What Happens If AI Falls Short of the Bet the Market Has Made

Every bull market eventually asks you to believe in something you cannot see yet. Right now the market is asking you to believe in artificial intelligence the way a rookie believes he will hit .320 in the show. Maybe he will. He has the swing, the bat speed, the scouting reports all say so. But you do not pencil him into the middle of your batting order and bet the season on it until he has done it against big league pitching, in September, with the games that count.

That is roughly where we sit with AI and the markets that have been built on top of it.

I want to be straight with you before I say anything alarming, because I do not believe in ringing bells that are not ready to ring.

Price trends across the major averages remain up.

Credit conditions, by every measure I track each week across the high yield spreads, the CCC tier, and the National Financial Conditions Index, remain loose and accommodating.

Corporate earnings are not just growing, they are growing broadly, with margins holding up in sectors that have nothing to do with a single graphics processor.

None of the four pillars I have used for 30 years to separate a real bull market from a mirage, value, credit, trend and momentum, is currently flashing red. Three of the four are flashing green. Value is the one pillar that has gone quiet on the mega cap names, and quiet is not the same thing as broken.

So no, I am not writing to tell you the sky is falling. I have watched enough market cycles fall apart to know that the crash never arrives on the schedule the doomsayers publish. It arrives later, and from an angle nobody was staring at. My job this month is to point you at the angle.

Here is the plain version of the problem. An enormous share of the market's advance over the past three years, and almost all of the market's advance over the past 12 months, rests on the expectation that artificial intelligence will generate a return on capital commensurate with what is being spent on it. That is not a controversial statement. It is closer to an accounting identity. Strip AI related capital spending and the earnings and share price gains tied to it out of the S&P 500, and you are left with a market that looks a good deal more ordinary, a good deal more like the slow, grinding, cyclical animal it has always been.

I read through a lengthy piece of research and had several lenghtyconversations this month on what the trade press has taken to calling the AI infrastructure virtuous cycle, and I want to walk you through what it actually says, because the plumbing matters more than the headlines.

The idea of the virtuous cycle is simple enough.

Capital gets raised by the frontier model companies and the hyperscalers. That capital gets spent on chips, data centers, power and networking. That spending becomes revenue for the chipmakers, the cloud providers and the colocation companies. That revenue and the backlog it creates support more capital raising. And the new capacity supports more model training and more customer adoption, which is supposed to justify the whole thing over again.

It is a flywheel, and flywheels are wonderful things right up until friction, or a missed step, slows one down.

The numbers behind this cycle are almost too large to hold in your head at once. Global data center investment reached roughly $500 billion in 2024, nearly double the 2022 level. Alphabet spent $91.4 billion of capital expenditure in 2025. Amazon has guided to roughly $200 billion of companywide capital spending in 2026. Meta's guidance for the year runs $130 billion to $145 billion. Microsoft has said that roughly two thirds of its recent capital spending has gone toward short lived assets, mostly chips, meaning equipment that will need to be replaced or upgraded on a cycle measured in a few years rather than a few decades.

None of that, by itself, is alarming. Big industries get built by big spending. Railroads, electrification and the interstate highway system were not funded out of pocket change either. What is worth your attention is not the size of the spending. It is the structure of who is financing whom.

Amazon has invested up to $25 billion in Anthropic while Anthropic has committed to spend more than $100 billion on Amazon Web Services technology over the next decade.

NVIDIA has said it intends to invest as much as $100 billion in OpenAI as OpenAI deploys NVIDIA hardware.

OpenAI has committed up to $11.9 billion to CoreWeave while receiving $350 million of CoreWeave stock in the same transaction.

Cerebras funded OpenAI a working capital loan of roughly $1 billion, received a warrant tied to that funding, and is repaying the loan not in cash but in non cash service credits, meaning capacity rather than currency.

I want to be careful with you here, because it would be easy and lazy to call this a shell game and leave it at that. The evidence does not support that stronger claim. These arrangements are disclosed, not hidden, in SEC filings and company reports.

