Equipment catalogue

What Qing knows how to buy — and what it will refuse to

categories
1478
freely routable
1283
no export gate on the item itself
classify before outreach
141
control depends on measured parameters
never routed
42
advanced-computing ECCNs

42 categories are refused at the catalogue, not escalated

A PRC-incorporated entity cannot lawfully be the seller of an advanced-computing item (ECCN 3A090 / 4A090) to any destination worldwide, non-US manufacture does not escape the Foreign Direct Product Rule, and every credible PRC accelerator vendor is Entity Listed. Qing refuses at the catalogue rather than opening a conversation and escalating, because the conversation itself creates a record. See the compliance perimeter →

12 near-duplicate pairs need curating — they are why a matched requirement sometimes comes back ambiguous

The importer collapses pure reorderings and empty qualifiers automatically. It will not decide whether “(x86)” distinguishes a product — that depends on whether an ARM variant exists, and guessing wrong destroys a real distinction. These are left for a human.

  • 1.001U dual-socket rack server (x86)↔2U dual-socket rack server (x86)
  • 1.001U tool-less ball-bearing slide rail kit↔2U tool-less ball-bearing slide rail kit
  • 1.00Active optical cable (AOC), 400G/800G↔Active optical cable (AOC), 400G/800G, 3-30 m
  • 1.00Cable management arm (CMA) 1U/2U↔Cable management arm (CMA)
  • 1.00Enterprise NVMe SSD, U.2/E1.S/E3.S, 3.84-30.72 TB↔Enterprise NVMe SSD, E3.S
  • 1.00Enterprise NVMe SSD, U.2/U.3↔Enterprise NVMe SSD, U.2/E1.S/E3.S, 3.84-30.72 TB
  • 1.00Enterprise NVMe SSD, U.2/U.3↔Enterprise NVMe SSD, E3.S
  • 1.00General-purpose 2U dual-socket rack server↔1U dual-socket rack server
  • 1.00High-capacity QLC NVMe SSD↔E1.L NVMe QLC high-capacity SSD
  • 1.00Type 2 surge protective device↔Surge protective device, Type 1+2 and Type 2
  • 0.88Rear-door heat exchanger, active (fan-assisted)↔Rear-door heat exchanger, active (fan-assisted), 60-100 kW
  • 0.88Single-phase immersion dielectric fluid, synthetic ester base↔Single-phase immersion dielectric fluid (synthetic hydrocarbon / ester)
All 1478Cooling 296Power 270IT hardware 240Facility systems 212White space 198Optics & cabling 111Compute 72Consumables 49Prefabricated 30

Optics & cabling

Each carries the questions that make two quotes comparable, and the traps specific to it

showing 60
  • 100/200/400GbE NIC100G/200G/400G网卡
    Optics & cablingHS 8517.62per pcCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • 'OEM' means the opposite of what you think. A Chinese seller saying OEM almost always means 代工/贴牌 — they will build or relabel to your spec, with no brand IP behind it. A Western buyer hears 'made by Dell/HPE'. Never accept the word; ask which factory designed the board and which built it, and get FCC ID / CMIIT ID and PCB silkscreen photos.
    • Engineering samples sold as production silicon. ES/QS CPUs are typically 30-50% below retail, carry no vendor warranty, and can be unstable in specific workloads. Intel states the S-spec is printed only on the IHS, that 'Intel Confidential' or a missing brand string identifies an ES, and that it may retain or destroy re-marked product. QS is nearer retail but is still a sample. Demand IHS photos and an ES/QS exclusion clause with lot-level return.
    • 拆机件 sold as new. Decommissioned DDR4 RDIMMs, RAID cards and drives are a huge, largely legitimate market — the fraud is the relabelling. Refurbishers are documented as resurfacing bad sectors, zeroing power-on hours and swapping model labels. Require original sealed packaging, single-lot date codes, and per-unit health logs.
    • SMART is not proof of a new drive. Commercial tools reset SMART attributes to factory defaults in one click. On Seagate drives the FARM log is a second, harder-to-scrub record: FARM hours exceeding SMART hours indicates tampering. On SSDs, firmware can misreport capacity and health outright, so a clean SMART page means very little.
    • Altered labels defeat verification portals. Scammers copy serials from newer drives onto old units and reprint QR codes so that vendor verification appears to pass. Verify serials yourself at the OEM portal (verify.seagate.com, cway.cisco.com/sncheck, Intel's FPO/ATPO lookup), and cross-check that the serial's warranty start date is consistent with a genuinely new unit.
    • Remarked memory. A DDR5-4800 module relabelled as 6400, or SK hynix dies under a Samsung label, is caught only by comparing three layers: printed module label, SPD dump, and the die markings on the chips themselves. Any disagreement across those three is disqualifying.
    • Vendor-locked parts that boot nowhere. AMD PSB-locked EPYCs, crossflashed RAID cards and OEM-keyed drives are common in pulled stock and are functionally worthless outside their original platform. Ask about lock status explicitly; a seller who does not understand the question is a reseller, not a builder.
  • 100GbE dual-port QSFP28 adapter100G 双口 QSFP28 网卡
    Optics & cablingHS 8517.62per pcCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • 100G leaf switch (32xQSFP28)100G 汇聚交换机 32口
    Optics & cablingHS 8517.62per pcCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • 100G QSFP28 optical transceiver100G光模块
    Optics & cablingHS 8517.62per pcCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • 'OEM' means the opposite of what you think. A Chinese seller saying OEM almost always means 代工/贴牌 — they will build or relabel to your spec, with no brand IP behind it. A Western buyer hears 'made by Dell/HPE'. Never accept the word; ask which factory designed the board and which built it, and get FCC ID / CMIIT ID and PCB silkscreen photos.
    • Engineering samples sold as production silicon. ES/QS CPUs are typically 30-50% below retail, carry no vendor warranty, and can be unstable in specific workloads. Intel states the S-spec is printed only on the IHS, that 'Intel Confidential' or a missing brand string identifies an ES, and that it may retain or destroy re-marked product. QS is nearer retail but is still a sample. Demand IHS photos and an ES/QS exclusion clause with lot-level return.
    • 拆机件 sold as new. Decommissioned DDR4 RDIMMs, RAID cards and drives are a huge, largely legitimate market — the fraud is the relabelling. Refurbishers are documented as resurfacing bad sectors, zeroing power-on hours and swapping model labels. Require original sealed packaging, single-lot date codes, and per-unit health logs.
    • SMART is not proof of a new drive. Commercial tools reset SMART attributes to factory defaults in one click. On Seagate drives the FARM log is a second, harder-to-scrub record: FARM hours exceeding SMART hours indicates tampering. On SSDs, firmware can misreport capacity and health outright, so a clean SMART page means very little.
    • Altered labels defeat verification portals. Scammers copy serials from newer drives onto old units and reprint QR codes so that vendor verification appears to pass. Verify serials yourself at the OEM portal (verify.seagate.com, cway.cisco.com/sncheck, Intel's FPO/ATPO lookup), and cross-check that the serial's warranty start date is consistent with a genuinely new unit.
    • Remarked memory. A DDR5-4800 module relabelled as 6400, or SK hynix dies under a Samsung label, is caught only by comparing three layers: printed module label, SPD dump, and the die markings on the chips themselves. Any disagreement across those three is disqualifying.
    • Vendor-locked parts that boot nowhere. AMD PSB-locked EPYCs, crossflashed RAID cards and OEM-keyed drives are common in pulled stock and are functionally worthless outside their original platform. Ask about lock status explicitly; a seller who does not understand the question is a reseller, not a builder.
  • 100G QSFP28 SR4 optical transceiver100G QSFP28 SR4 光模块
    Optics & cablingHS 8517.62per moduleCN supply: dominant
    unrestricted
    7 ways buyers get burned here
    • Assuming 'not US-origin' means 'not subject to the EAR'. The Advanced Computing FDPR at 734.9(h) captures foreign-produced 3A090/4A090/.z items made with listed US technology or software, or on a plant derived from it. That covers essentially all leading-edge logic on earth. De minimis analysis never even begins for an FDPR-captured item.
    • Assuming a Chinese trading company can lawfully resell a controlled accelerator abroad. It cannot: 742.6(b)(10)(iii) is presumption of denial for all reexports and for D:5/Macau-headquartered entities, and License Exception ACA at 740.8 expressly excludes entities headquartered in — or ultimately parented in — D:5 or Macau. There is no structure that fixes this.
    • Reading the January 2026 case-by-case rule as 'H200 is now sellable to China'. It permits a LICENSED US EXPORTER to apply, with four certifications including a 50%-of-US-volume cap and mandatory independent US-lab pre-export verification. It changed nothing about reexport from China, and left presumption of denial fully intact for D:5/Macau-headquartered buyers.
    • Believing the AI Diffusion Rule was rescinded and its machinery is gone. It was never rescinded in the Federal Register: 740.27 (AIA), 740.28 (ACM), 740.29 (LPP), Supplement 5 to Part 740 and the 790,000,000 TPP country allocations are all still codified at 2026-09-01. BIS announced non-enforcement of its NEW requirements in May 2025, but the 2026-05-31 guidance confirms the pre-existing November-2023 headquarters-based licence requirement IS enforced, and the 2026-06-17 FAQ extends it to .z items.
    • Treating the Affiliates Rule as dead. It is STAYED, not repealed — 2025-19846 stays the 2025-09-30 IFR until 2026-11-09, roughly two months from now. Build the 50%-ownership aggregation logic now behind a flag. Note also that FDPR footnote 5 already carries its own surviving 50%-ownership provision, and OFAC's own 50% Rule never went away.
    • Screening the brand instead of the legal entity. Entity List designations are entity-specific: 'Inspur (Beijing) Electronic Information Industry' is listed while other Inspur group companies are not, and several distinct Fudan Microelectronics entities were listed separately in September 2025. Screen the unified social credit code on the contract, plus every alias.
    • Ignoring alias sprawl. The September 2025 additions show entities with up to sixteen aliases, and the four NetForward / Nanfei entities were listed across four cities. A single-string name match will miss most of them.
  • 100G QSFP28 transceiver (SR4/CWDM4/LR4)100G QSFP28 光模块
    Optics & cablingHS 8517.62per pcCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • 100G single-lambda transceiver (DR/FR)100G 单波 DR/FR 光模块
    Optics & cablingHS 8517.62per pcCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • 10/25GbE SFP28 network adapter万兆/25G网卡
    Optics & cablingHS 8517.62per pcCN supply: strong
    unrestricted
    7 ways buyers get burned here
    • 'OEM' means the opposite of what you think. A Chinese seller saying OEM almost always means 代工/贴牌 — they will build or relabel to your spec, with no brand IP behind it. A Western buyer hears 'made by Dell/HPE'. Never accept the word; ask which factory designed the board and which built it, and get FCC ID / CMIIT ID and PCB silkscreen photos.
    • Engineering samples sold as production silicon. ES/QS CPUs are typically 30-50% below retail, carry no vendor warranty, and can be unstable in specific workloads. Intel states the S-spec is printed only on the IHS, that 'Intel Confidential' or a missing brand string identifies an ES, and that it may retain or destroy re-marked product. QS is nearer retail but is still a sample. Demand IHS photos and an ES/QS exclusion clause with lot-level return.
    • 拆机件 sold as new. Decommissioned DDR4 RDIMMs, RAID cards and drives are a huge, largely legitimate market — the fraud is the relabelling. Refurbishers are documented as resurfacing bad sectors, zeroing power-on hours and swapping model labels. Require original sealed packaging, single-lot date codes, and per-unit health logs.
    • SMART is not proof of a new drive. Commercial tools reset SMART attributes to factory defaults in one click. On Seagate drives the FARM log is a second, harder-to-scrub record: FARM hours exceeding SMART hours indicates tampering. On SSDs, firmware can misreport capacity and health outright, so a clean SMART page means very little.
    • Altered labels defeat verification portals. Scammers copy serials from newer drives onto old units and reprint QR codes so that vendor verification appears to pass. Verify serials yourself at the OEM portal (verify.seagate.com, cway.cisco.com/sncheck, Intel's FPO/ATPO lookup), and cross-check that the serial's warranty start date is consistent with a genuinely new unit.
    • Remarked memory. A DDR5-4800 module relabelled as 6400, or SK hynix dies under a Samsung label, is caught only by comparing three layers: printed module label, SPD dump, and the die markings on the chips themselves. Any disagreement across those three is disqualifying.
    • Vendor-locked parts that boot nowhere. AMD PSB-locked EPYCs, crossflashed RAID cards and OEM-keyed drives are common in pulled stock and are functionally worthless outside their original platform. Ask about lock status explicitly; a seller who does not understand the question is a reseller, not a builder.
  • 10GbE dual-port SFP+ network adapter10G 双口 SFP+ 网卡
    Optics & cablingHS 8517.62 / 8473.30contestedper pcCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • 10G SFP+ transceiver (SR/LR)10G SFP+ 光模块
    Optics & cablingHS 8517.62per pcCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • 10G ToR switch (48xSFP+ / 6xQSFP28)10G 接入交换机
    Optics & cablingHS 8517.62per pcCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • 1.6T OSFP / OSFP-XD transceiver1.6T OSFP 光模块
    Optics & cablingHS 8517.62per pcCN supply: moderate
    classify first

