Leaked footage shows Chinese factories converting RTX 5090s for AI
Footage shared by HKEPC reveals industrial-scale operations in China where technicians are dismantling hundreds of new NVIDIA RTX 5090 graphics cards. The process involves removing the GB202 silicon and GDDR7 memory chips to transplant them onto custom printed circuit boards. These modified cards feature dual-slot blower coolers designed specifically for dense AI server racks, allowing them to…
Key points
- Chinese factories are desoldering RTX 5090 chips to build custom AI server cards.
- Modified cards sell for $8,500 to $10,000, far exceeding standard retail prices.
- NVIDIA reportedly halted RTX 5090 production to prioritize professional GPUs.
The source of such a large volume of consumer hardware remains unclear, especially given US export restrictions on the RTX 5090. Reports suggest buyers are sourcing cards globally in bulk to bypass these limits. The converted units retain the original 32GB of memory, as no verified 64GB or 96GB versions exist. Companies prefer this route because the RTX 5090 offers high performance at a fraction of the cost of restricted enterprise accelerators, such as the $16,000 RTX PRO 6000.
This diversion has severe consequences for the consumer market. Modified cards sell for roughly $8,500 to $10,000 on secondary markets, making them more profitable for AI labs than standard retail sales. Consequently, US retail stock has been nearly wiped out. NVIDIA has reportedly halted RTX 5090 production to prioritize higher-margin professional GPUs, and has not announced measures to stop the conversions. Gamers face a shrinking supply and rising prices as high-end hardware is cannibalized for server use.
The story so far
6 episodes →- Leaked footage shows Chinese factories converting RTX 5090s for AIthis story
Leaked footage shows Chinese factories stripping hundreds of RTX 5090s for AI use
tweaktown.com · 25 October 2026
Loading the full article…
This text was published by tweaktown.com and written by https://www.facebook.com/hamid.murtaza.5. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Hardware & Compute
All →- TechCrunch lists 10 AI wearables including Bee, Friend, and Rabbit R1 · 1 src
- Nvidia CEO Jensen Huang says quantum computers could help train AI · 1 src
- Nvidia explores converting decommissioned German nuclear plants into AI data centers · 1 src
- Microsoft cancels UK AI data center project due to power delays · 2 src
- Cerebras explains how to run 3T models on WSE-3T chips · 1 src
Comments
via GitHub Discussions