There is real, outside, third party revenue flowing through this system too. Amazon Web Services says its AI business has topped a $25 billion annual run rate. Microsoft's Azure cloud business has topped $75 billion in annual revenue. Alphabet's cloud backlog stands at $514 billion. Those are not numbers you get from companies simply trading capital back and forth in a circle with no one else in the room.

What the evidence does establish, cleanly and without much room for argument, is that the companies at the center of this buildout occupy two or three roles at once. They are investor, customer, supplier and lender, often in the same transaction. When NVIDIA is both a greater than 5% owner of CoreWeave and a customer paying CoreWeave for services, and when OpenAI is simultaneously a customer of Cerebras, a lender to Cerebras, and a warrant holder in Cerebras, you have built something that only works cleanly as long as every leg of the chain keeps performing.

Pull one card out of that structure and you find out how many other cards were leaning on it.

Here is where the debt markets come into the picture, and this is the part of the story I think gets the least attention from investors who are watching only the Nasdaq ticker.

The major hyperscalers issued roughly $121 billion in new debt during 2025, more than four times their five year average. AI linked companies now make up roughly 14% of the JPMorgan investment grade bond index, a larger share than the banks.

Beneath the debt that is already on the books sits something bigger still. Microsoft, Meta, Oracle, Amazon and Alphabet together have disclosed roughly $1.09 trillion in future lease commitments tied largely to data centers, against only about $285 billion of lease liabilities currently recognized on their balance sheets. That gap exists because accounting rules do not require a lease to hit the balance sheet until the facility is actually up and running.

The obligation is real today. The accounting recognition of it is not, yet.

I have spent 30 years underwriting community banks by asking one question above all others: what happens to this institution's capital if the assumption behind its largest asset turns out to be wrong. I am asking the same question here, except the asset in question is a hyperscale data center full of graphics processors, and the assumption behind it is that AI demand, and AI pricing, will keep climbing at something close to its current pace for years to come.

Graphics processors are not like a bank branch or an office building. A data center's walls and cooling systems will last three or four decades. The chips inside it are closer to a fleet of delivery trucks. They depreciate fast, and a new generation of hardware can make the old generation economically obsolete well before it is physically worn out.

When that hardware is also the collateral behind a loan, as it is in several of CoreWeave's financing arrangements, the collateral value and the revenue generating capacity of the business can decline together, at the same time, for the same reason.

That is not a risk you want stacked underneath a piece of the investment grade bond index.

Now let me bring this back to something closer to home, because a Baltimore boy does not spend his whole column talking about hyperscalers and warrants without circling back to the batting order.

I have told you for years that price, credit, trend and momentum are your four outfielders, and when all four of them are playing shallow and moving well, you can trust the defense. Right now three of the four are playing well. Trend is up. Credit spreads are tight, not widening, which tells you the bond market has not yet started pricing real distress into the system, even with all that new AI linked debt sitting on the shelf. Earnings momentum, outside the handful of AI infrastructure names, remains broad and healthy.

That is not the profile of a market standing on the edge of a cliff. It is the profile of a market that is, for the moment, still playing good defense.

Value is the pillar sitting on the bench, and it deserves a longer look, because it is where this whole essay is actually pointed.

When you price a stock like NVIDIA, Microsoft, Oracle or CoreWeave today, you are not pricing a company.

You are pricing a set of forecasts about future utilization of data centers not yet built, future pricing of inference that has never existed at that scale before, future power availability in regions that do not currently have the grid capacity to support it, and future productivity gains for the broader economy that, as of a Federal Reserve staff note from July of this year, have not yet shown up in the aggregate data in any way that would justify the scale of the investment already committed.

The Fed's own researchers said plainly that AI related investment is contributing meaningfully to GDP growth right now, but that broad, economy wide productivity gains remain limited and unproven at the aggregate level. Investment is running well ahead of return.