    OSFP vs OSFP-RHS (riding heat sink) is a switch-cage decision made before the optics order

    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • 1.6T OSFP-XD DR8+ optical transceiver1.6T OSFP-XD DR8+ 光模块
    Optics & cablingHS 8517.62per moduleCN supply: dominant
    unrestricted
    7 ways buyers get burned here
    • Assuming 'not US-origin' means 'not subject to the EAR'. The Advanced Computing FDPR at 734.9(h) captures foreign-produced 3A090/4A090/.z items made with listed US technology or software, or on a plant derived from it. That covers essentially all leading-edge logic on earth. De minimis analysis never even begins for an FDPR-captured item.
    • Assuming a Chinese trading company can lawfully resell a controlled accelerator abroad. It cannot: 742.6(b)(10)(iii) is presumption of denial for all reexports and for D:5/Macau-headquartered entities, and License Exception ACA at 740.8 expressly excludes entities headquartered in — or ultimately parented in — D:5 or Macau. There is no structure that fixes this.
    • Reading the January 2026 case-by-case rule as 'H200 is now sellable to China'. It permits a LICENSED US EXPORTER to apply, with four certifications including a 50%-of-US-volume cap and mandatory independent US-lab pre-export verification. It changed nothing about reexport from China, and left presumption of denial fully intact for D:5/Macau-headquartered buyers.
    • Believing the AI Diffusion Rule was rescinded and its machinery is gone. It was never rescinded in the Federal Register: 740.27 (AIA), 740.28 (ACM), 740.29 (LPP), Supplement 5 to Part 740 and the 790,000,000 TPP country allocations are all still codified at 2026-09-01. BIS announced non-enforcement of its NEW requirements in May 2025, but the 2026-05-31 guidance confirms the pre-existing November-2023 headquarters-based licence requirement IS enforced, and the 2026-06-17 FAQ extends it to .z items.
    • Treating the Affiliates Rule as dead. It is STAYED, not repealed — 2025-19846 stays the 2025-09-30 IFR until 2026-11-09, roughly two months from now. Build the 50%-ownership aggregation logic now behind a flag. Note also that FDPR footnote 5 already carries its own surviving 50%-ownership provision, and OFAC's own 50% Rule never went away.
    • Screening the brand instead of the legal entity. Entity List designations are entity-specific: 'Inspur (Beijing) Electronic Information Industry' is listed while other Inspur group companies are not, and several distinct Fudan Microelectronics entities were listed separately in September 2025. Screen the unified social credit code on the contract, plus every alias.
    • Ignoring alias sprawl. The September 2025 additions show entities with up to sixteen aliases, and the four NetForward / Nanfei entities were listed across four cities. A single-string name match will miss most of them.
  • 1.6T OSFP-XD transceiver (2026 ramp)1.6T光模块 OSFP-XD
    Optics & cablingHS 8517.62per pieceCN supply: strong
    classify first

    Early generation — assume elevated first-year AFR and negotiate an advance-replacement pool rather than a spares purchase.

    7 ways buyers get burned here
    • Filter class fraud: media sold as ePM1 55% (F7) that tests at coarse-filter performance. Demand the ISO 16890 report number, the accredited laboratory and the report date, and spot-test media from the delivered lot — not from the sample.
    • Nominal versus actual filter dimensions: a filter quoted as 600x600 that measures 592x592 against a frame expecting 595x595 leaks unfiltered air around the element and voids the rated efficiency entirely. Always specify actual mm and tolerance.
    • Buying filters on unit price instead of Eurovent energy class: a filter 30 Pa higher in initial pressure drop can cost more in fan energy over one service life than the entire unit price difference. Compare kWh/yr at 0.944 m3/s, then price.
    • PTM7950 and high-end TIM counterfeits: this is the single most counterfeited item in the data-centre consumable catalogue. Relabelled generic phase-change film is endemic on Chinese marketplaces. Buy only via authorised distribution with lot traceability, and verify storage conditions in transit.
    • Coolant supplied as industrial glycol cut with tap water rather than certified DI water, with no inhibitor package. Conductivity on arrival is the tell — test every drum against a stated uS/cm limit and reject on the number, not on the label.
    • Automotive or HVAC inhibited coolant sold as data-centre coolant: silicates and borates drop out and foul cold-plate microchannels months later. Require a written silicate-free, borate-free, nitrite-free, amine-free declaration.
    • Fitting a mixed-bed DI polishing cartridge to an inhibited PG25 loop: the resin strips the inhibitor package out of the fluid and corrosion begins silently. Common and expensive design error, usually made by whoever specified the CDU rather than the fluid.
  • 1G/10G out-of-band management switch带外管理交换机
    Optics & cablingHS 8517.62per unitCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • NEARLINE HDD: there is no such thing as a Chinese nearline HDD manufacturer. The world has exactly three producers — Seagate, Western Digital and Toshiba. Seagate's Wuxi plant is the only foreign HDD assembly site in China and it is Seagate's own. Any PRC counterparty presenting itself as an HDD 'factory' is a distributor, a broker or a relabeller. Buy through the three producers' authorised channel and reconcile every serial number against the producer's own warranty database.
    • REFURBISHED AND RECERTIFIED DRIVES SOLD AS NEW. The dominant HDD fraud in the Shenzhen channel is OEM-pull or recertified drives with SMART power-on-hours reset and relabelled cases. Require SMART attribute dumps (power-on hours, power cycles, reallocated sectors) at incoming inspection on a sampled basis, plus original sealed anti-static packaging and matching carton lot codes.
    • LTO TAPE IS THE SECOND NO-PRC-MANUFACTURER POSITION. Tape drive mechanisms come from IBM; media from Fujifilm and Sony. A Chinese 'tape library manufacturer' is an integrator. Apply the same reasoning as HDD.
    • 'ARM SERVER' AMBIGUITY. A Chinese quote for an ARM server very often means Kunpeng (Huawei HiSilicon) or Phytium — both Entity Listed — not Ampere. Force the exact CPU part number into the RFQ and reject quotes that describe the CPU only as 'domestic ARM' or 'ARM 64-core'.
    • 'BARE SWITCH CHASSIS' MEANS TWO DIFFERENT THINGS. Sometimes it means a complete, ASIC-populated whitebox switch shipped with ONIE and no network OS — still a switch for HS and for encryption purposes. Sometimes it means an empty sheet-metal enclosure with PSU and fan bays and no switching board — a part. The price gap is roughly an order of magnitude and the customs treatment differs. Never accept the phrase without a BOM.
    • REBADGING ACROSS THE ENTITY LIST BOUNDARY. Machines built by listed vendors circulate under unlisted third-party brands, and unlisted brands sometimes subcontract to listed manufacturers. Require the manufacturer's legal entity name and factory address on the commercial invoice and the certificate of origin, not just a brand.
    • NAND PROVENANCE INSIDE ENTERPRISE SSD. A PRC-built enterprise SSD may be assembled by an unlisted company using YMTC NAND, and YMTC has been Entity Listed since December 2022. Ask for the NAND die vendor and the controller vendor by name for every SSD SKU.
  • 1G/10G out-of-band management switch (48x1G + 4x10G)带外管理交换机 48口千兆
    Optics & cablingHS 8517.62per pcCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • 200G/400G Ethernet NIC (non-accelerated)200G/400G 以太网网卡
    Optics & cablingHS 8517.62per cardCN supply: moderate
    classify first

    5A991/5A992 depending on crypto. Distinguish from BlueField-class DPUs, which are .z candidates.

