Have you noticed how quickly the AI debate slid from wonder to warning? One month everyone is counting tokens. The next month the same crowd is counting reasons the whole thing should snap. I keep hearing the same line at dinner tables and on trading desks: this looks like 2000, only bigger. That comparison is sticky. It is also incomplete. Excess can be real while the trade still refuses to die on schedule.
The Case Against AI Looks Serious For A Reason
I respect the skeptics. Some of them have spent decades watching technology cycles inflate and then puncture. When people that good line up on one side, you do not wave them off. You go back through the filings, the cash-flow statements, and the ugly concentration charts. Then you decide which parts of the warning are facts and which parts are borrowed from a different era.
The bear argument is not a vibe. It is a stack of claims. Valuations look stretched. A handful of mega-cap names now dominate the headline index in a way that makes passive ownership feel like an accidental concentrated bet. Chip vendors sit inside circular financing webs where equity stakes, cloud credits, and purchase commitments blur the line between organic demand and sponsored demand. Depreciation has not yet caught up with the capex wave. Corporate pilots, in many shops, still struggle to show clean revenue. That is a lot of smoke. Smoke does not always mean the building is already on fire.
Where The Skeptics Are Genuinely Right
Start with concentration, because it is the measure that tells you how fragile a market can become when leadership stumbles. The largest names now account for a share of the benchmark that sits well above the peak of the last tech mania. That is not a trivia point. It is a plumbing point. When ten stocks carry that much of the index, a rotation does not look like a polite handoff. It looks like a drop in the thing everyone thought was “the market.”
Equal-weight versions of the same index have already started to tell a different story from the cap-weighted headline. I have found that this kind of internal split often shows up before the megacaps wobble in public. It does not guarantee a crash. It does tell you the rally is narrower than the slogans imply.
Being early on excess is not the same as being right on timing. Markets punish those two mistakes in very different ways.
Circular financing is real too. When a chip supplier takes an equity stake in a customer that then commits to buying more chips, part of reported demand is the seller helping fund the sale. That does not make every contract fake. It does distort the signal. Anyone who pretends otherwise is selling comfort, not analysis.
Then there is the depreciation lag. Trailing capital spending has raced ahead of the depreciation running through income statements. The bill still arrives. It just arrives later, often in a cluster of years rather than in a single dramatic quarter. If revenue growth cools while that accounting catch-up accelerates, margins will feel it. The bears did not invent that math. They read it.
Surveys of professional managers keep labeling AI a bubble. Crowded long positioning in the same handful of names has been the street’s favorite confession for a long stretch. When the consensus already sounds bearish, the “obvious” crash has a habit of showing up late, or arriving in a shape nobody booked in advance. That does not erase the excess. It complicates the trade.
Why The Fiber Boom Analogy Breaks
The 2000 comparison turns on one variable that actually decided the outcome last time: who wrote the checks. A generation ago, a lot of the fiber was strung by firms that had no business borrowing what they borrowed. Vendor loans papered over customers who could not pay. When traffic and pricing failed to arrive on the promised calendar, the structure fell into bankruptcy court. That was a solvency story dressed up as a technology story.
This cycle is a different animal on that exact axis. A large share of near-term data-center spending is coming from operating cash flow and equity at a small group of extremely profitable platforms. Existing debt, where it appears, is mostly investment grade and still a minority of the spend. I am not saying leverage cannot creep in later. I am saying the balance sheets carrying the core buildout do not look like the weak carriers of the last bust.
That difference is close to the whole game. An overbuild funded with other people’s money collapses when the coupon cannot be paid. An overbuild funded with retained earnings can still destroy shareholder returns. It is less likely to produce a chain of defaults that takes the entire complex with it in one season. Those are not the same risk.
In my experience, investors mash those two outcomes together because the headlines rhyme. Rhyming is not analysis. Cash-funded waste is still waste. It is slower waste. Slower waste gives you time to resize, hedge, and pick layers of the stack instead of sprinting for the exit on a slogan.
Revenue Is Late, Not Imaginary
“No revenue” is a lazy headline. The more precise claim is that infrastructure spending is running ahead of the application-layer cash it is supposed to mint. That can still end badly. It is not the same as a zero-payback story. Inference already clears high gross margins in places where the product is actually in production. Cloud AI run rates at the largest platforms are no longer a rounding error. Some model labs have printed growth that would have sounded like fiction two years ago.
