> If NVIDIA's next generations deliver radically more tokens per watt and per dollar,

> an Oracle facility designed, financed, and contracted around an older GPU

> generation can be overtaken before it is fully online or fully monetized.

Every user of Nvidia's gear has thought of this, and Nvidia provides a solution: Newer components are (usually) compatible with older racks, and Nvidia will buy back parts you replace with newer hardware.

For example, let's say you have an NV72 with Blackwell-ultra. Rubin comes out, and you decide you want to upgrade. There is a SKU of Rubin designed specifically to work in the racks shipped with Blackwell. Use it, and you only need to replace the compute board (not the power system, cooling, networking, etc.)

Nvidia will buy back the Blackwell boards you remove (and sell them to someone else, with a warrantee).

If you want to do an upgrade that increases power use, you may have to get a new rack. All of the parts are modular, and all parts can be sold back to Nvidia.

Nvidia does this because they have to. Without it, many customers would consider other vendors when upgrading, or delay purchases in case something better comes along later.

I would be very very curious in some source of these Nvidia buybacks

> you only need to replace the compute board

"Only" is doing a lot of heavy lifting here, considering that this compute board makes up the majority of the cost. And if they start going with bigger designs, like they’re doing now (compute in memory, custom TPUs etc), then even this won't hold.

> If you want to do an upgrade that increases power use, you may have to get a new rack.

Blackwell to Rubin upgrade is exactly that case. 3x3U 110 kW units vs the 8×1U 33 kW shelves typical in GB300 NVL72, power whip rating rises to 100A from 60A, rack coolant flow doubles versus GB300, and both compute and NVSwitch trays are fanless, 190-230 kW per rack depending on power mode vs about 140 kW for GB300 NVL72 and NVLink switching is new too etc

And where is your source about buyback program? I found no Nvidia trade in or buyback program for data center hardware. Used gear is resold through third parties, ITAD vendors like Procurri and Bitpro buy and resell used GPUs from hyperscalers, and resellers like Alta Technologies sell refurbished units with warranties.

The closest real thing is a financing arrangement announced in August where Nvidia partnered with Apollo, BlackRock, Blackstone, and others and in SOME CASES may provide residual value support for up to 25% of an opportunity, assessed project by project. So in practice, if a borrower defaults and the lender has to liquidate GPUs that can't fetch their booked value, Nvidia covers part of the shortfall. That backstops lenders. It is not a buyback for customers who upgrade.

What does Nvidia do with the parts they buy back? Resell to others or grind to dust?

The real shame with all of this is that when (not if) the AI bubble pops, it's going to be like the 1990's, when there was a huge supply of unix workstations, etc.

The difference is that none of this Nvidia gear can run on normal 120V (or even 240V) circuits, so it'll end up going to the landfill or just being powered down in place.

I'm not really sure they have thought this through. Or at least their solution has been to essentially insulate the company and even the physical building from the debt used to buy the GPUs. These SPVs typically lease the building and secure the debt to the GPUs themselves so if it all goes pear-shaped, it'll be the debtholders who are left holding the bag, priamrily.

But I have questions:

1. How often do GPUs fail? I've seen figures that this is as high as 9% annual failure rate for DC GPUs;

2. How easy is it to get replacement stock? At what cost?

3. If you trade in, what value do you get for your GPUs?

4. What is the availability like? How does NVidia decide who gets the next wave of GPUs? That's NVidia's big power here. It's a bit like how Boeing and Airbus decide on customer order and if you upset them, you may lose other orders. This happened to Qatar Airways over the A380 [1].

5. How long are the current form factors and interfaces going to last? How long before the power requirements get larger than what DCs are currently built for? Blackwell GPUs use a lot more power than, say, 2 or even 1 generation ago did.

GPUs are a huge cost factor, allegedly more than 50% of the cost of a data center [2]. If you take a bath on the trade in and are forced to do it because you can't compete with the current generation's performance-per-watt, that's not a good situation to be in.

[1]: https://www.arabnews.com/business/qatar-airways-in-mystery-a...

[2]: https://orennia.com/insights/what-it-actually-costs-to-build...