    21 questions ready18 of them decide comparabilitysee the question set →
    7 ways buyers get burned here
    • Assuming 'not US-origin' means 'not subject to the EAR'. The Advanced Computing FDPR at 734.9(h) captures foreign-produced 3A090/4A090/.z items made with listed US technology or software, or on a plant derived from it. That covers essentially all leading-edge logic on earth. De minimis analysis never even begins for an FDPR-captured item.
    • Assuming a Chinese trading company can lawfully resell a controlled accelerator abroad. It cannot: 742.6(b)(10)(iii) is presumption of denial for all reexports and for D:5/Macau-headquartered entities, and License Exception ACA at 740.8 expressly excludes entities headquartered in — or ultimately parented in — D:5 or Macau. There is no structure that fixes this.
    • Reading the January 2026 case-by-case rule as 'H200 is now sellable to China'. It permits a LICENSED US EXPORTER to apply, with four certifications including a 50%-of-US-volume cap and mandatory independent US-lab pre-export verification. It changed nothing about reexport from China, and left presumption of denial fully intact for D:5/Macau-headquartered buyers.
    • Believing the AI Diffusion Rule was rescinded and its machinery is gone. It was never rescinded in the Federal Register: 740.27 (AIA), 740.28 (ACM), 740.29 (LPP), Supplement 5 to Part 740 and the 790,000,000 TPP country allocations are all still codified at 2026-09-01. BIS announced non-enforcement of its NEW requirements in May 2025, but the 2026-05-31 guidance confirms the pre-existing November-2023 headquarters-based licence requirement IS enforced, and the 2026-06-17 FAQ extends it to .z items.
    • Treating the Affiliates Rule as dead. It is STAYED, not repealed — 2025-19846 stays the 2025-09-30 IFR until 2026-11-09, roughly two months from now. Build the 50%-ownership aggregation logic now behind a flag. Note also that FDPR footnote 5 already carries its own surviving 50%-ownership provision, and OFAC's own 50% Rule never went away.
    • Screening the brand instead of the legal entity. Entity List designations are entity-specific: 'Inspur (Beijing) Electronic Information Industry' is listed while other Inspur group companies are not, and several distinct Fudan Microelectronics entities were listed separately in September 2025. Screen the unified social credit code on the contract, plus every alias.
    • Ignoring alias sprawl. The September 2025 additions show entities with up to sixteen aliases, and the four NetForward / Nanfei entities were listed across four cities. A single-string name match will miss most of them.
  • 200GbE single/dual-port QSFP112 adapter200G QSFP112 网卡
    Optics & cablingHS 8517.62per pcCN supply: moderate
    classify first

    Ask for ECCN self-classification in writing at this tier and above

    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • 200G QSFP56 transceiver (SR4/FR4)200G QSFP56 光模块
    Optics & cablingHS 8517.62per pcCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • 25.6T Ethernet switch (Tomahawk 4 class)25.6T 以太网交换机
    Optics & cablingHS 8517.62per unitCN supply: strong
    classify first

    Same analysis. Freely sourceable.

    7 ways buyers get burned here
    • Assuming 'not US-origin' means 'not subject to the EAR'. The Advanced Computing FDPR at 734.9(h) captures foreign-produced 3A090/4A090/.z items made with listed US technology or software, or on a plant derived from it. That covers essentially all leading-edge logic on earth. De minimis analysis never even begins for an FDPR-captured item.
    • Assuming a Chinese trading company can lawfully resell a controlled accelerator abroad. It cannot: 742.6(b)(10)(iii) is presumption of denial for all reexports and for D:5/Macau-headquartered entities, and License Exception ACA at 740.8 expressly excludes entities headquartered in — or ultimately parented in — D:5 or Macau. There is no structure that fixes this.
    • Reading the January 2026 case-by-case rule as 'H200 is now sellable to China'. It permits a LICENSED US EXPORTER to apply, with four certifications including a 50%-of-US-volume cap and mandatory independent US-lab pre-export verification. It changed nothing about reexport from China, and left presumption of denial fully intact for D:5/Macau-headquartered buyers.
    • Believing the AI Diffusion Rule was rescinded and its machinery is gone. It was never rescinded in the Federal Register: 740.27 (AIA), 740.28 (ACM), 740.29 (LPP), Supplement 5 to Part 740 and the 790,000,000 TPP country allocations are all still codified at 2026-09-01. BIS announced non-enforcement of its NEW requirements in May 2025, but the 2026-05-31 guidance confirms the pre-existing November-2023 headquarters-based licence requirement IS enforced, and the 2026-06-17 FAQ extends it to .z items.
    • Treating the Affiliates Rule as dead. It is STAYED, not repealed — 2025-19846 stays the 2025-09-30 IFR until 2026-11-09, roughly two months from now. Build the 50%-ownership aggregation logic now behind a flag. Note also that FDPR footnote 5 already carries its own surviving 50%-ownership provision, and OFAC's own 50% Rule never went away.
    • Screening the brand instead of the legal entity. Entity List designations are entity-specific: 'Inspur (Beijing) Electronic Information Industry' is listed while other Inspur group companies are not, and several distinct Fudan Microelectronics entities were listed separately in September 2025. Screen the unified social credit code on the contract, plus every alias.
    • Ignoring alias sprawl. The September 2025 additions show entities with up to sixteen aliases, and the four NetForward / Nanfei entities were listed across four cities. A single-string name match will miss most of them.
  • 25G/100G top-of-rack switch25G/100G接入交换机（ToR）
    Optics & cablingHS 8517.62per unitCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • NEARLINE HDD: there is no such thing as a Chinese nearline HDD manufacturer. The world has exactly three producers — Seagate, Western Digital and Toshiba. Seagate's Wuxi plant is the only foreign HDD assembly site in China and it is Seagate's own. Any PRC counterparty presenting itself as an HDD 'factory' is a distributor, a broker or a relabeller. Buy through the three producers' authorised channel and reconcile every serial number against the producer's own warranty database.
    • REFURBISHED AND RECERTIFIED DRIVES SOLD AS NEW. The dominant HDD fraud in the Shenzhen channel is OEM-pull or recertified drives with SMART power-on-hours reset and relabelled cases. Require SMART attribute dumps (power-on hours, power cycles, reallocated sectors) at incoming inspection on a sampled basis, plus original sealed anti-static packaging and matching carton lot codes.
    • LTO TAPE IS THE SECOND NO-PRC-MANUFACTURER POSITION. Tape drive mechanisms come from IBM; media from Fujifilm and Sony. A Chinese 'tape library manufacturer' is an integrator. Apply the same reasoning as HDD.
    • 'ARM SERVER' AMBIGUITY. A Chinese quote for an ARM server very often means Kunpeng (Huawei HiSilicon) or Phytium — both Entity Listed — not Ampere. Force the exact CPU part number into the RFQ and reject quotes that describe the CPU only as 'domestic ARM' or 'ARM 64-core'.
    • 'BARE SWITCH CHASSIS' MEANS TWO DIFFERENT THINGS. Sometimes it means a complete, ASIC-populated whitebox switch shipped with ONIE and no network OS — still a switch for HS and for encryption purposes. Sometimes it means an empty sheet-metal enclosure with PSU and fan bays and no switching board — a part. The price gap is roughly an order of magnitude and the customs treatment differs. Never accept the phrase without a BOM.
    • REBADGING ACROSS THE ENTITY LIST BOUNDARY. Machines built by listed vendors circulate under unlisted third-party brands, and unlisted brands sometimes subcontract to listed manufacturers. Require the manufacturer's legal entity name and factory address on the commercial invoice and the certificate of origin, not just a brand.
    • NAND PROVENANCE INSIDE ENTERPRISE SSD. A PRC-built enterprise SSD may be assembled by an unlisted company using YMTC NAND, and YMTC has been Entity Listed since December 2022. Ask for the NAND die vendor and the controller vendor by name for every SSD SKU.
  • 25GbE dual-port SFP28 network adapter25G 双口 SFP28 网卡
    Optics & cablingHS 8517.62per pcCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • 25GbE OCP 3.0 SFF adapter25G OCP 3.0 网卡
    Optics & cablingHS 8517.62per pcCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • 25G SFP28 transceiver (SR/LR)25G SFP28 光模块
    Optics & cablingHS 8517.62per pcCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • 25G ToR switch (48xSFP28 + 8xQSFP28)25G 接入交换机 48口
    Optics & cablingHS 8517.62per pcCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • 400GbE OSFP/QSFP112 adapter400G 网卡
    Optics & cablingHS 8517.62per pcCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • 400G leaf/spine data-centre switch400G数据中心交换机
    Optics & cablingHS 8517.62per unitCN supply: moderate
    classify first

    H3C S9825-64D is a concrete 64x400G reference SKU.