Even the chip name that bears love to use as a punching bag has seen forward earnings climb fast enough that the multiple compressed while the stock stayed in the conversation. That is the mirror image of the classic late-cycle hardware peak, when price-to-sales sat in the stratosphere and earnings then vanished. Multiples can still fall. Fundamentals catching up is not the same tape as fundamentals disappearing.
Perhaps the most interesting aspect is how uneven the monetization looks underneath the slogans. A pilot that never leaves a slide deck is not a product. A copilot feature that reduces handle time in a call center is a product, even if the CFO cannot isolate every dollar. Markets will eventually demand the second thing in bulk. Until then, the gap between capex and clean revenue will keep feeding the bear case. Fair enough. Treat it as a timing risk, not as proof that nobody is getting paid.
Shortage Now, Glut Later?
Bears forecast a glut. The binding constraint on the ground, right now, often looks like the opposite. Power, interconnection queues, transformers, and permitting are rationing capacity. There are stories of idle accelerators waiting on electricity. There are order books that cannot be filled on the customer’s preferred calendar. A large slice of capacity planned for the back half of the decade is not even under construction yet.
If demand is being rationed by the grid rather than by customers walking away, you do not have a classic inventory dump. You have a bottleneck. Bottlenecks produce price power in the near term and overbuild risk in the medium term. That is the honest objection, and it is the strongest card in the bear deck. Build two or three years of generation and transmission, and today’s shortage can become tomorrow’s empty hall.
I will say this plainly. That is more of a later-decade question than a this-year question. Investors who sell everything because 2028 might be ugly can miss a lot of compounding in 2026. Investors who ignore 2028 because 2026 looks tight can get trapped in stranded assets. The job is to hold both thoughts without pretending they cancel.
What A Disciplined Investor Actually Does
The question is not whether there is excess. There is. The question is what you do with a genuine, extreme, cash-funded overbuild. Closing your eyes and buying the whole complex is not a strategy. Dumping every winner because a smart skeptic called a top is how people miss years of compounding while they wait for a crash that arrives late, arrives sideways, or arrives in a different ticker than the one they shorted.
Position sizing comes first. It is the only tactic that still works after a 50 percent drawdown. Several of the leading chip and platform names have already shown they can fall that far more than once and still grind to new highs later. That history is not a promise. It is a reminder that survival weight matters more than forecast elegance. Own the exposure at a size where a brutal correction is uncomfortable rather than fatal.
- Favor self-funders over balance sheets that need perfect capital markets.
- Spread exposure across chips, clouds, and the power layer instead of one ticker.
- Pay for current earnings power, not a total-addressable-market slide.
- Keep cash so volatility becomes a shopping list instead of a panic button.
Where you take the risk matters as much as how much you take. Levered edges of the trade, the vendors stretching to finance customers, the smaller builders who need cheap credit and perfect utilization, look the most like 2000. Those are the first positions to shrink when the story wobbles. The profitable compounders funding their own buildout sit in a different bucket. Selling them because someone declared a top is a different kind of error.
Five Signals That Matter More Than A Hot Take
You do not need to call the exact top. You need a short list of signals that fire before the top is obvious in every group chat. Then you need the discipline to act on the list instead of arguing with it. I like lists that are boring on purpose. Boring lists survive contact with ego.
- Cash-flow funded capex starts giving way to aggressive debt at the core platforms.
- Power and interconnection constraints ease faster than demand, and utilization slips.
- Circular deals grow as a share of reported chip demand rather than shrinking.
- Cloud AI growth stays strong in press releases but weakens in unit economics.
- Equal-weight leadership keeps failing while a few megacaps carry the whole tape.
None of those items requires a crystal ball. They require reading. If three of them flash at once, you do not need a manifesto. You cut the levered stuff, you tighten size, and you stop treating every dip as a gift. If none of them flash, selling a cash-generating compounder because a conference panel used the word bubble is theater.
The Index Is Already The Bet
People ask why they should bother with stock picking when they can just own the benchmark. Here is the uncomfortable answer. With a thin group of names driving such a large slice of the index, buying “the market” is already a concentrated wager on the AI complex. Nobody on the committee had to decide that. The weights did it for them.
