Most coverage of GPU export controls focuses on buying and deploying chips. Almost none addresses what happens when you retire them. But the same rules that govern importing an H100 govern reselling it, and the retirement side is where many operators have the least visibility. Here’s the ITAD-perspective primer.
Important: This article is educational and does not constitute legal advice. Export control law is complex, fact-specific, and changes frequently. The rules described here were accurate to the best of our understanding as of mid-2026 but may have changed. Always consult qualified export-control counsel before exporting, reexporting, or transferring controlled hardware. ROC Telecom is not a law firm and does not provide legal or export-classification opinions.
TL;DR
When AI infrastructure reaches end of life and gets resold, traded, or shipped, export controls do not disappear. They follow the hardware. The key points for anyone retiring GPU systems in 2026:
- ECCN 3A090.a is the classification that matters. It captures advanced AI accelerators above specific performance thresholds, including NVIDIA H100, H200, A100, B100, B200, and GB200 series, AMD MI300X and MI325X, and systems containing them (under 4A090 and corresponding .z classifications).
- Controls apply to resale and reexport, not just original export. Reexports from third countries and in-country transfers are covered. A retired GPU resold across a border is an export transaction subject to the same framework as the original sale.
- A worldwide license requirement now applies to the most advanced chips. The January 2025 framework imposed worldwide licensing on 3A090.a advanced computing items, with a tiered system of country groups, license exceptions, and presumptions.
- The January 15, 2026 rule shifted some China/Macau applications to case-by-case review, with strict conditions. Now finalized, it moved H200 and AMD MI325X exports (those below a total processing performance of 21,000 and a total DRAM bandwidth of 6,500 GB/s) from presumption of denial to case-by-case review, subject to a 25 percent tariff, a 50 percent volume cap, third-party testing, and rigorous know-your-customer requirements. Licenses are still required and the bar is high.
- The H200-to-China opening has not translated into actual shipments. Commerce cleared roughly ten Chinese firms (including Alibaba, ByteDance, and Tencent) to buy H200s, each capped at about 75,000 units, but as of mid-2026 none had shipped, reportedly because China is blocking the imports to favor its domestic chip industry. The legal path opened; the trade did not follow.
- The Chip Security Act is advancing. H.R. 3447 passed the House Foreign Affairs Committee 42-0 in March 2026, and would require chip-level location verification for covered chips (under ECCNs 3A090 and 3A001.z). If enacted, it adds serial-number-level inventory, periodic audits, and end-user attestation obligations for license holders globally.
- Two adjacent 2026 developments tightened the picture. The House passed the Remote Access Security Act (369-22) to close the cloud GPU rental loophole, and in March 2026 the DOJ National Security Division issued a Corporate Enforcement Policy creating affirmative obligations to voluntarily disclose potential export-control violations.
For ITAD specifically, the practical takeaway is that retiring GPU infrastructure requires export-classification awareness, careful documentation, denied-party screening, and (often) restricting the buyer pool to compliant destinations. The recovery value on retired AI hardware is substantial, but capturing it compliantly requires treating export controls as part of the disposition plan, not an afterthought.
Why Export Controls Matter at Retirement, Not Just Purchase
The conversation about GPU export controls is almost entirely framed around acquisition: who can buy advanced chips, which countries can receive them, what licenses importing requires. That framing misses half the lifecycle.
Export controls attach to the hardware, not to a single transaction. When a controlled GPU is resold, traded in, consigned, or shipped across a border at end of life, that movement is itself an export, reexport, or in-country transfer subject to the Export Administration Regulations (EAR). The classification that applied when the chip was new still applies when it is retired.
This creates a blind spot. An operator that carefully managed export compliance when deploying a GPU fleet may treat the same fleet as ordinary scrap or surplus at retirement, handing it to an ITAD vendor or broker without considering that reselling those chips internationally carries the same obligations as the original import. The result can be an inadvertent violation, and it is a violation of the EAR even for non-U.S. persons to proceed with a transaction knowing U.S. export controls have been or will be violated.
The retirement side is arguably higher-risk precisely because it gets less attention. Procurement and legal teams scrutinize acquisition. Disposition often runs through facilities or IT operations with less export-control awareness, and through brokers and secondary-market channels where the buyer’s ultimate destination may be opaque.
