The Buyback Is the Tell

A supplier running roughly 75% gross margins does not offer to buy back its own unsold product because it's feeling generous. It does it because, at the price it needs, the demand isn't standing on its own — so it steps in and holds the risk itself to keep the machine running. Nvidia just did exactly that, and the buyback guarantee is the most honest number in the whole announcement.

Here's the mechanism, stripped down. On July 1 Nvidia unveiled a revenue-sharing and credit-support model with two neocloud partners, Sharon AI and Firmus, scaling toward something like 210,000 Grace Blackwell GB300 GPUs between them. The providers get the hardware without carrying the full upfront capital load. In return Nvidia collects its usual chip revenue plus a recurring cut of whatever cloud income those GPUs eventually generate. And underneath it, the load-bearing piece: if a partner can't fill its racks with paying tenants, Nvidia buys the unsold capacity back at an agreed price. It absorbs the demand risk.

Read those three moves together and the shape is unmistakable. Nvidia is financing the customers who buy Nvidia hardware, then reaching into their revenue on the way out. Michael Burry and Jim Chanos both called it what it looks like — a company funding purchases of its own product — and the analogy everyone reached for is dot-com vendor financing, Lucent and Nortel lending telecoms the money to buy Lucent and Nortel gear right up until the music stopped.

That analogy is the easy read. It's also where most people stop, and stopping there misses the more interesting thing.

The one risk only Nvidia can price

Sit with the buyback for a second, because it's stranger than the circular-financing headline makes it sound. Nvidia is writing an insurance policy on the residual value of a GPU. Promise to repurchase capacity at a set price two or three years out and you've made a bet on what that silicon is worth down the road — which means you've made a bet on how fast it goes obsolete.

Now ask who controls that obsolescence schedule.

Nvidia does. It sets the cadence. Rubin lands, the generation-over-generation efficiency jump is real, and the moment it ships the older chip stops being worth premium power for frontier work — I've made the case before that a GPU's useful life is an opinion, not a measurement. The company underwriting the residual value is the same company that decides, by shipping the next thing, when that residual value collapses. It's insuring a fire while holding the matches.

That's not a flaw in the plan. It's the whole reason the plan can exist. No bank, no leasing shop, no outside lender can price a GB300 buyback three years out, because none of them know the roadmap. Only Nvidia does. So the buyback isn't reckless — it's the one financial product in this market that Nvidia is uniquely equipped to sell, precisely because it owns both ends of the trade. The risk it's absorbing is a risk it also authors.

And there's a genuine counterargument I won't wave away, because it's the strongest thing the defenders have. This isn't dark fiber. The telecom bust financed capacity that sat black, lit at a few percent of what got built. AI compute is being consumed right now, hard, at real utilization. Tokens are getting generated and paid for today. If the demand is actually there, the buyback is a backstop that almost never gets called, and the revenue share is just Nvidia getting paid twice for making the market possible. That's a coherent story. It might even be the right one.

But notice what the strong version concedes. The whole structure only earns its keep if utilization holds. Every part of it — the deferred payment, the revenue cut, the repurchase promise — is a bet that the racks stay full. Nvidia isn't building this because filling the racks is a sure thing. You don't backstop a certainty.

What actually changed

Step back to the system level. Nvidia used to be a component vendor. You bought the chip, the risk was yours, the relationship ended at the loading dock. After July 1 it's something else — supplier, lender, revenue partner, and buyer-of-last-resort to the same counterparty, all at once. If you're a neocloud in this program, Nvidia now sits at four different points in your capital stack, and every one of them is a place where its interests and yours can quietly diverge.

That's the part worth carrying out of this week. The dependency story in AI keeps getting deeper and less visible. First it was "you rent the model and someone else holds the off-switch." Now it's "you rent the compute, from a financier who's also your supplier, your revenue partner, and the only party that knows when your collateral turns into a paperweight." The lock-in didn't go away. It moved up the stack and put on a suit.

The buildout is real and the compute is genuinely getting used. Both things can be true while a 75%-margin supplier decides the market won't clear on its own and steps in to make it clear. When the seller has to guarantee the resale to move the inventory, the guarantee is the data point. Everything else is the pitch.

When the house starts covering your losses, stop admiring the generosity and go find out what it knows about the odds.

— Dustin