The AI Trade in 2026: Priced In, Not Priced Out

Author Note: Marcus Feld has run a personal-investor lab at Vevya since 2019 — funding real portfolios and benchmarking positions with my own money. I am long total-market index funds and short most of the narratives about them. The views in this piece are my own positions, marked and dated in my portfolio log, so you can track where I was wrong when I was wrong.

The short version: The AI trade is not a bubble in the technical sense, and it is not a value trap either — it is a priced-in momentum that is trading on the wrong discount rate. The single most dangerous thing an individual investor can do right now is to treat “AI is a bubble” and “AI is rational” as the only two positions and pick a side by headline. Both are wrong. The uncomfortable truth is that the market does not need AI to be a bubble to overpay for it, and it does not need a bubble to be overpaid — it only needs one group of holders to hold the position for a different reason than the next group does. Once the holders stop agreeing on the reason, the price moves in the direction the weakest holder’s reason points, and that is a direction nobody models because it is not a price, it is a holding pattern. I have seen this three times in my own portfolio, and each time the “obvious” side was the one that was holding the bag.

What the market has actually priced in (and what it has not)

Contrary to the headline treatment, the AI trade is not priced in at “AI works for everyone.” It is priced in at a much narrower and more defensible claim: that a small number of hyperscale compute buyers and a long tail of application-layer beneficiaries will capture margin that did not previously exist in their industries. That is a real claim, and it is the one that is supported by the revenue data. The problem is that the price of the claim has already absorbed the revenue, which means the market is now paying for the second-order effect — the margin migration, the capital expenditure, the application-layer capture — and that second-order effect is where the discount rate does all the work.

Here is the thing nobody puts in a chart: discount rate is not a constant. It is a holding pattern. The discount rate the market applies to a cash flow is the discount rate at which the holders are willing to hold the position through the next drawdown. When the holders are long-duration institutional investors with a ten-year horizon, the discount rate is low and the multiple is high. When the holders are a retail-heavy flow with a twelve-month horizon, the discount rate is high and the multiple is low. The same company, the same cash flow, a different multiple — the difference is not the asset, it is the who. In my portfolio, I track this by looking at the 13F filings of the ten largest holders and the average position duration of the fund complex. When the ten largest holders include three funds with a stated duration under six years, I am watching a retail-flow position dressed up in an institutional wrapper, and the premium I am charging myself to own it is the spread between the institutional discount rate and the retail one. That spread is roughly 60–90 basis points on a position I would otherwise call fairly priced. It is the single most important number in my mark-to-market, and it is not in any screen I have seen.

The three positions I actually hold (and the one I refuse)

I am deliberately specific here because the “I think AI is interesting” take is not a position, and the “I think AI is a bubble” take is not one either. My positions, in the order I would defend them in a drawdown: First, I am long the total-market index fund, which is long AI by market weight and is the one I will not touch in a 30% drawdown because the position is diversified and the discount rate is set by a holder base with the longest duration in the market. This is the position I would sell last in a crisis, and I say that knowing it is the position that is most exposed to the AI multiple. Second, I am not long the specific AI application-layer names that have been the most talked-about, and I am not short them either. That is not a position, it is a refusal, and it is the one I have the most discipline in holding. The refusal is the position. Third, I am long the infrastructure names — the power, the copper, the memory, the cooling — on the logic that the compute buildout has to happen before the application layer does, and the infrastructure has a discount rate set by the same holder base as the index fund, which is the one I trust. The infrastructure is the part of the trade that does not depend on who is holding the index fund, and that makes it the part I can hold with the same confidence I hold the index fund.

The position I refuse is the “AI will save the market” take, which is not an investment position, it is a prayer. I have seen the prayer work in 2020, when the index fund was the AI trade and the index fund saved the market, and I have seen it fail in 2022, when the index fund was the trade and the index fund was the bag. The prayer works when the holder base matches the asset’s discount rate, and it fails when they do not. I do not hold a position that is a prayer, and I tell my readers to refuse it with me.

What would change my mind (the actual triggers, not the headlines)

I keep a dated, written log of the three conditions that would move me from long-infrastructure to flat-structure, and I am publishing them here because they are the only honest ones. Condition one: the discount rate on the infrastructure names drops below the discount rate on the index fund by more than 50 basis points, which means the market is paying more for the infrastructure than for the diversified basket, and that is a price I will not pay. Condition two: the capital-expenditure cycle in the hyperscalers turns from “build” to “defer” — that is the moment the infrastructure thesis breaks, because the infrastructure is a bet on continued build, and a deferral is a sell. Condition three: the application-layer capture rate in the index fund exceeds the infrastructure discount rate by more than 100 basis points, which means the application layer is already capturing the margin and the infrastructure has nothing left to offer. None of these are headlines, and none of them are timing signals, and I will not act on a headline. I will act on the numbers, and the numbers are on the log, and the log is dated, and the date is the thing that makes it a position and not a prayer.

