How will NVIDIA repurpose the AI chips it can't sell to China?

High-Performance Computing Machine Learning, LLMs, & AI
1

Since NVIDIA can’t sell its chips to China, will they repurpose them and sell them to developers?

How would that work? Get some boutique builders like Falcon Northwest or System76 to create some new machines? Or would they just go with mainline suppliers like Dell? Workstations or rack mounts?

What kind of motherboard / system parameters would it take? What would an ideal setup look like for a developer? Or, for a small to medium business?

I have seen a few YouTube videos of repurposed AI accelerators.

AI chips just sitting in warehouse won’t help anyone. And they are losing value by the day.

2

Appending my original post. Still waking up.

For this, you need a system that can scale its BAR to meet the Vram needs. Eypc level CPUs are a hard req at this level.

———-
Old post:

A user on the Gamers Nexus Discord had an interesting thought on this. Some claims are already coming true while others I’m not too sure off, like Nvidia destroying GPUs to keep demand high. That said, Nvidia could try and bake a buyback agreement into new contracts going forward that buyers can only sell back working GPUs to Nvidia in exchange for above market value for their used hardware.

DF — 5/16/26, 3:04 AM
Steve Burke

AI is getting more efficient. More efficient AI is driving more demand for AI, but this effect in not always in sync.

Result of a combination of:

  • S oftware side AI inference efficiency gains

  • More GPU designers and fabs entering into the loop

  • next gen hardware gains AI datacenters will find themselves in a situation where post trainning they will be having idling capacity.

And this idling capacity + next gen hardwares entering datacenters is going to create a perfect storm of current gen enterprise gpu’s getting dumped into retailers like Newegg.

GPU shortage of today is not a permamnent shift. It’s a pendulum, and it will swing in the opposite direction hard. A glut of datacenter gpu firesales is coming.

BUT! THAT’S A GOOD THING FOR US.

AND! What’s good for us is bad for Nvidia profit margins.
This scenario of gpu glut is exactly the opposite of what Nvidia wants.

So what does companies do when their margins are pressured by supply? THEY DESTROY THE PERFECTLY USABLE EXISTING SUPPLY JUST SO THAT THEY CAN KEEP SELLING THE NEW SUPPLY AT EXORBITANT PRICES.

This is what luxury brands do? Where does unsold Gucci or whatever clothes go? They go to incinerator. They would rather burn those overpriced clothes than see people accessing them at affordable prices.

So this is exactly what Nvidia and it’s close partners like CoreWeave are going to do.

BUT! GPUs and luxury clothes are fundamentally different. And GPU buyers and overpriced clothing buyers are also ussually fundamentally different.

GPUs are usefull tools and GPU buyers, at least enthusiast part of them are eagerly reading the news and questioning the situation.

So the question is: Is there a.. some sort of “Preserve Useful Tools Act” or something that would prevent Nvidia&Partners from destroying perfectly usable gpus so that they can sell new ones for a higher price?

What’s stopping them from doing this evil act of burning those beauties just so that they preserve the elevated shortage prices for the next gen?
DF — 5/16/26, 3:23 AM
Average enterprise GPU utilization is stuck at 5%
Organizations that panic-procured H100s under 3–5 year depreciation cycles now own depreciating assets that are mostly idle.
5% GPU utilization: The $401 billion AI infrastructure problem enterprises can't keep ignoring | VentureBeat

Per-token inference prices have fallen between 9× and 900× per year for various performance milestones, with Gartner forecasting a further 90% cost reduction by 2030
NVIDIA’s own Groq 3 LPX rack (announced at GTC 2026) combined with the Vera Rubin GPU promises 35× token efficiency improvement and 10× inference cost reduction on-premise
AI Inference Cost Crisis 2026: Why Your AI Bill Is Exploding

On-device inference (Gemini Nano, Apple Intelligence, Microsoft’s Fara-7B) is removing entire classes of queries from the datacenter entirely
https://medium.com/@milesk_33/the-cost-of-scale-why-2026-may-be-the-year-we-shrink-our-models-7af26155271e

Huawei is on track to capture the largest share of China’s AI chip market in 2026, with sales jumping at least 60% amid demand for domestic alternatives
Huawei could seize China’s AI chip crown in 2026 as Nvidia's H200 shipments stall in regulatory limbo — Beijing pushes homegrown AI hardware dominance in a market projected to hit $67 billion by 2030 | Tom's Hardware

GLUT IS COMING:
A100 prices are in gradual decline as organizations upgrade to Hopper/Blackwell. Used A100 80GB units trade at $12,000–$18,000 (down from new pricing), with further depreciation expected as Hopper/Blackwell expand
Buying or Selling GPUs in 2026: Prices, Tips & Rent vs Sell Analysis | GPUnex Blog

The big enterprise gpu firesale glut is coming because AI datacenters will want to become profitable + next gen hardware will arrive + ai inference software will become more efficient + chinese gpu wave will hit full force. These 4 effects will create the situation of pendulum shifting from gpu shortage to gpu glut.