What ECCN 3A090 Actually Covers
ECCN (Export Control Classification Number) 3A090 is the classification at the center of advanced-chip export controls. Understanding it in outline (not as a substitute for a formal classification) helps frame the disposition question.
The Performance Thresholds
ECCN 3A090.a controls integrated circuits with digital processing units meeting specific performance criteria. As described in the current framework, it captures ICs having either a “total processing performance” (TPP) of 4800 or more, or a TPP of 1600 or more combined with a “performance density” of 5.92 or more. These thresholds are designed to capture advanced AI-training accelerators while leaving lower-performance chips less restricted.
BIS (the Bureau of Industry and Security, the agency within the U.S. Department of Commerce that administers the EAR) has signaled these thresholds will be reviewed periodically as new chip generations enter production, so the specific numbers are subject to change.
The Related Classifications
The control framework extends beyond the bare chip:
| Classification | What It Covers |
|---|---|
| 3A090.a | The advanced integrated circuits (the chips) meeting the performance thresholds |
| 4A090.a | Computers, electronic assemblies, and components containing 3A090.a-level chips (the systems) |
| Corresponding .z ECCNs | Items in other categories that incorporate or reference 3A090.a/4A090.a controls (e.g., 3A001.z, 5A002.z) |
| 4E091 | Model weights of certain advanced closed-weight AI models (a separate but related control) |
The practical implication for ITAD: it is not just loose GPUs that are controlled. Complete AI server systems, HGX baseboards, and assemblies containing covered chips fall within scope under 4A090 and the .z classifications. A retired AI server is not automatically outside the control framework just because it is a “system” rather than a “chip.”
The Covered Hardware
The chips and systems commonly understood to fall within these classifications include NVIDIA’s H100, H200, A100, B100, B200, and GB200 series, and AMD’s MI300X and MI325X series, along with products and complete server systems containing them. This is exactly the hardware filling AI data centers in 2026, and exactly the hardware that will dominate the AI retirement wave as these fleets refresh.
The 2026 Regulatory Landscape
The export-control framework for advanced computing has been among the most actively revised areas of U.S. trade regulation. A few anchor points current to mid-2026, again as general orientation rather than legal guidance:
The Worldwide License Requirement
The January 2025 framework established a worldwide license requirement for advanced computing ICs under 3A090.a, 4A090.a, and corresponding .z classifications, along with associated software and technology and certain AI model weights. This replaced a narrower destination-based approach with a global framework built around country groups, license exceptions, and presumptions of approval or denial depending on destination.
In broad strokes: applications for close U.S. allies and partners (a defined list including countries in specific country groups) enjoy a presumption of approval, applications for China, Macau, and arms-embargoed destinations face a presumption of denial, and other destinations fall into intermediate treatment, in some cases presumptively approved up to a country-wide compute quota and presumptively denied thereafter.
The January 15, 2026 Adjustment
A January 15, 2026 BIS final rule created an opening for certain previously near-prohibited transactions. License applications to export specific chips to China and Macau can now be reviewed under a case-by-case policy rather than the prior categorical presumption of denial. This followed a December 8, 2025 announcement that the United States would allow the H200 and similar products to ship to approved customers in China.
The relaxed posture is narrow and heavily conditioned. It applies only to chips below defined performance thresholds: a total processing performance (TPP) under 21,000 and a total DRAM bandwidth under 6,500 GB/s, which correspond roughly to the NVIDIA H200 and AMD MI325X. (The MI325X sits close to the ceiling, with a TPP around 20,800, leaving little headroom.) And qualifying for case-by-case review comes loaded with requirements: a 25 percent tariff, a 50 percent volume cap (China-bound shipment volume cannot exceed half the exporter’s U.S. sales volume for the same product line), a demonstration that no domestic orders were delayed or deprioritized to fulfill China-bound shipments, third-party testing, and strict know-your-customer screening.
This did not remove the license requirement. A license is still required for 3A090.a items to China and Macau, and the underlying 3A090/4A090 framework, its performance caps and interconnect thresholds, remains fully intact. What changed is the government’s willingness to grant licenses for hardware within the defined window, not the rules themselves. The most advanced chips (Blackwell-class and anything above the specified thresholds) remain under the tighter posture.