    7 ways buyers get burned here
    • NEARLINE HDD: there is no such thing as a Chinese nearline HDD manufacturer. The world has exactly three producers — Seagate, Western Digital and Toshiba. Seagate's Wuxi plant is the only foreign HDD assembly site in China and it is Seagate's own. Any PRC counterparty presenting itself as an HDD 'factory' is a distributor, a broker or a relabeller. Buy through the three producers' authorised channel and reconcile every serial number against the producer's own warranty database.
    • REFURBISHED AND RECERTIFIED DRIVES SOLD AS NEW. The dominant HDD fraud in the Shenzhen channel is OEM-pull or recertified drives with SMART power-on-hours reset and relabelled cases. Require SMART attribute dumps (power-on hours, power cycles, reallocated sectors) at incoming inspection on a sampled basis, plus original sealed anti-static packaging and matching carton lot codes.
    • LTO TAPE IS THE SECOND NO-PRC-MANUFACTURER POSITION. Tape drive mechanisms come from IBM; media from Fujifilm and Sony. A Chinese 'tape library manufacturer' is an integrator. Apply the same reasoning as HDD.
    • 'ARM SERVER' AMBIGUITY. A Chinese quote for an ARM server very often means Kunpeng (Huawei HiSilicon) or Phytium — both Entity Listed — not Ampere. Force the exact CPU part number into the RFQ and reject quotes that describe the CPU only as 'domestic ARM' or 'ARM 64-core'.
    • 'BARE SWITCH CHASSIS' MEANS TWO DIFFERENT THINGS. Sometimes it means a complete, ASIC-populated whitebox switch shipped with ONIE and no network OS — still a switch for HS and for encryption purposes. Sometimes it means an empty sheet-metal enclosure with PSU and fan bays and no switching board — a part. The price gap is roughly an order of magnitude and the customs treatment differs. Never accept the phrase without a BOM.
    • REBADGING ACROSS THE ENTITY LIST BOUNDARY. Machines built by listed vendors circulate under unlisted third-party brands, and unlisted brands sometimes subcontract to listed manufacturers. Require the manufacturer's legal entity name and factory address on the commercial invoice and the certificate of origin, not just a brand.
    • NAND PROVENANCE INSIDE ENTERPRISE SSD. A PRC-built enterprise SSD may be assembled by an unlisted company using YMTC NAND, and YMTC has been Entity Listed since December 2022. Ask for the NAND die vendor and the controller vendor by name for every SSD SKU.
  • 400G leaf/spine switch (32x400G, 12.8-25.6T)400G 交换机 32口
    Optics & cablingHS 8517.62per pcCN supply: moderate
    classify first

    QSFP-DD vs OSFP cage choice locks your entire optics BOM; decide before RFQ

    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • 400G OSFP transceiver (DR4/2xFR4)400G OSFP 光模块
    Optics & cablingHS 8517.62per pcCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • 400G QSFP112 transceiver400G QSFP112 光模块
    Optics & cablingHS 8517.62per pcCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • 400G QSFP-DD DR4 optical transceiver400G QSFP-DD DR4 光模块
    Optics & cablingHS 8517.62per moduleCN supply: dominant
    unrestricted
    7 ways buyers get burned here
    • Assuming 'not US-origin' means 'not subject to the EAR'. The Advanced Computing FDPR at 734.9(h) captures foreign-produced 3A090/4A090/.z items made with listed US technology or software, or on a plant derived from it. That covers essentially all leading-edge logic on earth. De minimis analysis never even begins for an FDPR-captured item.
    • Assuming a Chinese trading company can lawfully resell a controlled accelerator abroad. It cannot: 742.6(b)(10)(iii) is presumption of denial for all reexports and for D:5/Macau-headquartered entities, and License Exception ACA at 740.8 expressly excludes entities headquartered in — or ultimately parented in — D:5 or Macau. There is no structure that fixes this.
    • Reading the January 2026 case-by-case rule as 'H200 is now sellable to China'. It permits a LICENSED US EXPORTER to apply, with four certifications including a 50%-of-US-volume cap and mandatory independent US-lab pre-export verification. It changed nothing about reexport from China, and left presumption of denial fully intact for D:5/Macau-headquartered buyers.
    • Believing the AI Diffusion Rule was rescinded and its machinery is gone. It was never rescinded in the Federal Register: 740.27 (AIA), 740.28 (ACM), 740.29 (LPP), Supplement 5 to Part 740 and the 790,000,000 TPP country allocations are all still codified at 2026-09-01. BIS announced non-enforcement of its NEW requirements in May 2025, but the 2026-05-31 guidance confirms the pre-existing November-2023 headquarters-based licence requirement IS enforced, and the 2026-06-17 FAQ extends it to .z items.
    • Treating the Affiliates Rule as dead. It is STAYED, not repealed — 2025-19846 stays the 2025-09-30 IFR until 2026-11-09, roughly two months from now. Build the 50%-ownership aggregation logic now behind a flag. Note also that FDPR footnote 5 already carries its own surviving 50%-ownership provision, and OFAC's own 50% Rule never went away.
    • Screening the brand instead of the legal entity. Entity List designations are entity-specific: 'Inspur (Beijing) Electronic Information Industry' is listed while other Inspur group companies are not, and several distinct Fudan Microelectronics entities were listed separately in September 2025. Screen the unified social credit code on the contract, plus every alias.
    • Ignoring alias sprawl. The September 2025 additions show entities with up to sixteen aliases, and the four NetForward / Nanfei entities were listed across four cities. A single-string name match will miss most of them.
  • 400G QSFP-DD / OSFP transceiver, DR4 / FR4 / SR8 / LR4400G光模块 QSFP-DD/OSFP DR4/FR4
    Optics & cablingHS 8517.62per pieceCN supply: dominant
    classify first

    Treat as a consumable, not an asset: field AFR is around 1-2%/yr for mature generations and materially higher in the first year of a new one. China (Innolight, Eoptolink, Accelink, HGTech) is the global volume source.

    7 ways buyers get burned here
    • Filter class fraud: media sold as ePM1 55% (F7) that tests at coarse-filter performance. Demand the ISO 16890 report number, the accredited laboratory and the report date, and spot-test media from the delivered lot — not from the sample.
    • Nominal versus actual filter dimensions: a filter quoted as 600x600 that measures 592x592 against a frame expecting 595x595 leaks unfiltered air around the element and voids the rated efficiency entirely. Always specify actual mm and tolerance.
    • Buying filters on unit price instead of Eurovent energy class: a filter 30 Pa higher in initial pressure drop can cost more in fan energy over one service life than the entire unit price difference. Compare kWh/yr at 0.944 m3/s, then price.
    • PTM7950 and high-end TIM counterfeits: this is the single most counterfeited item in the data-centre consumable catalogue. Relabelled generic phase-change film is endemic on Chinese marketplaces. Buy only via authorised distribution with lot traceability, and verify storage conditions in transit.
    • Coolant supplied as industrial glycol cut with tap water rather than certified DI water, with no inhibitor package. Conductivity on arrival is the tell — test every drum against a stated uS/cm limit and reject on the number, not on the label.
    • Automotive or HVAC inhibited coolant sold as data-centre coolant: silicates and borates drop out and foul cold-plate microchannels months later. Require a written silicate-free, borate-free, nitrite-free, amine-free declaration.
    • Fitting a mixed-bed DI polishing cartridge to an inhibited PG25 loop: the resin strips the inhibitor package out of the fluid and corrosion begins silently. Common and expensive design error, usually made by whoever specified the CDU rather than the fluid.
  • 400G QSFP-DD transceiver (DR4/FR4/LR4)400G QSFP-DD 光模块
    Optics & cablingHS 8517.62per pcCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • 400G ZR/ZR+ coherent transceiver400G ZR/ZR+ 相干光模块
    Optics & cablingHS 8517.62per pcCN supply: moderate
    classify first

    ZR and ZR+ are different link budgets; specify amplification and span loss, not just kilometres

    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • 51.2T Ethernet leaf/spine switch (merchant silicon)51.2T 以太网交换机（商用芯片）
    Optics & cablingHS 8517.62per unitCN supply: strong
    classify first

    Normally 5A991 or EAR99; becomes 5A992.z / 5A002.z only if it contains a 3A090-class IC — a merchant Ethernet ASIC does not. Encryption features may pull it into 5A002. Freely sourceable with classification evidence.

    7 ways buyers get burned here
    • Assuming 'not US-origin' means 'not subject to the EAR'. The Advanced Computing FDPR at 734.9(h) captures foreign-produced 3A090/4A090/.z items made with listed US technology or software, or on a plant derived from it. That covers essentially all leading-edge logic on earth. De minimis analysis never even begins for an FDPR-captured item.
    • Assuming a Chinese trading company can lawfully resell a controlled accelerator abroad. It cannot: 742.6(b)(10)(iii) is presumption of denial for all reexports and for D:5/Macau-headquartered entities, and License Exception ACA at 740.8 expressly excludes entities headquartered in — or ultimately parented in — D:5 or Macau. There is no structure that fixes this.
    • Reading the January 2026 case-by-case rule as 'H200 is now sellable to China'. It permits a LICENSED US EXPORTER to apply, with four certifications including a 50%-of-US-volume cap and mandatory independent US-lab pre-export verification. It changed nothing about reexport from China, and left presumption of denial fully intact for D:5/Macau-headquartered buyers.
    • Believing the AI Diffusion Rule was rescinded and its machinery is gone. It was never rescinded in the Federal Register: 740.27 (AIA), 740.28 (ACM), 740.29 (LPP), Supplement 5 to Part 740 and the 790,000,000 TPP country allocations are all still codified at 2026-09-01. BIS announced non-enforcement of its NEW requirements in May 2025, but the 2026-05-31 guidance confirms the pre-existing November-2023 headquarters-based licence requirement IS enforced, and the 2026-06-17 FAQ extends it to .z items.
    • Treating the Affiliates Rule as dead. It is STAYED, not repealed — 2025-19846 stays the 2025-09-30 IFR until 2026-11-09, roughly two months from now. Build the 50%-ownership aggregation logic now behind a flag. Note also that FDPR footnote 5 already carries its own surviving 50%-ownership provision, and OFAC's own 50% Rule never went away.
    • Screening the brand instead of the legal entity. Entity List designations are entity-specific: 'Inspur (Beijing) Electronic Information Industry' is listed while other Inspur group companies are not, and several distinct Fudan Microelectronics entities were listed separately in September 2025. Screen the unified social credit code on the contract, plus every alias.
    • Ignoring alias sprawl. The September 2025 additions show entities with up to sixteen aliases, and the four NetForward / Nanfei entities were listed across four cities. A single-string name match will miss most of them.
  • 800GbE AI back-end adapter (SuperNIC class)800G AI 后端网卡
    Optics & cablingHS 8517.62per pcCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • 800G CPO silicon-photonics switch (51.2T)800G CPO 硅光交换机
    Optics & cablingHS 8517.62per pcCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • 800G linear-drive transceiver (LPO/LRO)800G LPO/LRO 线性直驱光模块
    Optics & cablingHS 8517.62per pcCN supply: moderate
    classify first

    LPO only works against a qualified host SerDes; without an interop list it is unbuyable