My personal take: I think that we are starting to see signs of the bubble straining.

  • Most Datacenters have been put on hold/canceled.

  • Open weight models are catching up to closed weights very quickly.

  • Open AI/Anthropic going IPO means they ran out of PE/VC money to burn and need IPO in order to get one more injection of cash.

  • Public pushback, especially from younger generations, for “Ai all the things!” is getting more and more vocal and impossible to ignore.

  • Companies are getting burned on token costs, data spills, HIPAA concerns and pulling back on many Ai initiatives to stop cash burns that show little to negative ROI.

3

A few months ago I read reports of the Chinese car industry over producing cars then scraping them with 0 km drives. All so they can report high production numbers. With the right kind of accounting what seems insane some how can make sense.

My speculation is that PCIE based cards will not hit the market. If anything will escape to ebay etc it will be SXM based gpus

4

Without getting into topics that will spawn debates…

Under current conditions, it’s not unthinkable that parties will look the other way to prevent a total bubble burst. Right now A100s are showing up on ebay but still at $14K, but with enough of them hitting eBay it could dip into the price point where a home lab user that needs to do real work (Or small company that is ok maintaining used hardware) could entertain the idea.

Nvidia and OpenAI needs that not to happen.

5

That suggests an interesting possibility. Selling them to universities and high schools.

A lot of these places already have computer labs for students. A government program to expand access could be useful to both NVIDIA and students.

6

Military drones for the highest bidder

7

Most of the chips it’s sold to datacentres won’t actually make it into service anyway - by the time the buildings are operational and fully powered, the GPUs that have been sold will already be obsolete (in DC terms).

This just adds to that pile of chips that are more likely to be destroyed than not. There’s an outside chance that they’ll manage to repurpose them into a different market, but it’s pretty unlikely they’ll ever end up in a form that enthusiasts will be able to use IMO…especially since, as has been mentioned, any GPUs sold at a low-cost price now will reduce the market for the next-generation units they actually want to maintain the pricing for.

8

Could be. There is another option. The tax write-off option. “Donating“ to a school or university or other 501c non-profit corporation could allow for a dollar-for-dollar write down of the chips without directly affecting the market.

A lot of Open Source projects are non-profit (including mine).

9

Especially true for electric cars and bycicles too. There are fields and fields of abandoned new (but very cheaply made) electric cars and bycicles. Also entire cities worth of buildings.

In their case it is mostly a way to scam either government or investors.

Imho this situation applies to ai hardware too, the actual hardware is just a physical item used as part of a scam, and wont see the actual use it was intended for.

10

“Here, this is the stuff that we are trying to use to replace you. GL;HF!”

I get what you are saying that LLMs are a useful tool that people going into tech should learn. but that’s the issue, it’s tool, not a endgame solution that tech bros are lead to believe.

The other thing is, outside of high-scoped research, those systems don’t do much good and Jenson needs to make leather jacket money. A Uni would need maybe two racks at best.

11

Wasn’t Trump in China like yesterday and Jensen too, and they agreed that China is allowed to buy chips?

1 Like
12

I have only read “H200 chips for select Chinese clients”.

13

Yes but China isn’t buying them. Or; the chinese government isn’t letting companies buy them.

Which makes sense for them long-term and isn’t really a surprise. They want to develop their own chips so they want AI CapEx to flow to chinese chip manufacturers rather than Nvidia & TSMC.

14

From what I’m seeing, they may be more interested in getting into an agreement with AMD over Nvidia due to how much Jenson is a wildcard with the current Admin.

Edit: That said, their long-term goal is to make their own AI chips that are just as good (if not better than) US based options. AMD likely knows this as well. I have my issues with Lisa, but she doesn’t run the ship on ego like Jenson does.