardware becomes economically obsolete sooner than expected, then the accounting assumptions can become painful.

At some point, companies may have to recognize that parts of their AI hardware base are worth less than expected. That can mean higher depreciation, impairments, weaker returns on invested capital, and lower real profitability than the market originally assumed.

This matters because the whole AI boom is being financed and justified by future returns on massive physical investment. If the assets depreciate faster than expected, the economics become much less attractive.

A data center full of expensive GPUs is not a software subscription. It is not pure margin. It is hardware-heavy, power-heavy, cooling-heavy, and capital-heavy. If the revenue does not show up fast enough, or if pricing weakens, the numbers can deteriorate quickly.

The Market Is Getting What It Asked For

The market wanted more AI exposure. Now it is getting exactly that.

Massive IPOs. Massive equity offerings. Massive data center capex. Massive GPU orders. Massive infrastructure buildouts. More AI companies preparing to go public. More ways for investors to pay high prices for the same story.

Alphabet raising tens of billions for AI infrastructure. SpaceX raising a record amount of capital at a huge valuation. OpenAI and Anthropic moving toward public markets. Microsoft, Amazon, Meta, Google, and others spending gigantic sums on data centers and compute.

This is what often happens near the hot part of a cycle. The capital markets open, and smart companies use the window. They sell expensive equity because the market is willing to pay. From the company’s point of view, it is rational. If investors are ready to pay very high multiples, take the money.

But from the buyer’s point of view, it may be a poor deal.

The sellers are not stupid. They know when capital is cheap. They know when the story is hot. They know when the market is willing to fund almost anything with “AI” attached to it.

That is another classic bubble sign.

Great Companies Can Still Be Terrible Investments

This is where people often get confused.

Saying something is in a bubble does not mean the companies are worthless. Nvidia is not worthless. ASML is not worthless. Lam Research is not worthless. SpaceX is not worthless. Tesla is not worthless. Palantir is not worthless.

That is not the argument.

The argument is that price matters.

A great company bought at an excessive price can be a bad investment. A cyclical company bought at peak earnings can be a disaster. A dominant company can still disappoint if the market has already priced in perfection.

The AI bulls often frame the question as whether AI is real.

That is not the only relevant question.

The more important questions are:

• How much future growth is already priced in?
• Are current margins sustainable?
• Are current revenues cyclical or normalized?
• How much capex is being pulled forward?
• What happens when customers slow spending?
• What happens when GPU supply catches up?
• What happens when depreciation starts biting?
• What happens when pricing power disappears?
• What happens when the market stops paying very high multiples?

That is where the risk is.

The Bottom Line

My view is that the AI bubble is not just about high valuation multiples. The deeper risk is that the earnings base itself may be inflated by an extraordinary semiconductor and infrastructure cycle.

First, the multiples can crash.
Then the stocks can look cheap.
Then the fundamentals can roll over.
Then the value trap becomes obvious.

That is the sequence I am watching.

The market is currently pricing parts of the AI ecosystem as if extraordinary demand, extraordinary margins, and extraordinary investment growth can continue for a very long time. I am skeptical of that assumption. Semiconductor history does not strongly support it. Capital cycle history does not strongly support it. Basic economics does not strongly support it.

AI may change the world.

That still does not mean investors should pay any price for the companies selling the picks and shovels.

And it does not mean that today’s peak-cycle semiconductor earnings should automatically be treated as a permanent baseline.

That, in my view, is the real trap.