    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • 800G LPO / LRO linear-drive optical module800G LPO 线性直驱光模块
    Optics & cablingHS 8517.62per moduleCN supply: dominant
    unrestricted
    7 ways buyers get burned here
    • Assuming 'not US-origin' means 'not subject to the EAR'. The Advanced Computing FDPR at 734.9(h) captures foreign-produced 3A090/4A090/.z items made with listed US technology or software, or on a plant derived from it. That covers essentially all leading-edge logic on earth. De minimis analysis never even begins for an FDPR-captured item.
    • Assuming a Chinese trading company can lawfully resell a controlled accelerator abroad. It cannot: 742.6(b)(10)(iii) is presumption of denial for all reexports and for D:5/Macau-headquartered entities, and License Exception ACA at 740.8 expressly excludes entities headquartered in — or ultimately parented in — D:5 or Macau. There is no structure that fixes this.
    • Reading the January 2026 case-by-case rule as 'H200 is now sellable to China'. It permits a LICENSED US EXPORTER to apply, with four certifications including a 50%-of-US-volume cap and mandatory independent US-lab pre-export verification. It changed nothing about reexport from China, and left presumption of denial fully intact for D:5/Macau-headquartered buyers.
    • Believing the AI Diffusion Rule was rescinded and its machinery is gone. It was never rescinded in the Federal Register: 740.27 (AIA), 740.28 (ACM), 740.29 (LPP), Supplement 5 to Part 740 and the 790,000,000 TPP country allocations are all still codified at 2026-09-01. BIS announced non-enforcement of its NEW requirements in May 2025, but the 2026-05-31 guidance confirms the pre-existing November-2023 headquarters-based licence requirement IS enforced, and the 2026-06-17 FAQ extends it to .z items.
    • Treating the Affiliates Rule as dead. It is STAYED, not repealed — 2025-19846 stays the 2025-09-30 IFR until 2026-11-09, roughly two months from now. Build the 50%-ownership aggregation logic now behind a flag. Note also that FDPR footnote 5 already carries its own surviving 50%-ownership provision, and OFAC's own 50% Rule never went away.
    • Screening the brand instead of the legal entity. Entity List designations are entity-specific: 'Inspur (Beijing) Electronic Information Industry' is listed while other Inspur group companies are not, and several distinct Fudan Microelectronics entities were listed separately in September 2025. Screen the unified social credit code on the contract, plus every alias.
    • Ignoring alias sprawl. The September 2025 additions show entities with up to sixteen aliases, and the four NetForward / Nanfei entities were listed across four cities. A single-string name match will miss most of them.
  • 800G OSFP 2xFR4 optical transceiver800G OSFP 2xFR4 光模块
    Optics & cablingHS 8517.62per moduleCN supply: dominant
    unrestricted
    7 ways buyers get burned here
    • Assuming 'not US-origin' means 'not subject to the EAR'. The Advanced Computing FDPR at 734.9(h) captures foreign-produced 3A090/4A090/.z items made with listed US technology or software, or on a plant derived from it. That covers essentially all leading-edge logic on earth. De minimis analysis never even begins for an FDPR-captured item.
    • Assuming a Chinese trading company can lawfully resell a controlled accelerator abroad. It cannot: 742.6(b)(10)(iii) is presumption of denial for all reexports and for D:5/Macau-headquartered entities, and License Exception ACA at 740.8 expressly excludes entities headquartered in — or ultimately parented in — D:5 or Macau. There is no structure that fixes this.
    • Reading the January 2026 case-by-case rule as 'H200 is now sellable to China'. It permits a LICENSED US EXPORTER to apply, with four certifications including a 50%-of-US-volume cap and mandatory independent US-lab pre-export verification. It changed nothing about reexport from China, and left presumption of denial fully intact for D:5/Macau-headquartered buyers.
    • Believing the AI Diffusion Rule was rescinded and its machinery is gone. It was never rescinded in the Federal Register: 740.27 (AIA), 740.28 (ACM), 740.29 (LPP), Supplement 5 to Part 740 and the 790,000,000 TPP country allocations are all still codified at 2026-09-01. BIS announced non-enforcement of its NEW requirements in May 2025, but the 2026-05-31 guidance confirms the pre-existing November-2023 headquarters-based licence requirement IS enforced, and the 2026-06-17 FAQ extends it to .z items.
    • Treating the Affiliates Rule as dead. It is STAYED, not repealed — 2025-19846 stays the 2025-09-30 IFR until 2026-11-09, roughly two months from now. Build the 50%-ownership aggregation logic now behind a flag. Note also that FDPR footnote 5 already carries its own surviving 50%-ownership provision, and OFAC's own 50% Rule never went away.
    • Screening the brand instead of the legal entity. Entity List designations are entity-specific: 'Inspur (Beijing) Electronic Information Industry' is listed while other Inspur group companies are not, and several distinct Fudan Microelectronics entities were listed separately in September 2025. Screen the unified social credit code on the contract, plus every alias.
    • Ignoring alias sprawl. The September 2025 additions show entities with up to sixteen aliases, and the four NetForward / Nanfei entities were listed across four cities. A single-string name match will miss most of them.
  • 800G OSFP DR8 optical transceiver800G OSFP DR8 光模块
    Optics & cablingHS 8517.62per moduleCN supply: dominant
    unrestricted
    19 questions ready13 of them decide comparabilitysee the question set →
    7 ways buyers get burned here
    • Assuming 'not US-origin' means 'not subject to the EAR'. The Advanced Computing FDPR at 734.9(h) captures foreign-produced 3A090/4A090/.z items made with listed US technology or software, or on a plant derived from it. That covers essentially all leading-edge logic on earth. De minimis analysis never even begins for an FDPR-captured item.
    • Assuming a Chinese trading company can lawfully resell a controlled accelerator abroad. It cannot: 742.6(b)(10)(iii) is presumption of denial for all reexports and for D:5/Macau-headquartered entities, and License Exception ACA at 740.8 expressly excludes entities headquartered in — or ultimately parented in — D:5 or Macau. There is no structure that fixes this.
    • Reading the January 2026 case-by-case rule as 'H200 is now sellable to China'. It permits a LICENSED US EXPORTER to apply, with four certifications including a 50%-of-US-volume cap and mandatory independent US-lab pre-export verification. It changed nothing about reexport from China, and left presumption of denial fully intact for D:5/Macau-headquartered buyers.
    • Believing the AI Diffusion Rule was rescinded and its machinery is gone. It was never rescinded in the Federal Register: 740.27 (AIA), 740.28 (ACM), 740.29 (LPP), Supplement 5 to Part 740 and the 790,000,000 TPP country allocations are all still codified at 2026-09-01. BIS announced non-enforcement of its NEW requirements in May 2025, but the 2026-05-31 guidance confirms the pre-existing November-2023 headquarters-based licence requirement IS enforced, and the 2026-06-17 FAQ extends it to .z items.
    • Treating the Affiliates Rule as dead. It is STAYED, not repealed — 2025-19846 stays the 2025-09-30 IFR until 2026-11-09, roughly two months from now. Build the 50%-ownership aggregation logic now behind a flag. Note also that FDPR footnote 5 already carries its own surviving 50%-ownership provision, and OFAC's own 50% Rule never went away.
    • Screening the brand instead of the legal entity. Entity List designations are entity-specific: 'Inspur (Beijing) Electronic Information Industry' is listed while other Inspur group companies are not, and several distinct Fudan Microelectronics entities were listed separately in September 2025. Screen the unified social credit code on the contract, plus every alias.
    • Ignoring alias sprawl. The September 2025 additions show entities with up to sixteen aliases, and the four NetForward / Nanfei entities were listed across four cities. A single-string name match will miss most of them.
  • 800G OSFP/QSFP-DD optical transceiver800G光模块
    Optics & cablingHS 8517.62per pcCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • 'OEM' means the opposite of what you think. A Chinese seller saying OEM almost always means 代工/贴牌 — they will build or relabel to your spec, with no brand IP behind it. A Western buyer hears 'made by Dell/HPE'. Never accept the word; ask which factory designed the board and which built it, and get FCC ID / CMIIT ID and PCB silkscreen photos.
    • Engineering samples sold as production silicon. ES/QS CPUs are typically 30-50% below retail, carry no vendor warranty, and can be unstable in specific workloads. Intel states the S-spec is printed only on the IHS, that 'Intel Confidential' or a missing brand string identifies an ES, and that it may retain or destroy re-marked product. QS is nearer retail but is still a sample. Demand IHS photos and an ES/QS exclusion clause with lot-level return.
    • 拆机件 sold as new. Decommissioned DDR4 RDIMMs, RAID cards and drives are a huge, largely legitimate market — the fraud is the relabelling. Refurbishers are documented as resurfacing bad sectors, zeroing power-on hours and swapping model labels. Require original sealed packaging, single-lot date codes, and per-unit health logs.
    • SMART is not proof of a new drive. Commercial tools reset SMART attributes to factory defaults in one click. On Seagate drives the FARM log is a second, harder-to-scrub record: FARM hours exceeding SMART hours indicates tampering. On SSDs, firmware can misreport capacity and health outright, so a clean SMART page means very little.
    • Altered labels defeat verification portals. Scammers copy serials from newer drives onto old units and reprint QR codes so that vendor verification appears to pass. Verify serials yourself at the OEM portal (verify.seagate.com, cway.cisco.com/sncheck, Intel's FPO/ATPO lookup), and cross-check that the serial's warranty start date is consistent with a genuinely new unit.
    • Remarked memory. A DDR5-4800 module relabelled as 6400, or SK hynix dies under a Samsung label, is caught only by comparing three layers: printed module label, SPD dump, and the die markings on the chips themselves. Any disagreement across those three is disqualifying.
    • Vendor-locked parts that boot nowhere. AMD PSB-locked EPYCs, crossflashed RAID cards and OEM-keyed drives are common in pulled stock and are functionally worthless outside their original platform. Ask about lock status explicitly; a seller who does not understand the question is a reseller, not a builder.
  • 800G OSFP / QSFP-DD transceiver, 2xDR4 / 2xFR4 / SR8 / DR8800G光模块 OSFP/QSFP-DD
    Optics & cablingHS 8517.62per pieceCN supply: dominant
    classify first

    The largest single recurring spend line in an AI hall — a rail-optimised fabric carries thousands of modules behind a few hundred GPUs. Carry 3-5% spares in year one of a new generation, 1-2% thereafter.

