Wenyi Huang

Founder & CEO of Mercatus AI @mercatus_ai — the financial layer for AI inference tokens — transparent, liquid, and open to all. Ex-Jump, ex-Meta, ex-Citadel

[Wenyi Huang @harrywy]

We've been building the token layer at Mercatus AI, so possibly an unusual angle.

Every instrument in this thread hedges what compute costs. Nothing here hedges what it sells for.

An inference provider leases GPUs and sells tokens. Offtake, ROFR blocks, @ronitrjain's 3y lease with a put on the back half — those all hedge the lease. None of them touch the revenue, which arrives as token sales at prices they don't set. Same asymmetry on the buy side: the company whose exposure is an API invoice never touches a GPU at all.

@BrettHarrison's critique applies one level down, too. Token bills already have an offtake market — enterprise commit deals — negotiated case by case, opaque, size-gated, no transfer. Two companies pay different rates by size and negotiating power. That's not a market, it's an offtake deal wearing a rate card.

And a token bill isn't a GPU bill with a markup. A lab sets its rate card as a pricing decision — margin, competition, positioning — not as a pass-through of what H100s cost. So a GPU hedge pays when rents rise and does nothing when the vendor reprices. That's not a hedge, it's a correlated bet.

What a book changes: one lot is one model, one month, 1M output-standardized tokens. One public quote, so the discount isn't a function of who negotiated hardest. And you can sell it back before expiry, which no commit deal allows.

Early days, but it's the other leg, not a thinner version of this one.

[Wenyi Huang @harrywy]

@4lili_lili4 We are already doing so🤣 building DC for AI, building compute market for AI. There is no way back!

[Wenyi Huang @harrywy]

We've been building the token layer at Mercatus AI, so possibly an unusual angle.

Every instrument in this thread hedges what compute costs. Nothing here hedges what it sells for.

An inference provider leases GPUs and sells tokens. Offtake, ROFR blocks, @ronitrjain's 3y lease with a put on the back half — those all hedge the lease. None of them touch the revenue, which arrives as token sales at prices they don't set. Same asymmetry on the buy side: the company whose exposure is an API invoice never touches a GPU at all.

@BrettHarrison's critique applies one level down, too. Token bills already have an offtake market — enterprise commit deals — negotiated case by case, opaque, size-gated, no transfer. Two companies pay different rates by size and negotiating power. That's not a market, it's an offtake deal wearing a rate card.

And a token bill isn't a GPU bill with a markup. A lab sets its rate card as a pricing decision — margin, competition, positioning — not as a pass-through of what H100s cost. So a GPU hedge pays when rents rise and does nothing when the vendor reprices. That's not a hedge, it's a correlated bet.

What a book changes: one lot is one model, one month, 1M output-standardized tokens. One public quote, so the discount isn't a function of who negotiated hardest. And you can sell it back before expiry, which no commit deal allows.

Early days, but it's the other leg, not a thinner version of this one.