Dan Niles Says AI Hasn't Peaked. His Cash Advice Says Something Else.

I've followed Dan Niles long enough to know he doesn't usually hedge. The former hedge fund manager turned research guy plants a flag. So when I saw he's spent the better part of a month telling clients to sit on more cash, I figured something was up.

Here's the part that makes it interesting. He doesn't think the AI trade is dead. He says so directly, and he points to agentic AI as the reason.

The setup

Agentic AI is the shorthand for models that do things instead of just talking about them. Booking travel, writing and testing code, running multi-step workflows without a human babysitting every turn. That's a real shift, and Niles is right that it extends the runway. A chatbot answers a question once. An agent burns tokens in a loop until the job is done.

So the thesis holds. The timing is where he gets nervous.

His caution lands just days before Microsoft, Alphabet, and Amazon report earnings. That matters more than it sounds. Those three are the demand side of the AI chip trade. When they guide capex up, semis go up. When they hesitate, the whole complex wobbles.

And the wobble already started. The PHLX Semiconductor Index dropped 3.2% last week while the S&P 500 gave back 0.5%. That's more than six times the decline. Granted, one week isn't a trend. But chips falling faster than the broad tape usually means somebody out there's trimming expectations before the numbers go public.

The 1998 comparison is the part people keep skimming past. Back then the internet thesis was correct. Dead right, in fact. And the Nasdaq still fell roughly 30% from July to October of that year. Right idea, brutal entry point. History suggests otherwise only if you think this time the drawdown skips a step.

Color me skeptical, but I don't think it does.

What it means past the headline

Agentic AI has a margin problem nobody's talking about enough. Agents multiply inference calls. More inference means more compute, more power, more depreciation on GPU fleets. That cost hits the income statement before the revenue from all those autonomous workflows shows up. Proponents wave it away. The math says otherwise.

For crypto, the transmission channel is simple. AI tokens and decentralized compute markets don't have their own demand curve. They rent one from Nvidia's stock chart. If semis crack, those tokens crack harder. Admittedly, that cuts both ways when the cycle turns.

My read

The 1998 analogy is the most useful thing Niles has said, and also the most dangerous. In 1998 you had to sit through a nasty drawdown and then wait nearly two more years for the blowoff top. Being right about the technology didn't protect you from being wrong about the price.

So which one actually determines your returns. If the thesis is correct and the entry point is wrong, what did the thesis get you.

I don't think a correction is imminent, not yet. But the risk is real, and holding some dry powder isn't cowardice. The question worth asking: are you buying AI because you believe it, or because it's gone up. Those are different positions, and only one of them survives a drawdown.

What to watch next: capex guidance from Microsoft, Alphabet, and Amazon, not the headline earnings beat. Watch the ratio between the SOX and the S&P. And listen closely to any commentary on depreciation and gross margins tied to GPU fleets, because that's where the agentic story either gets funded or stalls. Time will tell, though.