    7 ways buyers get burned here
    • Filter class fraud: media sold as ePM1 55% (F7) that tests at coarse-filter performance. Demand the ISO 16890 report number, the accredited laboratory and the report date, and spot-test media from the delivered lot — not from the sample.
    • Nominal versus actual filter dimensions: a filter quoted as 600x600 that measures 592x592 against a frame expecting 595x595 leaks unfiltered air around the element and voids the rated efficiency entirely. Always specify actual mm and tolerance.
    • Buying filters on unit price instead of Eurovent energy class: a filter 30 Pa higher in initial pressure drop can cost more in fan energy over one service life than the entire unit price difference. Compare kWh/yr at 0.944 m3/s, then price.
    • PTM7950 and high-end TIM counterfeits: this is the single most counterfeited item in the data-centre consumable catalogue. Relabelled generic phase-change film is endemic on Chinese marketplaces. Buy only via authorised distribution with lot traceability, and verify storage conditions in transit.
    • Coolant supplied as industrial glycol cut with tap water rather than certified DI water, with no inhibitor package. Conductivity on arrival is the tell — test every drum against a stated uS/cm limit and reject on the number, not on the label.
    • Automotive or HVAC inhibited coolant sold as data-centre coolant: silicates and borates drop out and foul cold-plate microchannels months later. Require a written silicate-free, borate-free, nitrite-free, amine-free declaration.
    • Fitting a mixed-bed DI polishing cartridge to an inhibited PG25 loop: the resin strips the inhibitor package out of the fluid and corrosion begins silently. Common and expensive design error, usually made by whoever specified the CDU rather than the fluid.
  • 800G OSFP SR8 optical transceiver800G OSFP SR8 多模光模块
    Optics & cablingHS 8517.62per moduleCN supply: dominant
    unrestricted
    7 ways buyers get burned here
    • Assuming 'not US-origin' means 'not subject to the EAR'. The Advanced Computing FDPR at 734.9(h) captures foreign-produced 3A090/4A090/.z items made with listed US technology or software, or on a plant derived from it. That covers essentially all leading-edge logic on earth. De minimis analysis never even begins for an FDPR-captured item.
    • Assuming a Chinese trading company can lawfully resell a controlled accelerator abroad. It cannot: 742.6(b)(10)(iii) is presumption of denial for all reexports and for D:5/Macau-headquartered entities, and License Exception ACA at 740.8 expressly excludes entities headquartered in — or ultimately parented in — D:5 or Macau. There is no structure that fixes this.
    • Reading the January 2026 case-by-case rule as 'H200 is now sellable to China'. It permits a LICENSED US EXPORTER to apply, with four certifications including a 50%-of-US-volume cap and mandatory independent US-lab pre-export verification. It changed nothing about reexport from China, and left presumption of denial fully intact for D:5/Macau-headquartered buyers.
    • Believing the AI Diffusion Rule was rescinded and its machinery is gone. It was never rescinded in the Federal Register: 740.27 (AIA), 740.28 (ACM), 740.29 (LPP), Supplement 5 to Part 740 and the 790,000,000 TPP country allocations are all still codified at 2026-09-01. BIS announced non-enforcement of its NEW requirements in May 2025, but the 2026-05-31 guidance confirms the pre-existing November-2023 headquarters-based licence requirement IS enforced, and the 2026-06-17 FAQ extends it to .z items.
    • Treating the Affiliates Rule as dead. It is STAYED, not repealed — 2025-19846 stays the 2025-09-30 IFR until 2026-11-09, roughly two months from now. Build the 50%-ownership aggregation logic now behind a flag. Note also that FDPR footnote 5 already carries its own surviving 50%-ownership provision, and OFAC's own 50% Rule never went away.
    • Screening the brand instead of the legal entity. Entity List designations are entity-specific: 'Inspur (Beijing) Electronic Information Industry' is listed while other Inspur group companies are not, and several distinct Fudan Microelectronics entities were listed separately in September 2025. Screen the unified social credit code on the contract, plus every alias.
    • Ignoring alias sprawl. The September 2025 additions show entities with up to sixteen aliases, and the four NetForward / Nanfei entities were listed across four cities. A single-string name match will miss most of them.
  • 800G OSFP transceiver (DR8/2xFR4/SR8)800G OSFP 光模块
    Optics & cablingHS 8517.62per pcCN supply: dominant
    unrestricted
    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • 800G QSFP-DD800 transceiver800G QSFP-DD800 光模块
    Optics & cablingHS 8517.62per pcCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • 800G silicon-photonics optical module800G 硅光光模块
    Optics & cablingHS 8517.62per moduleCN supply: strong
    unrestricted
    7 ways buyers get burned here
    • Assuming 'not US-origin' means 'not subject to the EAR'. The Advanced Computing FDPR at 734.9(h) captures foreign-produced 3A090/4A090/.z items made with listed US technology or software, or on a plant derived from it. That covers essentially all leading-edge logic on earth. De minimis analysis never even begins for an FDPR-captured item.
    • Assuming a Chinese trading company can lawfully resell a controlled accelerator abroad. It cannot: 742.6(b)(10)(iii) is presumption of denial for all reexports and for D:5/Macau-headquartered entities, and License Exception ACA at 740.8 expressly excludes entities headquartered in — or ultimately parented in — D:5 or Macau. There is no structure that fixes this.
    • Reading the January 2026 case-by-case rule as 'H200 is now sellable to China'. It permits a LICENSED US EXPORTER to apply, with four certifications including a 50%-of-US-volume cap and mandatory independent US-lab pre-export verification. It changed nothing about reexport from China, and left presumption of denial fully intact for D:5/Macau-headquartered buyers.
    • Believing the AI Diffusion Rule was rescinded and its machinery is gone. It was never rescinded in the Federal Register: 740.27 (AIA), 740.28 (ACM), 740.29 (LPP), Supplement 5 to Part 740 and the 790,000,000 TPP country allocations are all still codified at 2026-09-01. BIS announced non-enforcement of its NEW requirements in May 2025, but the 2026-05-31 guidance confirms the pre-existing November-2023 headquarters-based licence requirement IS enforced, and the 2026-06-17 FAQ extends it to .z items.
    • Treating the Affiliates Rule as dead. It is STAYED, not repealed — 2025-19846 stays the 2025-09-30 IFR until 2026-11-09, roughly two months from now. Build the 50%-ownership aggregation logic now behind a flag. Note also that FDPR footnote 5 already carries its own surviving 50%-ownership provision, and OFAC's own 50% Rule never went away.
    • Screening the brand instead of the legal entity. Entity List designations are entity-specific: 'Inspur (Beijing) Electronic Information Industry' is listed while other Inspur group companies are not, and several distinct Fudan Microelectronics entities were listed separately in September 2025. Screen the unified social credit code on the contract, plus every alias.
    • Ignoring alias sprawl. The September 2025 additions show entities with up to sixteen aliases, and the four NetForward / Nanfei entities were listed across four cities. A single-string name match will miss most of them.
  • 800G spine switch800G骨干交换机
    Optics & cablingHS 8517.62per unitCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • NEARLINE HDD: there is no such thing as a Chinese nearline HDD manufacturer. The world has exactly three producers — Seagate, Western Digital and Toshiba. Seagate's Wuxi plant is the only foreign HDD assembly site in China and it is Seagate's own. Any PRC counterparty presenting itself as an HDD 'factory' is a distributor, a broker or a relabeller. Buy through the three producers' authorised channel and reconcile every serial number against the producer's own warranty database.
    • REFURBISHED AND RECERTIFIED DRIVES SOLD AS NEW. The dominant HDD fraud in the Shenzhen channel is OEM-pull or recertified drives with SMART power-on-hours reset and relabelled cases. Require SMART attribute dumps (power-on hours, power cycles, reallocated sectors) at incoming inspection on a sampled basis, plus original sealed anti-static packaging and matching carton lot codes.
    • LTO TAPE IS THE SECOND NO-PRC-MANUFACTURER POSITION. Tape drive mechanisms come from IBM; media from Fujifilm and Sony. A Chinese 'tape library manufacturer' is an integrator. Apply the same reasoning as HDD.
    • 'ARM SERVER' AMBIGUITY. A Chinese quote for an ARM server very often means Kunpeng (Huawei HiSilicon) or Phytium — both Entity Listed — not Ampere. Force the exact CPU part number into the RFQ and reject quotes that describe the CPU only as 'domestic ARM' or 'ARM 64-core'.
    • 'BARE SWITCH CHASSIS' MEANS TWO DIFFERENT THINGS. Sometimes it means a complete, ASIC-populated whitebox switch shipped with ONIE and no network OS — still a switch for HS and for encryption purposes. Sometimes it means an empty sheet-metal enclosure with PSU and fan bays and no switching board — a part. The price gap is roughly an order of magnitude and the customs treatment differs. Never accept the phrase without a BOM.
    • REBADGING ACROSS THE ENTITY LIST BOUNDARY. Machines built by listed vendors circulate under unlisted third-party brands, and unlisted brands sometimes subcontract to listed manufacturers. Require the manufacturer's legal entity name and factory address on the commercial invoice and the certificate of origin, not just a brand.
    • NAND PROVENANCE INSIDE ENTERPRISE SSD. A PRC-built enterprise SSD may be assembled by an unlisted company using YMTC NAND, and YMTC has been Entity Listed since December 2022. Ask for the NAND die vendor and the controller vendor by name for every SSD SKU.
  • 800G ToR/spine switch (64x800G, 51.2T)800G 交换机 64口 51.2T
    Optics & cablingHS 8517.62per pcCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • Active electrical cable (AEC) / active copper cable, 200-800G有源电缆 AEC
    Optics & cablingHS 8544.42per pcCN supply: strong
    unrestricted
    7 ways buyers get burned here
    • 'OEM' means the opposite of what you think. A Chinese seller saying OEM almost always means 代工/贴牌 — they will build or relabel to your spec, with no brand IP behind it. A Western buyer hears 'made by Dell/HPE'. Never accept the word; ask which factory designed the board and which built it, and get FCC ID / CMIIT ID and PCB silkscreen photos.
    • Engineering samples sold as production silicon. ES/QS CPUs are typically 30-50% below retail, carry no vendor warranty, and can be unstable in specific workloads. Intel states the S-spec is printed only on the IHS, that 'Intel Confidential' or a missing brand string identifies an ES, and that it may retain or destroy re-marked product. QS is nearer retail but is still a sample. Demand IHS photos and an ES/QS exclusion clause with lot-level return.
    • 拆机件 sold as new. Decommissioned DDR4 RDIMMs, RAID cards and drives are a huge, largely legitimate market — the fraud is the relabelling. Refurbishers are documented as resurfacing bad sectors, zeroing power-on hours and swapping model labels. Require original sealed packaging, single-lot date codes, and per-unit health logs.
    • SMART is not proof of a new drive. Commercial tools reset SMART attributes to factory defaults in one click. On Seagate drives the FARM log is a second, harder-to-scrub record: FARM hours exceeding SMART hours indicates tampering. On SSDs, firmware can misreport capacity and health outright, so a clean SMART page means very little.
    • Altered labels defeat verification portals. Scammers copy serials from newer drives onto old units and reprint QR codes so that vendor verification appears to pass. Verify serials yourself at the OEM portal (verify.seagate.com, cway.cisco.com/sncheck, Intel's FPO/ATPO lookup), and cross-check that the serial's warranty start date is consistent with a genuinely new unit.
    • Remarked memory. A DDR5-4800 module relabelled as 6400, or SK hynix dies under a Samsung label, is caught only by comparing three layers: printed module label, SPD dump, and the die markings on the chips themselves. Any disagreement across those three is disqualifying.
    • Vendor-locked parts that boot nowhere. AMD PSB-locked EPYCs, crossflashed RAID cards and OEM-keyed drives are common in pulled stock and are functionally worthless outside their original platform. Ask about lock status explicitly; a seller who does not understand the question is a reseller, not a builder.
  • Active electrical cable (AEC) / active copper cable (ACC), 800G, 3-7 m有源铜缆 AEC/ACC 800G
    Optics & cablingHS 8544.42 / 8517.62contestedper pieceCN supply: strong
    classify first

    Contains active silicon, so it fails like an optic rather than like a cable — budget spares accordingly.