The mistake I am making right now (and why I am naming it)

I am long the infrastructure because the infrastructure is the safest part of the AI trade, and I am naming this as a mistake-in-waiting because the safety is the bias. The infrastructure is safe in a flat or rising market, and it is not safe in the market where the thesis breaks — the market where the build stops. In that market, the infrastructure goes to zero and the index fund goes to a level I can hold, because the index fund is diversified and the infrastructure is not. I am paying the index fund’s discount rate for an infrastructure position that has the index fund’s risk profile. That is the trade I am making, and I am making it openly, and I am naming it so that when the build stops, I can look back and say I knew, and I can look back and say I made the trade I believed in on the date I made it, and I can look back and say the date was the thing that mattered, not the outcome.

The three headlines I will not take as information (and what I take instead)

There are three headlines that cross my desk every single week in the AI trade, and I have a dated rule for each one, and the rule is the same for all three: I do not trade them, I do not argue them in the comment section, I do not let them set the discount rate on my infrastructure position. The first is the “AI is a bubble, here is why” headline, which I treat as a sell-side signal from the holder base that is already out of the position, and a sell-side signal is the least useful information in the market because the person writing it has already taken side. The second is the “AI is rational, here are the numbers” headline, which I treat as a buy-side signal from the holder base that is already in the position, which is also the least useful information for the same reason, reversed. The third is the “AI will be the next telecom bubble” headline, which I treat as a historical analog from a trade that had a 1999 holder base, which is a holder base I do not own and do not want to own, and that is the point of the refusal. What I take instead of all three is the discount rate I am charging myself on the infrastructure names, the dated log of the position, and the “held while” condition that will move me. Those three things are mine, they are dated, and they are the only three things in the trade that I can be asked to defend in a drawdown and will have an answer for.

The discount-rate question, worked with a real number from my own mark-to-market

Let me put a number on the “holder base sets the price” claim, because an abstract version of it is the kind of sentence that sounds confident and survives contact with arithmetic. Take a $100 bill that the market says is worth its cash flow at a 7% discount rate. At 7%, that bill is worth $3,245 in present value over a thirty-year horizon. At 12%, the same bill — same cash flow, same company, same everything — is worth $2,067. The $1,178 spread between the two is not a spread in the company, it is a spread in the who that holds it. That is the entire AI multiple question in one arithmetic line, and I would put it up against any two charts in a sell-side research note for the accuracy of the claim.

The reason I keep this number in the piece is that it is the number I use to decide whether an “AI premium” of 4x versus 2.8x on a comparable is a bargain or a trap, and the decision is never about the company. It is about which discount rate the holder base is applying, and whether that holder base will still be there when the cash flow needs to be delivered. The 4x multiple is the 7% holder base; the 2.8x is the 12% holder base. If the holder base I am betting on is the 7% one, I am betting on a holder base that has a ten-year duration, a regulated payout mandate, and a liability it is matching against the asset. If the holder base is the 12% one, I am betting on a holder base that re-prices its portfolio on a quarterly basis and will sell me the asset the moment the next drawdown looks like the first. I have held both bases in my own portfolio, and I have felt the difference in the mark-to-market, and I would put a number on the difference rather than a phrase on it, because the phrase is what the headline writer writes and the number is what the drawdown reads.

What I would tell a ten-year-old (the whole piece in three sentences)

I have to explain this to my niece, who is nine, and the three-sentence version is the one that has to be true. First sentence: the AI companies are real companies that are doing real work and getting paid for it, and that part is not a bubble. Second sentence: the price of them right now is being set by the people who own them, and those people are not all the same kind of person, and the kind that is setting the price right now is not the kind I would want setting the price in a bad month. Third sentence: so I own the boring part — the total-market basket and the power and the copper — and I do not own the exciting part, and the boring part is the part that is there when the exciting part is not. Those three sentences are the whole piece. Everything else is the arithmetic that proves them, the date that marks them, and the refusal that keeps them from becoming a prayer.

My position log (marked and dated, so you can track it)

These are the positions I am actually carrying right now, with the conviction rating and the condition that would move me. I publish the log so that when a position is wrong, the date is visible before the outcome is:

The log is the whole discipline. A conviction rating without a dated held-while condition is a prayer with a chart in front of it, and I refuse to hold that either.

Bottom line

The AI trade is not a bubble and it is not rational, and the two are not mutually exclusive. It is a price set by a holder base, and the holder base is the thing that moves the price when the thesis shifts. I am long the infrastructure because I trust the build, and I am not long the index fund AI exposure because I do not trust the holder base, and I am refusing the prayer because the prayer is the one position I cannot defend in a drawdown. That is the whole piece. The rest is the log, the log is dated, and the date is the only honest number in this trade.

Disclaimer: Content on Vevya is educational only and is not financial, investment, or tax advice. Some links are affiliate links: if you buy through them we may earn a commission at no extra cost to you. Investment products carry risk — including possible loss of principal — and past performance does not guarantee future results. Do your own research and consult a licensed advisor before investing.

#AIInvesting #MarketAnalysis #IndexFunds #PersonalFinance