Grayscale's October 23 Note Splits the AI Trade Into Three Separate Bets

Grayscale published a research note on October 23 that does what most AI-token coverage avoids. It pulls the label apart and looks at what's actually underneath.

The note, led by research head Zach Pandl, walks through three examples. DATA handles contributed training data. USD.AI runs GPU-backed lending. Pearl is trying to pair AI computing with blockchain rewards. Three names, three business models, one crowded narrative.

The Sequence, in Order

Start with what the note says and what it doesn't. Pandl frames these as selected examples and expectations, not as proven outcomes for an entire sector. That qualifier carries weight. Grayscale isn't claiming these three have durable unit economics. It's claiming they occupy distinct roles in the buildout, and that each one should be judged on its own terms.

Then the USD.AI documentation makes the credit piece explicit. Loans are secured by GPUs and the cash flows attached to them. Lenders generally get recourse to the collateral itself, plus a few named carve-outs. That's a familiar structure dressed in new clothes.

The structure employs a split that holders should read carefully. There's USDai, which doesn't bear yield. There's sUSDai, which accrues income from loan payments and reserve-related activity. And there's CHIP, the governance piece. Three instruments, three sets of rights. Naming the protocol tells you almost nothing about which exposure you actually hold.

Bitdeer's mining and AI disclosures fit the same pattern. Machines, funded capacity and realized customer revenue get reported side by side. They aren't interchangeable. One is a capital number. One is an operating number. One is the only figure that tells you whether a customer paid.

Where the Risk Actually Sits

Here's the part that deserves more attention. GPU-backed lending imports credit risk into a blockchain wrapper. But the wrapper doesn't remove the risk. It makes the claim easier to see.

Three questions decide whether these loans hold up. What happens if the equipment loses value faster than the amortization schedule assumes? How concentrated are the borrowers paying to use that capacity? And who can actually seize and sell a depreciating GPU rack after a default?

A ledger can improve visibility into a claim. It can't answer any of those three. That's underwriting, and underwriting is a human business with a long, unglamorous history.

My read: most of the AI-token category gets sorted by one question. Does the underlying business have a customer who pays cash? Not a customer farming a reward. A customer who renews.

Take data networks. Paying contributors can bootstrap supply. Demand is a different problem entirely. Buyers have to find that data lawful, relevant and hard to obtain elsewhere. Token incentives can fund the first six months of activity. They can't fund the seventh if nobody is paying full price.

So who wins? Operators with real revenue lines and disclosed customer concentration. Who loses? Anything whose only revenue is emissions. That split is coming, and it's coming faster than the sector's marketing suggests.

What to Watch Next

Three things, and all of them are measurable.

First, loan performance at USD.AI. Watch for disclosure on defaults, recoveries and collateral marks. If the first stress test arrives and losses are contained, that's a real data point. If disclosure goes quiet, that's a data point too.

Second, revenue mix at the listed miners and GPU landlords. According to 13F filings, institutional positioning in these names will surface in mid-February. That filing window is the cleanest read on whether Wall Street is treating AI capacity as infrastructure or as a trade.

Third, incentive schedules. The moment emissions taper is the moment you learn which data networks and computing markets have genuine demand. Wall Street is moving. Quietly. But it's moving toward cash flow, not toward narratives.

Grayscale's note isn't a buy list. It's a checklist, and the order matters. Service first. Credit second. Token rights third. The first transaction of its kind won't be the one that prints the highest yield. It'll be the one where a borrower repays on schedule and a customer renews without a subsidy attached.

Until that happens, the AI label is doing a lot of work it hasn't earned.