    7 ways buyers get burned here
    • Filter class fraud: media sold as ePM1 55% (F7) that tests at coarse-filter performance. Demand the ISO 16890 report number, the accredited laboratory and the report date, and spot-test media from the delivered lot — not from the sample.
    • Nominal versus actual filter dimensions: a filter quoted as 600x600 that measures 592x592 against a frame expecting 595x595 leaks unfiltered air around the element and voids the rated efficiency entirely. Always specify actual mm and tolerance.
    • Buying filters on unit price instead of Eurovent energy class: a filter 30 Pa higher in initial pressure drop can cost more in fan energy over one service life than the entire unit price difference. Compare kWh/yr at 0.944 m3/s, then price.
    • PTM7950 and high-end TIM counterfeits: this is the single most counterfeited item in the data-centre consumable catalogue. Relabelled generic phase-change film is endemic on Chinese marketplaces. Buy only via authorised distribution with lot traceability, and verify storage conditions in transit.
    • Coolant supplied as industrial glycol cut with tap water rather than certified DI water, with no inhibitor package. Conductivity on arrival is the tell — test every drum against a stated uS/cm limit and reject on the number, not on the label.
    • Automotive or HVAC inhibited coolant sold as data-centre coolant: silicates and borates drop out and foul cold-plate microchannels months later. Require a written silicate-free, borate-free, nitrite-free, amine-free declaration.
    • Fitting a mixed-bed DI polishing cartridge to an inhibited PG25 loop: the resin strips the inhibitor package out of the fluid and corrosion begins silently. Common and expensive design error, usually made by whoever specified the CDU rather than the fluid.
  • Active optical cable (AOC), 400G/800G有源光缆 AOC 400G/800G
    Optics & cablingHS 8544.70contestedper pcCN supply: dominant
    classify first

    Missing — the report has DAC and AEC but no AOC, which is the standard choice beyond DAC reach and below transceiver-plus-fibre cost. T&S 300570 is a verified source. Same host-coding trap as DAC: an AOC coded for the wrong switch vendor is scrap.

    7 ways buyers get burned here
    • The 'manufacturer' that is a trading company. On Alibaba and Made-in-China this is the default state for racks, containment, rack PDUs, cabling and small cooling accessories. The distinguishing evidence: (a) 经营范围 in the business licence says 销售/贸易 but not 生产; (b) the 'factory address' equals the registered office and is in an office tower, not an industrial park; (c) no 生产许可 or product certification is held in the seller's own name — certificates show a different company; (d) they are not in any CIOE/China Refrigeration/CDCC exhibitor list; (e) they quote across five unrelated categories. Always cross-check the 统一社会信用代码 on gsxt.gov.cn.
    • The reverse trap: assuming a trader is always bad. For a 20-line mixed BOM, a competent Guangdong trading company is often the only party that will handle consolidation, export documentation and one invoice. The trap is paying manufacturer margins for trader service, or letting the trader hide who actually made the goods so warranty has no counterparty.
    • Cooling capacity quoted at undeclared conditions. A '100 kW' CRAH at 15/21C chilled water is a different machine from a '100 kW' CRAH at 12/18C, and a '35 kW' passive rear door at 12C water is not the same product as one at 18C. Chinese quotes routinely omit the conditions and Western buyers routinely compare the headline number.
    • CDU kW ratings with only one loop temperature. A CDU is defined by primary in, primary out, secondary in, secondary out and both flow rates. A single number is not a specification and cannot be compared across vendors.
    • Residual air heat in liquid-cooled racks. Vendors state '90% heat captured to liquid' and buyers size the room for 10%. In practice the uncaptured fraction is often 15-25% under real workloads, and a 120 kW rack then dumps 25 kW into a hall designed for 12. Ask for kW of residual air heat per chassis, in writing.
    • Filled floor loading for immersion tanks. A filled immersion tank can exceed 2,000 kg/m2. Chinese quotes give the dry weight. The building, not the tank, is what fails.
    • Aluminium substituted for copper in busbar, busway and switchgear. It is legal, it meets the current rating, and it is not what the buyer priced. Specify conductor material and cross-section, and weigh a sample on arrival.
  • Active optical cable (AOC), 400G/800G, 3-30 m有源光缆 AOC 400G/800G
    Optics & cablingHS 8544.70 / 8517.62contestedper pieceCN supply: dominant
    classify first

    Whole-cable replacement on any end failure, which makes its effective failure cost higher than two transceivers plus a patch cord.

    7 ways buyers get burned here
    • Filter class fraud: media sold as ePM1 55% (F7) that tests at coarse-filter performance. Demand the ISO 16890 report number, the accredited laboratory and the report date, and spot-test media from the delivered lot — not from the sample.
    • Nominal versus actual filter dimensions: a filter quoted as 600x600 that measures 592x592 against a frame expecting 595x595 leaks unfiltered air around the element and voids the rated efficiency entirely. Always specify actual mm and tolerance.
    • Buying filters on unit price instead of Eurovent energy class: a filter 30 Pa higher in initial pressure drop can cost more in fan energy over one service life than the entire unit price difference. Compare kWh/yr at 0.944 m3/s, then price.
    • PTM7950 and high-end TIM counterfeits: this is the single most counterfeited item in the data-centre consumable catalogue. Relabelled generic phase-change film is endemic on Chinese marketplaces. Buy only via authorised distribution with lot traceability, and verify storage conditions in transit.
    • Coolant supplied as industrial glycol cut with tap water rather than certified DI water, with no inhibitor package. Conductivity on arrival is the tell — test every drum against a stated uS/cm limit and reject on the number, not on the label.
    • Automotive or HVAC inhibited coolant sold as data-centre coolant: silicates and borates drop out and foul cold-plate microchannels months later. Require a written silicate-free, borate-free, nitrite-free, amine-free declaration.
    • Fitting a mixed-bed DI polishing cartridge to an inhibited PG25 loop: the resin strips the inhibitor package out of the fluid and corrosion begins silently. Common and expensive design error, usually made by whoever specified the CDU rather than the fluid.
  • Application delivery controller / load balancer appliance应用交付控制器 / 负载均衡设备
    Optics & cablingHS 8517.62per unitCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • NEARLINE HDD: there is no such thing as a Chinese nearline HDD manufacturer. The world has exactly three producers — Seagate, Western Digital and Toshiba. Seagate's Wuxi plant is the only foreign HDD assembly site in China and it is Seagate's own. Any PRC counterparty presenting itself as an HDD 'factory' is a distributor, a broker or a relabeller. Buy through the three producers' authorised channel and reconcile every serial number against the producer's own warranty database.
    • REFURBISHED AND RECERTIFIED DRIVES SOLD AS NEW. The dominant HDD fraud in the Shenzhen channel is OEM-pull or recertified drives with SMART power-on-hours reset and relabelled cases. Require SMART attribute dumps (power-on hours, power cycles, reallocated sectors) at incoming inspection on a sampled basis, plus original sealed anti-static packaging and matching carton lot codes.
    • LTO TAPE IS THE SECOND NO-PRC-MANUFACTURER POSITION. Tape drive mechanisms come from IBM; media from Fujifilm and Sony. A Chinese 'tape library manufacturer' is an integrator. Apply the same reasoning as HDD.
    • 'ARM SERVER' AMBIGUITY. A Chinese quote for an ARM server very often means Kunpeng (Huawei HiSilicon) or Phytium — both Entity Listed — not Ampere. Force the exact CPU part number into the RFQ and reject quotes that describe the CPU only as 'domestic ARM' or 'ARM 64-core'.
    • 'BARE SWITCH CHASSIS' MEANS TWO DIFFERENT THINGS. Sometimes it means a complete, ASIC-populated whitebox switch shipped with ONIE and no network OS — still a switch for HS and for encryption purposes. Sometimes it means an empty sheet-metal enclosure with PSU and fan bays and no switching board — a part. The price gap is roughly an order of magnitude and the customs treatment differs. Never accept the phrase without a BOM.
    • REBADGING ACROSS THE ENTITY LIST BOUNDARY. Machines built by listed vendors circulate under unlisted third-party brands, and unlisted brands sometimes subcontract to listed manufacturers. Require the manufacturer's legal entity name and factory address on the commercial invoice and the certificate of origin, not just a brand.
    • NAND PROVENANCE INSIDE ENTERPRISE SSD. A PRC-built enterprise SSD may be assembled by an unlisted company using YMTC NAND, and YMTC has been Entity Listed since December 2022. Ask for the NAND die vendor and the controller vendor by name for every SSD SKU.
  • ASIC DPU 200G (domestic)国产 200G DPU
    Optics & cablingHS 8517.62per pcCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • Bare-metal ONIE whitebox switch with SONiC白牌裸机交换机 + SONiC
    Optics & cablingHS 8517.62per pcCN supply: moderate
    classify first

    SAI/platform-driver quality is the whole product; ask which SONiC branch is actually tested

    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • Bare switch chassis / whitebox switch (ONIE, no NOS)白盒交换机 / 裸交换机（无网络操作系统）
    Optics & cablingHS 8517.62 / 8473.30 / 8538contestedper unitCN supply: moderate
    classify first

    'Bare switch chassis' means two different things and the catalogue must not conflate them: (a) a fully-functional switch with populated ASIC but no network OS — still a switch for HS purposes, and the standard whitebox purchase; (b) a genuinely empty sheet-metal enclosure with PSU/fan bays and no switching board — that is a part, not a switch. Always ask which is being quoted.

