Leaked Video Shows RTX 5090s Dismantled En Masse for AI Servers in Chinese Factories

A leaked video reportedly filmed inside a Chinese factory appears to show the industrial-scale dismantling of consumer RTX 5090 graphics cards for conversion into AI server accelerators, offering a stark visual explanation for the card's near-total absence from retail shelves worldwide. The footage, posted by HKEPC on X and reported by Tom's Hardware, depicts workers systematically stripping three-fan coolers and heatsinks from what looks like hundreds, possibly thousands, of brand-new GPUs before stacking the bare PCBs for dual-slot cooler transplants. For anyone who has spent months refreshing retailer pages or watching prices climb into absurd territory, this video connects the dots between AI infrastructure demand and the empty product pages gamers have been staring at since launch.

The operation captured on camera goes far beyond a simple cooler swap. According to HKEPC, technicians desolder the GPU die and VRAM modules from the original consumer boards and transplant them onto entirely new PCBs designed for server rack density. Once reassembled with dual-slot coolers, the converted cards are validated on vast arrays of benchmarking rigs before entering AI workloads. This is not a handful of enthusiasts modding cards in a garage — it is a pipeline. The scale suggests a coordinated supply chain diverting consumer silicon into the data center market, where margins and volume dwarf the gaming segment entirely.

What Was Leaked

The video originated from HKEPC's X account and was surfaced internationally by Tom's Hardware, showing a facility floor packed with RTX 5090 inventory despite the card not being officially sold in China. Workers remove the elaborate three-fan cooling solutions that define the consumer Founders Edition and partner designs, revealing reference PCBs that are then queued for component-level rework. The desoldering and transplantation of the GB202 GPU and its VRAM onto server-targeted boards implies specialized tooling and skilled labor — this is a repeatable industrial process, not an ad-hoc salvage job. The final frames show rack upon rack of testing stations validating the converted units, confirming the output volume matches the input scale.

Notably, the RTX 5090 has no official retail presence in the Chinese market, yet this facility appears to have secured a massive allocation. That discrepancy points to grey market channels moving pallets of cards from authorized distribution streams directly into conversion shops. HKEPC notes these facilities are scattered throughout the country, suggesting the footage represents a single node in a broader network. If each site processes even a fraction of the volume shown, the aggregate diversion of consumer supply into AI infrastructure could account for a significant portion of the global shortage.

What It Could Mean

Consumers feel the sting instantly: every card that jets off to the pipeline never lands on a gamer's rig. Nvidia keeps saying production is rolling, but if a sizable chunk of output is snatched before it reaches retail shelves, the claim that production is “ongoing” holds in theory while meaning nothing for the average player. The math is clear – an AI server buyer will shell out a premium no gamer can afford, and the cost to convert the hardware is a drop in the bucket compared to the markup. Unless Nvidia or its board partners tighten channel controls or hand out dedicated server SKUs that cover the demand, the grey market will continue to exploit the price difference.

"The horror, the horror, etc."

Andy, PC Gamer

"A gaming GPU for $8,000+ simply makes no sense at all, even if it happens to be Nvidia's top-level card for this generation"

Andy, PC Gamer

"chalk up another way AI server demand is making 2026 a year to forget for consumer PC hardware."

Andy, PC Gamer

We've stopped tracking the GPU in recent months on our GPU price watch page, as the only US listings we can find are from third-party sellers for vastly inflated prices. When a major outlet abandons price tracking because the market has detached from reality, the signal is clear: the consumer launch has effectively failed.

History and Context

This is not the first time high-end consumer GPUs have been repurposed for compute workloads. During the cryptocurrency mining booms of 2017–2018 and 2020–2021, similar diversion occurred, though typically at the hands of individual miners buying retail stock rather than factory-level rework operations. The difference now is the sophistication: these are not cards running in open-air frames with riser cables. They are being surgically rebuilt for density, thermals, and reliability in server racks — a process that requires capital equipment and engineering support. That shift from retail arbitrage to supply chain interception raises the stakes considerably.

HKEPC has a track record of sourcing hardware leaks from the Asian supply chain, and Tom's Hardware's decision to amplify the footage lends it additional credibility. But the video remains unverified by Nvidia or its board partners, and the exact locations, volumes, and commercial relationships involved are unconfirmed. Nvidia has not commented on the specific allegations. The next official milestone to watch is Computex in June, where Nvidia typically addresses channel strategy and may unveil dedicated server products that could alleviate pressure on consumer SKUs — if the company chooses to prioritize that segment.

The footage ultimately reframes the RTX 5090 shortage not as a production shortfall but as a demand collision. AI infrastructure buyers have the capital to outbid the entire consumer market, and a network of conversion facilities has emerged to bridge the form-factor gap. Until that structural imbalance shifts — through policy, dedicated silicon, or a cooling of AI capital expenditure — the flagship gaming GPU risks becoming a theoretical product: announced, benchmarked, and then immediately swallowed by the data center. Gamers waiting for a fair price should not hold their breath.