How is GPU recycling priced differently from regular IC recycling?

GPU/AI cards are priced by "path": the same card is valued separately as whole-card refurb, GDDR die recovery, or board materials — a 3-10x spread. Regular IC recycling mainly tracks package, grade, and refurb-channel resale; GPUs add power-bin quality, multi-card consistency, mining history, and BIOS/memory mods. Functional flagship cards go the refurb route, dead cards are stripped for GDDR dies first, with metals as the floor.

How do I tell if a used GPU was a mining card?

Three signals. Visually, mining cards show discolored chokes and PCB staining. Behaviorally, they artifact under load, reboot randomly, and throw memory errors. And a 30-minute full-load burn-in will expose erratic temperature curves. Mining vs personal-use cards typically differ by 20%+ in price — always disclose mining status before trading.

What do tested pulled GDDR dies fetch now?

October 2026 reference: GDDR6 8Gb pull dies run roughly CNY 3-8 each, GDDR6X 8Gb CNY 5-10, with another 10-20% premium when shipped with test records. An A100 80GB carries 80 on-board dies — on dead cards, die recovery value often beats scrap value, making it the steadiest-margin route in GPU recycling.

How far behind NVIDIA are domestic cards like HUAWEI Ascend 910B on residual value?

2026 H2 reference: a verified Ascend 910B whole card runs CNY 40k-60k, about 60-70% of the previous-gen NVIDIA flagship; Cambricon MLU590 is CNY 15k-25k with strong inference cost efficiency. Domestic cards must be graded by host architecture and software-stack compatibility rather than a single NVIDIA-style price table.

Do you issue test reports for recycled GPUs?

Yes. ChipReclaim performs visual inspection, full-load burn-in, functional and multi-card consistency verification on whole cards, and individually tests every GDDR die with results retained. Function cards shipped with test records carry a 10-20% premium; reports are issued against the original part number with any modifications noted.

IGBT功率模块回收和普通芯片回收有什么不一样？

两类差异：一是形态——IGBT模块是大功率封装，内部是IGBT芯片+FRD续流二极管+铜基板DBC+键合铝线，不是单颗小封装，测试靠万用表二极管档与双脉冲测试台，而不是芯片测试座；二是渠道——普通芯片流向维修与翻新市场，IGBT模块流向功率模块翻新、电驱维修与变频器维修渠道，炸管件按废料铜基板计价。两者混料报价会损失大量利润。

IGBT功率模块回收价大概多少？

2026年参考区间（好模块）：车规英飞凌HybridPACK Drive FS820R08A6P2B约120-220元/块、FS400R07A1E3约60-110元；工业模块FF600R17ME4约80-150元、FF400R12KT4约50-100元；三菱CM400DY-24A约60-110元、CM600DY-24A约90-160元；单管IGBT（TO-247）IKW75N60T约3-6元/颗；国产同规格模块按英飞凌6-7折计价；炸管件5-15元/块按废料铜基板回收。

怎么判断IGBT模块是好的还是炸管的？

三步：一看外观——烧痕、外壳鼓包爆裂、引脚发黑的直接废料档；二用万用表二极管档测FRD续流二极管正反向压降，再测栅极-发射极，压降异常或开路短路的降档；三上双脉冲测试台实测开关特性，能正常开关、无拖尾的按好模块价，带测试记录的好模块可溢价15-25%。拆解时避免撬伤DBC基板与键合线，模块完整度直接决定价格。

国产IGBT模块和英飞凌/三菱的回收价差多少？

同规格国产模块（斯达GD系列、时代电气、士兰微、宏微）按英飞凌的6-7折计价，好批次成色完好的甚至更高。2026年国产IGBT回收量已占三成以上且持续上升：国产化率提升意味着供给更充足、价格更透明，出货渠道也更稳定——车规模块里HybridPACK Drive仍是流通量最大的单品，但国产模块是增长最快的档位。

GPU & AI Accelerator Recycling 2026 H2: NVIDIA A100/H100 Refurbished Residual Values, HUAWEI Ascend & Cambricon Substitution Premium, and GDDR6 Die Recovery

From 3-Year to 15-Month Refresh Cycles: What Really Drives the GPU Retirement Wave

In 2026, AI accelerator refresh cycles have collapsed. As NVIDIA's Blackwell ramps up and Hopper steps aside, and as domestic Chinese AI chips fill the export-control gap, data centers are replacing cards roughly 15 months after purchase — long before their 3-year depreciation schedules expire. The result is a flood of A100/H800 inventory hitting secondary markets. Recyclers see it directly: GPU-related inquiries nearly doubled quarter-over-quarter in Q3 2026, making AI accelerators the highest-value, most volatile category in the reverse supply chain — above even server memory.