    7 ways buyers get burned here
    • NEARLINE HDD: there is no such thing as a Chinese nearline HDD manufacturer. The world has exactly three producers — Seagate, Western Digital and Toshiba. Seagate's Wuxi plant is the only foreign HDD assembly site in China and it is Seagate's own. Any PRC counterparty presenting itself as an HDD 'factory' is a distributor, a broker or a relabeller. Buy through the three producers' authorised channel and reconcile every serial number against the producer's own warranty database.
    • REFURBISHED AND RECERTIFIED DRIVES SOLD AS NEW. The dominant HDD fraud in the Shenzhen channel is OEM-pull or recertified drives with SMART power-on-hours reset and relabelled cases. Require SMART attribute dumps (power-on hours, power cycles, reallocated sectors) at incoming inspection on a sampled basis, plus original sealed anti-static packaging and matching carton lot codes.
    • LTO TAPE IS THE SECOND NO-PRC-MANUFACTURER POSITION. Tape drive mechanisms come from IBM; media from Fujifilm and Sony. A Chinese 'tape library manufacturer' is an integrator. Apply the same reasoning as HDD.
    • 'ARM SERVER' AMBIGUITY. A Chinese quote for an ARM server very often means Kunpeng (Huawei HiSilicon) or Phytium — both Entity Listed — not Ampere. Force the exact CPU part number into the RFQ and reject quotes that describe the CPU only as 'domestic ARM' or 'ARM 64-core'.
    • 'BARE SWITCH CHASSIS' MEANS TWO DIFFERENT THINGS. Sometimes it means a complete, ASIC-populated whitebox switch shipped with ONIE and no network OS — still a switch for HS and for encryption purposes. Sometimes it means an empty sheet-metal enclosure with PSU and fan bays and no switching board — a part. The price gap is roughly an order of magnitude and the customs treatment differs. Never accept the phrase without a BOM.
    • REBADGING ACROSS THE ENTITY LIST BOUNDARY. Machines built by listed vendors circulate under unlisted third-party brands, and unlisted brands sometimes subcontract to listed manufacturers. Require the manufacturer's legal entity name and factory address on the commercial invoice and the certificate of origin, not just a brand.
    • NAND PROVENANCE INSIDE ENTERPRISE SSD. A PRC-built enterprise SSD may be assembled by an unlisted company using YMTC NAND, and YMTC has been Entity Listed since December 2022. Ask for the NAND die vendor and the controller vendor by name for every SSD SKU.
  • Cat6A/Cat6 patch cords and 24-port RJ45 patch panel六类/超六类跳线与24口配线架
    Optics & cablingHS 8544.42 / 8536.90contestedper pcCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • Transceiver vendor lock is a software problem sold as a hardware problem. The compatibility code sits in the module's EEPROM and is read over I2C by the switch, which compares it before bringing the port up. Chinese factories will burn whatever code you name — so 'is this Cisco compatible' is the wrong question; 'can you code each unit to our exact switch model list and bench-verify' is the right one. Quotes without a coding answer are not comparable.
    • The same trap applies at both ends of every DAC, ACC, AEC and AOC — each end carries its own EEPROM. A cable between a Cisco leaf and an H3C spine needs asymmetric coding, which most buyers never specify and most quotes never mention.
    • The .z sweep makes visual inspection worthless for compliance. BIS created .z paragraphs in 3A001.z, 4A003.z, 4A004.z, 4A005.z, 5A002.z, 5A004.z, 5A992.z, 5D002.z and 5D992.z that capture items purely because they incorporate an IC meeting 3A090/4A090 parameters. Two identical-looking 800G switches, or two SmartNICs in the same chassis, can sit on opposite sides of a licence requirement. Only the internal ASIC/DSP part number resolves it, so the ECCN question must be a hard RFQ field, not a footnote.
    • Lane structure is invisible in a part number and fatal in deployment. An 800G module can be 8x100G or 4x200G; a 400G can be 8x50G (QSFP-DD) or 4x100G (QSFP112). If the lane structure does not match the switch SerDes generation, the link never trains and both parties blame the other's hardware.
    • MPO-12 versus MPO-16 quietly invalidates entire cable plants. 800G DR8 needs MPO-16; most existing trunking is MPO-12. Combine that with UPC-versus-APC end-face and polarity method A/B/C and you have three independent ways to buy a fibre plant that physically connects and optically fails.
    • OSFP flat-top versus finned-top, and OSFP versus OSFP-RHS, are mechanical fit decisions inside the switch cage. Ordering the wrong heatsink profile means modules that will not seat in a switch you have already bought.
    • Passive DAC reach collapses as lane rate rises. What worked at 3-5m at 25G/lane is roughly 2m at 100G/lane. Buyers reuse an old length survey, order passive DAC, and discover mid-install that half the runs need ACC or AEC at several times the price and with an added export-classification question.
  • Cat6A/Cat8 copper patch cords六类/八类网线
    Optics & cablingHS 8544.42per pcCN supply: moderate
    unrestricted
    7 ways buyers get burned here
    • 'OEM' means the opposite of what you think. A Chinese seller saying OEM almost always means 代工/贴牌 — they will build or relabel to your spec, with no brand IP behind it. A Western buyer hears 'made by Dell/HPE'. Never accept the word; ask which factory designed the board and which built it, and get FCC ID / CMIIT ID and PCB silkscreen photos.
    • Engineering samples sold as production silicon. ES/QS CPUs are typically 30-50% below retail, carry no vendor warranty, and can be unstable in specific workloads. Intel states the S-spec is printed only on the IHS, that 'Intel Confidential' or a missing brand string identifies an ES, and that it may retain or destroy re-marked product. QS is nearer retail but is still a sample. Demand IHS photos and an ES/QS exclusion clause with lot-level return.
    • 拆机件 sold as new. Decommissioned DDR4 RDIMMs, RAID cards and drives are a huge, largely legitimate market — the fraud is the relabelling. Refurbishers are documented as resurfacing bad sectors, zeroing power-on hours and swapping model labels. Require original sealed packaging, single-lot date codes, and per-unit health logs.
    • SMART is not proof of a new drive. Commercial tools reset SMART attributes to factory defaults in one click. On Seagate drives the FARM log is a second, harder-to-scrub record: FARM hours exceeding SMART hours indicates tampering. On SSDs, firmware can misreport capacity and health outright, so a clean SMART page means very little.
    • Altered labels defeat verification portals. Scammers copy serials from newer drives onto old units and reprint QR codes so that vendor verification appears to pass. Verify serials yourself at the OEM portal (verify.seagate.com, cway.cisco.com/sncheck, Intel's FPO/ATPO lookup), and cross-check that the serial's warranty start date is consistent with a genuinely new unit.
    • Remarked memory. A DDR5-4800 module relabelled as 6400, or SK hynix dies under a Samsung label, is caught only by comparing three layers: printed module label, SPD dump, and the die markings on the chips themselves. Any disagreement across those three is disqualifying.
    • Vendor-locked parts that boot nowhere. AMD PSB-locked EPYCs, crossflashed RAID cards and OEM-keyed drives are common in pulled stock and are functionally worthless outside their original platform. Ask about lock status explicitly; a seller who does not understand the question is a reseller, not a builder.
  • Coherent 400G ZR / ZR+ transceiver for DCI相干 400G ZR / ZR+ 光模块（数据中心互联）
    Optics & cablingHS 8517.62contestedper pcCN supply: strong
    classify first

    Missing as a line — the report treats ZR only as a reach variant of 400G. DCI is a distinct design problem with distinct suppliers (Accelink 002281, CIG Shanghai 603083, Eoptolink). Power draw versus host cage thermal limit is the practical gate, not price.

    7 ways buyers get burned here
    • The 'manufacturer' that is a trading company. On Alibaba and Made-in-China this is the default state for racks, containment, rack PDUs, cabling and small cooling accessories. The distinguishing evidence: (a) 经营范围 in the business licence says 销售/贸易 but not 生产; (b) the 'factory address' equals the registered office and is in an office tower, not an industrial park; (c) no 生产许可 or product certification is held in the seller's own name — certificates show a different company; (d) they are not in any CIOE/China Refrigeration/CDCC exhibitor list; (e) they quote across five unrelated categories. Always cross-check the 统一社会信用代码 on gsxt.gov.cn.
    • The reverse trap: assuming a trader is always bad. For a 20-line mixed BOM, a competent Guangdong trading company is often the only party that will handle consolidation, export documentation and one invoice. The trap is paying manufacturer margins for trader service, or letting the trader hide who actually made the goods so warranty has no counterparty.
    • Cooling capacity quoted at undeclared conditions. A '100 kW' CRAH at 15/21C chilled water is a different machine from a '100 kW' CRAH at 12/18C, and a '35 kW' passive rear door at 12C water is not the same product as one at 18C. Chinese quotes routinely omit the conditions and Western buyers routinely compare the headline number.
    • CDU kW ratings with only one loop temperature. A CDU is defined by primary in, primary out, secondary in, secondary out and both flow rates. A single number is not a specification and cannot be compared across vendors.
    • Residual air heat in liquid-cooled racks. Vendors state '90% heat captured to liquid' and buyers size the room for 10%. In practice the uncaptured fraction is often 15-25% under real workloads, and a 120 kW rack then dumps 25 kW into a hall designed for 12. Ask for kW of residual air heat per chassis, in writing.
    • Filled floor loading for immersion tanks. A filled immersion tank can exceed 2,000 kg/m2. Chinese quotes give the dry weight. The building, not the tank, is what fails.
    • Aluminium substituted for copper in busbar, busway and switchgear. It is legal, it meets the current rating, and it is not what the buyer priced. Specify conductor material and cross-section, and weigh a sample on arrival.
All sections
China Export中国出口