That is the sentence I want you to sit with. Investment is running well ahead of return. It has been for a while. It can keep running ahead of return for a good while longer, because markets are patient with growth stories right up until they are not. But when a valuation already assumes flawless execution, on demand, on pricing, on power, on productivity, and on financing, there is no room left in the price for anything going merely fine instead of perfect.

Fine is a good outcome in most of life. In a stock priced for perfection, fine is a disappointment, and disappointment gets punished in a way that has nothing to do with whether the underlying business is actually healthy.

I think about Marty Whitman here, more than almost any other of the old value investors I lean on. Whitman used to say safe and cheap, in that order, because safe protects you from the downside and cheap is what gives you the upside. A great deal of the AI infrastructure complex today is neither.

It is not cheap, by any conventional measure of price to earnings, price to cash flow or price to tangible assets.

It is not safe either, not when the collateral behind billions in debt is a chip that a competitor's next product cycle could make obsolete, and not when a meaningful share of reported revenue is flowing between companies that also happen to be one another's investors.

None of this means the AI buildout is fake, and I want to be honest about that, because I have no patience for people who dress up a legitimate warning in conspiracy language to get attention. The revenue is largely real. The demand is largely real. Enterprises are actually buying and using these tools, and cloud providers are actually monetizing capacity that did not exist three years ago.

What is unresolved, and what the Federal Reserve itself says is unresolved, is whether the return on the capital already committed will match the scale of what has been spent. That is a different question than whether AI works. AI can work, genuinely and durably, and the stocks that were priced on the assumption of perfection can still fall hard if reality merely turns out to be very good instead of flawless.

So where does that leave you, sitting at your kitchen table trying to decide what to do with your own capital?

It leaves you exactly where a disciplined investor should always be standing, which is watching the data rather than the narrative. The narrative right now says AI changes everything, and it probably does change quite a lot. The data says trend is up, credit is loose, and earnings outside the AI infrastructure complex are broad and healthy. Trust the data for as long as the data holds.

Do not confuse a strong tape with a safe valuation, because those are two different animals wearing the same uniform.

If you own the picks and shovels of this buildout, the chipmakers, the hyperscalers, the data center operators, ask yourself honestly whether you are being paid an adequate margin of safety for the assumption of continued perfection, or whether you are simply along for the ride because the ride has been good so far.

If the answer is the second one, there is no shame in it, plenty of rides have been good for a long time, but know what you actually own. You own a bet on flawless execution across financing, power, hardware and demand, all at once, for years running. That is a demanding standard for any industry to clear.

If instead you are sitting in the parts of the market this cycle has largely ignored, community banks trading below tangible book value, REITs priced at a real discount to replacement cost, the deep value corners where Ben Graham's arithmetic still works the way it always has, you are in a different and, I think, better position.

Those holdings do not need artificial intelligence to hit .320 in order for you to earn an acceptable return. They need earnings, book value and dividends to keep doing what they have always done. That is a much lower bar to clear, and a much more comfortable one to be standing behind if the AI story turns out to need a few more innings before it delivers on its promise.

Stoicism teaches you to separate what is within your control from what is not. You cannot control whether data center utilization comes in at 90% or 60% five years from now. You cannot control whether the next generation of chips makes today's fleet obsolete a year early.

You can control how much of your portfolio is priced for perfection versus how much of it is priced with a margin of safety built in, in case perfection does not show up on schedule. That distinction, more than any prediction about where AI earnings land in 2028, is the one decision actually available to you today.

I will leave you with this. The market has not broken. The trend has not broken. Credit has not broken. I am not telling you to sell everything and hide in a bunker, because that has been terrible advice at every point in my career when someone has offered it to me.

I am telling you that the tallest tower in this market has been built on a foundation that has not yet been tested by a real storm, and that a great many of the companies pouring the concrete are also lending each other the money to buy it.

Keep your outfielders in position. Keep your discipline. And when everybody else is leaning left on this story, it remains, as it always has, worth your time to look right.