That does not make indexing dumb. It makes indexing less neutral than the brochure. If you want ballast, you have to build it on purpose: smaller cyclicals, cash, quality compounders outside the trade, maybe a hedge that costs a little and sleeps well. If you want the AI bet, own it with eyes open and a size you can live with. Accidental concentration is how calm people get surprised.
When all the experts and forecasts agree, something else usually happens.
– A market rule that keeps earning its keep
More than half of surveyed managers calling AI a bubble does not prove the bears are wrong. It does suggest the crash may refuse to arrive on the obvious calendar. One of the better lines in investing is that being too far ahead of your time is indistinguishable from being wrong. Timing is not a footnote. Timing is the trade.
Hedge The Tail, Own The Compounders
Notice how some of the sharpest long-time skeptics still treat the short side as a small hedge rather than a full conviction book. That posture is worth borrowing. Own the businesses that can fund their own mistakes. Hedge the left tail. Let revenue prove or disprove itself on the tape instead of in a thread.
Everyone is eventually right about a drawdown. Markets draw down. The interesting question is narrower. What happens when the richest companies on earth overbuild with their own money rather than borrowed money? That is the question on the table. Answer it before you sell the whole stack because a chart from 2000 looks familiar.
I keep coming back to a simple split. Excess is a diagnosis. The trade is a decision. You can accept the diagnosis and still refuse the decision that says “get out now and wait.” Waiting has a cost. So does stubbornness. The grown-up version of this argument lives in sizing, funding quality, power timelines, and a watchlist you actually honor.
A Practical Map Of The Stack
If you want something more concrete than mood, sort the complex into layers and assign different patience to each layer. Chips are cyclical even when the story is secular. Cloud platforms can bury a lot of capex inside a broader software machine. Power and grid equipment can stay tight long after model hype cools. Application software is where the revenue disappointment, if it comes, will show up first in ugly cohort data.
| Layer | Funding Quality | Near-Term Constraint | Main Risk |
| Frontier chips | Mostly cash plus some customer finance | Packaging, power, lead times | Circular demand and multiple compression |
| Hyperscale cloud | Strong internal cash flow | Electricity and interconnection | Returns on invested capital fading |
| Power and grid | Mixed, often regulated | Permits and equipment | Policy delays, then overbuild |
| App-layer AI | Uneven | Proof of paid usage | Pilots that never become budgets |
Use the table as a bias, not a religion. A self-funded cloud platform can still overpay for growth. A grid name can still sit in a political traffic jam for years. The point is to stop treating “AI” as one ticker with one destiny.
How I Think About Drawdowns Without Playing Hero
I do not try to be the person who sold the exact high. That person is mostly a myth with a good memory. I try to be the person who already decided what a 20 percent, 35 percent, and 50 percent hit would force me to do. Pre-commitment beats improvisation when the tape is red and the group chat is loud.
If a name is a core compounder with a clean balance sheet, a deep drawdown can be a chance to rebalance toward it, not away from it, provided the original thesis on funding and demand has not cracked. If a name is a levered satellite, the same drawdown is a chance to leave. Same percentage drop. Different job.
That sounds obvious written down. It is not obvious at 9:32 in the morning when a chip stock is limit-down on a rumor about delayed racks. Write the rules when you are calm. Follow them when you are not.
The Uncomfortable Middle
Most useful takes live in the middle, which is why they travel poorly. The middle says the buildout is too fast, the index is too top-heavy, the circular deals deserve side-eye, and the companies paying cash can still grind through a lot of waste before the structure breaks. The middle also says power is tight, inference is getting cheaper in a way that can lift volume, and selling quality because a cycle will someday turn is a habit that looks smart in recap pieces and expensive in real accounts.
Will there be a nasty correction in the leaders? History says yes. Will it look like the telecom wipeout, with vendor financing corpses and equity going to zero across a whole cohort? That is a higher bar. Cash changes the shape of the pain. It does not cancel the pain.
If you take nothing else, take this. Treat the AI bear case as a risk memo, not as a sell ticket. Size the trade so you can stay in the names that can fund themselves. Watch funding mix, utilization, and power more than you watch adjectives. Keep dry powder. And remember that a market can be both excessive and not yet finished. Those two facts have shared a room before. They are sharing one now.