GPU recycling is not simply "buy and resell." The same card can be worth 3-10x more on one path than another: whole-card refurbishing, GDDR die recovery, or PCB precious-metal refining. Based on October 2026 market data, this article lays out reference residual values, inspection essentials, and teardown economics.

Reference Residual Values — Mainstream GPUs & AI Accelerators (CNY, Oct 2026)

Model	Condition	Reference Price (CNY)
NVIDIA H100 SXM	80GB, fully functional	60,000-90,000
NVIDIA H800 SXM	80GB, complete	55,000-80,000
NVIDIA A100 80GB	PCIe, boots & stable	22,000-35,000
NVIDIA A800 80GB	PCIe, functional	20,000-30,000
NVIDIA RTX 4090	Clean, no mining	8,500-11,500
NVIDIA RTX 3090	24GB, functional	2,800-3,800
HUAWEI Ascend 910B	Whole card, verified	40,000-60,000
Cambricon MLU590	SW590, tested	15,000-25,000
GDDR6 8Gb die (pull)	Tested OK	CNY 3-8 per die
GDDR6X 8Gb die (pull)	Tested OK	CNY 5-10 per die

Note: Ranges are October 2026 industry references. Actual pricing varies with condition, batch, export-control policy, and domestic supply; always confirm a live quote.

Three Diversion Paths: Whole-Card Refurb, GDDR Die Recovery, Materials

Path 1 — Whole-card refurbishing (highest value): Fully functional H100/A100 units pass burn-in and multi-card consistency checks, then flow into small/medium AI compute operators, university labs, and Southeast Asia compute hosting. Cards shipped with a test record command a 10-20% premium over bare cards.
Path 2 — GDDR die recovery (steadiest margin): Dies from dead or missing-memory boards are reworked and individually tested; good GDDR6/GDDR6X dies sell at CNY 3-10 each into GPU repair and memory-upsize shops. An A100 80GB carries 80 on-board dies — on a dead card, die recovery value often exceeds the card's scrap value.
Path 3 — Board materials (floor price): Power stages, tantalum caps, gold fingers, copper, and palladium/gold in the PCB are recovered at metal-market rates. Lowest unit price, but the essential safety net for dead stock.

The Domestic Substitution Premium: Ascend & Cambricon Upside

Under export controls, domestic Chinese AI chips have moved from "temporary fallback" to "must-have." A HUAWEI Ascend 910B commands roughly 60-70% of the previous-gen NVIDIA flagship's residual value on secondary channels, while Cambricon MLU590 offers clear cost efficiency in inference workloads. Two implications for recyclers: first, domestic-card volumes keep rising through 2026 and require verification skills tuned to multi-card consistency and heterogeneous co-packaging; second, domestic cards are priced by host architecture and software-stack compatibility, not by a single across-the-board table.

Buyer's Checklist: Three Moves Before Buying Used GPUs

Run the temperature curve: 30 minutes of full-load burn-in. Erratic thermals or memory errors knock 30%+ off the price.
Check mining wear: Long low-voltage mining leaves discolored chokes and PCB; typical symptoms are artifacting and random reboots. "Mined" vs "personal-use" cards differ by 20%+ in price.
Verify mods: Cards with replaced coolers, reflashed BIOS, or reworked memory must pass functional testing against the original part number, with modifications noted on the test report.

The harder AI hardware turns over, the more undervalued retired cards become. Whoever turns an "obsolete" A100 into 80 usable GDDR dies, or routes a verified Ascend 910B into a mid-size compute cluster, captures the fattest slice of the 2026 H2 AI hardware cycle.

ChipReclaim (Mr. Wu, 188-2424-1693) specializes in GPU/AI accelerator, MCU/FPGA/IGBT/RF chip and full-category electronics recycling — whole-card functional verification, GDDR die sorting, board-level material recovery, and compliant redistribution end to end.