The scene
To understand where we are, look at what's coming public. SpaceX went to market this year, heavily advertising their role as an AI infrastructure company, rather than just a rocket company. And the main event hasn't even shown up yet. The Anthropic and OpenAI IPOs are still ahead. Those are the actual service layer of AI, which is a different thing from the silicon and steel being poured to serve them. That distinction matters more than almost anything else in this market, and I keep coming back to it. There's the handful of companies building the intelligence, and there's the giant complex building the infrastructure underneath it. So far the market has paid lavishly for the second one while it waits to price the first.
The race that built too much
We're living through one of the most violent capital reallocations in modern market history. Over the past twelve months, anything with a plausible claim to "AI infrastructure" got bid into the stratosphere. The numbers say it better than I can:
| Company | Role | ~1-year move |
|---|---|---|
| Sandisk | Flash storage | +4,781% |
| Bloom Energy | Datacenter power | +1,490% |
| Western Digital | Storage | +1,135% |
| Micron | Memory | +880% |
| Seagate | Storage | +735% |
| Intel | Silicon (turnaround) | +568% |
| AMD | Accelerators | +330% |
| HPE | Servers | +169% |
| TSMC | Foundry | +123% |
| Broadcom | Networking/ASICs | +56% |
| NVIDIA | GPUs | +45% |
There's a clean logic to the ordering, and it's worth sitting with. The biggest movers aren't the household names. They're the picks and shovels of the picks and shovels, which is to say storage, memory and power. NVIDIA, the company everyone associates with the AI trade, is up "only" 45%, because it was already enormous and already priced for a great deal of success. The crazy gains went instead to the second and third order suppliers. The people who sell the flash that the GPU servers write to, and the fuel cells that keep the buildings powered when the grid can't.
That pattern, money cascading outward and downward from the obvious winners into their less-obvious suppliers, is itself a classic late-cycle signature. When the front-line names are "fully valued," capital doesn't leave. It hunts for the next derivative exposure. By 2026 it has run out of derivatives that aren't datacenter-adjacent.
Two things are true at once
The hard part of thinking clearly about this moment is holding two ideas in tension without collapsing into either cynicism or euphoria.
The first one is that AI is a real, fundamental change to the world economy. This isn't 1999 tulips. The technology works, it's being adopted, and the productivity story is genuine. There will be two or three dominant platform winners, and they'll be among the most valuable companies ever built.
The second one is that the infrastructure to serve that future is being built faster than the demand to fill it. Capacity is being laid down on the assumption that everyone building it will be a winner. Most won't. The datacenters, the storage arrays, the power contracts, they're real assets. But real assets bought at bubble prices against speculative demand curves are exactly the thing that gets stranded in a consolidation.
Reconciling those two truths is the whole game. The future arrives and most of the capital deployed to meet it is impaired. Railroads electrified the continent and most railroad investors got wiped out. Fiber wired the internet and most of the fiber companies went bankrupt, and then the cable got bought for cents on the dollar by the survivors. The asset survives. The first owner rarely does.
The likely endgame for the infrastructure buildout looks the same to me. The eventual service winners, flush with the appreciated currency of their own equity, spend that haussed money to buy depreciated infrastructure off the overbuilders. The compute gets used. The original balance sheets that funded it don't get made whole.
Meanwhile, the wallflowers
While the infrastructure complex levitated, a whole tier of genuinely dominant internet-era businesses sat the dance out, or actively went backwards:
| Company | ~1-year move |
|---|---|
| SAP | −47% |
| Salesforce | −42% |
| Netflix | −40% |
| Microsoft | −23% |
| Meta | −17% |
| IBM | −10% |
There's something almost ridiculous here. Microsoft, which is basically the commercial AI distribution channel of this era, is down 23%, while a flash-memory supplier is up nearly 5,000%. Whatever you believe about fair value, that spread is not a steady state. It's a stretched rubber band.
These aren't failing companies. They're durable franchises with real cash flows that just didn't fit the single narrative the market wanted to express. Which is exactly what makes them the natural reservoir for the next move. Not because they'll necessarily outperform operationally, but because they're where capital can hide and still feel like it's "in the market."
Three ways this resolves
I don't think the future is knowable here, and I'm deeply skeptical of anyone who claims to know the timing. But the shape of the outcomes is more legible than the schedule. I see three broad paths.
Path 1, the classic crash
A broad global drawdown of 20 to 25%, sticking around a year or more. The infrastructure names lead the fall because that's where the leverage, the speculation and the most heroic demand assumptions are concentrated. This is the textbook resolution, and it's the one every veteran's instinct reaches for. It's real. But it's also the most anticipated outcome, and heavily anticipated outcomes have a way of not arriving on schedule.
Path 2, the rotation, not the rout
This is the path I find most underappreciated, and possibly the most likely. Money doesn't leave the market, it migrates. Capital rotates out of the parabolic winners and into the beaten-down durable names and the classic defensives. The index itself doesn't fall 50%. It "rumbles," churns, and grinds sideways while the internal composition violently reshuffles.
In this world the headline pain is concentrated in the 2025 and 2026 champions, while the broad tape gets held up by an influx into safer ground. The important bit is that the defensive names can rise just from the flow itself, from being the destination of rotating capital, more or less independent of whether their fundamentals improved at all. Valuation here is a function of where the money has to go, not of earnings surprises.
The candidates for that reservoir are exactly the names that anchor every conservative mandate:
- Staples: Walmart, Nestlé, Coca-Cola, Kraft Heinz
- Payments rails: Visa, Mastercard
- The sat-out tech franchises: Microsoft, Meta, the others above
- Classic safety: gold and energy, but, and this matters, not all of it. The thesis is that capital flows into fundamental goods and services broadly, not a panicked dash into bullion alone.
The tell for this path would be the big institutional allocators, the largest banks and asset managers, pre-positioning into safety ahead of a crash they also fear, in a way that weirdly prevents the index-level crash from fully happening. They rotate early enough and broadly enough that the market gets carried by the rotation even as the former leaders are gutted.
Path 3, the genuine regime change
A shift in the world economy big enough that our current frameworks just don't price it. If AI delivers a step-change in productivity, the relationship between earnings, employment, rates and valuation could get rewritten in ways that make both the bull and bear cases above look quaint. This is the lowest-confidence path precisely because it's the one we're least equipped to model. I name it here not to dismiss it, but to stay humble about the tails.
What I actually believe
A few convictions, held with appropriate humility.
The bubble is real, and the market can still run a lot further. Both at once. Bubbles kill the people who short them early far more reliably than they reward the people who call them. Trying to time the top is more dangerous than being wrong about whether it exists.
The split between infrastructure euphoria and franchise neglect is the defining anomaly of this market, and anomalies that extreme tend to close. They can close by the leaders falling, by the laggards rising, or both.
Path 2, rotation over rout, deserves far more weight than it gets. It's the outcome that lets the great institutional allocators protect themselves without triggering the systemic break they're protecting against. It's the resolution that's in almost everyone's interest to engineer.
The infrastructure assets will get used. The first owners often won't be made whole. The consolidation is the mechanism by which a real technological revolution coexists with the impairment of most of the capital that funded its first buildout.
This is a map of the terrain as it currently is. I think the single most useful posture in a moment like this isn't a prediction, it's a discipline. Respect that the bubble can inflate further, refuse to confuse the durability of the technology with the durability of any given valuation, and watch the internal rotation of the market more closely than its headline level, because that's where this regime shows its hand first.
We'll see.