Chinese Workshops Dismantle RTX 5090 GPUs for AI Servers

Leaked video footage from Chinese conversion facilities confirms that industrial workshops are acquiring brand-new GeForce RTX 5090 graphics cards by the thousands and stripping them down for artificial intelligence compute infrastructure. According to reporting from Tom’s Hardware, specialized technicians systematically desolder Nvidia’s flagship GB202 silicon dies and GDDR7 memory modules from original retail circuit boards, transplanting the components onto custom dual-slot printed circuit boards (PCBs) equipped with blower-style coolers. The operation converts gaming flagships into server-ready hardware, bypassing official supply limitations and escalating worldwide consumer hardware shortages.

Factory-Scale Harvesting of GB202 Silicon

The circulating workshop footage documents an assembly-line environment structured specifically for component salvaging. Workbenches are shown stacked with retail retail packaging and fresh bare cards from major board partners, alongside Nvidia’s own reference Founders Edition models. Rather than operating as informal modifications, the conversion facilities use dedicated workstations with heat guns and thermal rework stations to lift the complex ball grid array (BGA) packages intact.

Once technicians extract the GB202 die and the surrounding GDDR7 memory chips, the silicon is cleaned, re-balled, and soldered onto a specialized PCB. Completed boards are then fitted with dual-slot cooling assemblies featuring thick copper baseplates and rear-exhaust centrifugal blowers. Following assembly, the converted accelerators move directly to continuous open-air validation benches where dozens of modified units run multi-hour compute and graphics stress tests simultaneously before being boxed for shipment to commercial data centers and private AI laboratories.

Thermal and Chassis Constraints in AI Clusters

The primary motivation behind this destructive conversion process is physical form-factor incompatibility. Retail GeForce RTX 5090 graphics cards are engineered for desktop consumer cases, featuring massive axial cooling solutions that span three to four PCIe slots. In high-density multi-GPU enterprise environments, such designs present fatal operational bottlenecks:

• Slot Spacing: Standard 4U server chassis and multi-GPU motherboards require dual-slot spacing to accommodate four to eight accelerator cards side-by-side. Thick consumer cards physically obstruct adjacent PCIe slots.
• Exhaust Direction: Axial triple-fan assemblies disperse hot air laterally into the interior of the system chassis. In high-density server enclosures lacking massive open airflow, recirculated heat causes rapid thermal throttling. Dual-slot blower coolers channel intake air across a vapor chamber and exhaust 100% of the thermal load directly through the rear I/O bracket.
• Memory Expansion: Specialized workshops have expanded beyond simple thermal conversions. Hardware modifications documented on regional marketplaces like Alibaba have produced custom RTX 5090 variants outfitted with 96GB of memory—triple the factory 32GB GDDR7 configuration—offering specialized enterprise clusters substantial memory capacity at roughly 65% of standard retail cost.

Retail Exhaustion and Enterprise Silicon Reallocation

The scale of these conversion lines has exacerbated an already strained international supply chain. Consumer availability has evaporated from authorized distribution channels, driving third-party marketplace pricing for custom partner cards to between $8,500 and $10,000. In domestic consumer channels, prebuilt gaming desktop configurations have occasionally emerged as an unexpected pricing hedge, frequently selling below the standalone price of a grey-market card while retaining vendor system warranties.

Compounding this shortfall, industry reports indicate that Nvidia has suspended standard GeForce RTX 5090 manufacturing lines. Production allocations of the monolithic GB202 die are reportedly shifting to higher-margin commercial products, notably the $16,000 RTX Pro 6000 Blackwell accelerator. Between corporate silicon reallocation at the foundry level and aggressive grey-market purchasing by overseas modification lines, consumer inventory for top-tier silicon faces an extended supply freeze.

What happens next

The continued operation of large-scale conversion factories highlights an enduring structural imbalance: commercial demand for generative AI training and inference silicon continues to outpace enterprise production lines, making high-end consumer GPUs lucrative targets for secondary modification. As long as the market spread between consumer retail pricing and certified enterprise accelerators remains wide, third-party workshops will retain strong financial incentives to harvest and repackage gaming dies.

For enterprise buyers and infrastructure operators, these gray-market blower conversions introduce severe reliability and support liabilities. Transplanted BGAs lose all manufacturer warranty coverage and run without official driver validation or enterprise-grade error-correcting memory guarantees. Meanwhile, consumer PC builders and workstation operators face prolonged component shortages, leaving prebuilt systems and secondary-tier cards like the RTX 5080 as the only accessible avenues for desktop Blackwell hardware in the near term.