One striking wrinkle underscores that a legal opening is not the same as a transaction. Commerce cleared roughly ten Chinese firms, including Alibaba, ByteDance, Tencent, and JD.com, to buy H200s, each capped at about 75,000 units. Yet as of mid-2026, not one of those chips had actually shipped, reportedly because China moved to block the imports in order to push buyers toward its domestic semiconductor industry. For anyone tracking where covered hardware can actually flow, the lesson is that the map of legal destinations and the map of real-world markets do not always match, and both matter for disposition planning.
The Cloud-Access and Enforcement Layer
Two developments in early 2026 tightened the framework around, rather than through, the chips themselves. The House passed the Remote Access Security Act (by a 369-22 vote), aimed at closing the “cloud loophole” whereby a restricted party could rent remote access to controlled GPUs hosted in a permitted country rather than importing the hardware directly. Separately, in March 2026 the Department of Justice National Security Division issued a Corporate Enforcement Policy creating affirmative obligations around the voluntary disclosure of potential export-control violations. Both signal the same direction of travel: the compliance perimeter is widening to cover access and disclosure, not just physical shipment, and enforcement expectations are rising across the supply chain, including for intermediaries and service providers.
The Chip Security Act
Looking forward, the Chip Security Act (H.R. 3447) advanced on March 26, 2026, when the House Foreign Affairs Committee voted unanimously, 42-0, to move it forward, with a Senate companion bill introduced. If enacted, it would direct Commerce to require covered chips (those under ECCNs 3A090, 3A001.z, 4A090, and 4A003.z) to incorporate a hardware-level location-verification mechanism before export, and would impose chip-level, serial-number inventory, periodic on-site audits, and end-user attestation obligations on license holders globally. As of mid-2026 it had cleared committee but was not yet enacted, and its trajectory is worth watching closely.
For ITAD, the Chip Security Act’s most relevant feature is the serial-number-level tracking it contemplates. If location-verification, periodic audit, and chip-level inventory obligations become law, disposition documentation would need to account for covered chips individually, at the serial-number level, from receipt through final disposition. That reinforces the per-asset, serialized documentation discipline that serious ITAD already practices, and it would make the chain-of-custody record for a retired GPU a compliance artifact in its own right, not just a data-security one.
What This Means for GPU Disposition
Translating the regulatory framework into ITAD practice, several disposition realities follow.
The Buyer Pool Is Constrained
The single biggest practical effect on recovery is that export controls narrow the buyer pool for retired covered GPUs. The most valuable secondary-market demand for advanced accelerators includes international operators, but many international destinations are restricted, licensed, or quota-limited. A compliant disposition process screens buyers and destinations against the framework, which means some of the highest-bidding potential buyers may be ineligible without a license.
This is not a reason to avoid international resale. It is a reason to manage it deliberately, matching covered hardware to compliant destinations and buyers, and pursuing licenses where the value justifies the process.
Classification Comes First
Before any retired GPU hardware is marketed for resale, the disposition process should establish the export classification of the items. Whether a specific chip or system falls under 3A090.a/4A090.a, a less-restricted classification, or outside the advanced-computing controls entirely determines what the resale options are. This classification is the foundation for every subsequent compliance step, and it is exactly the kind of determination that warrants qualified export-control input rather than assumption.
Denied-Party and Destination Screening
Compliant disposition screens all parties to a resale transaction against the relevant restricted-party lists (the Entity List, Denied Persons List, and Specially Designated Nationals list) and confirms the destination’s treatment under the framework. Gaps in a transaction, intermediaries who refuse to provide end-user information, or inconsistencies in the supply chain are red flags that a compliant process does not ignore.
Documentation Carries the Compliance Story
Export-compliant disposition produces documentation: the classification determination, the screening results, the end-user and destination information, and the records substantiating that the transaction was permissible. This documentation is what demonstrates good-faith compliance if a transaction is ever questioned, and it dovetails with the serialized, per-asset chain-of-custody documentation that data-security-driven ITAD already requires.
The Domestic Path Is Often Cleaner
For many operators, the lowest-risk high-value path for retired covered GPUs is the domestic U.S. secondary market, where the export dimension does not arise. Strong domestic demand from U.S. enterprises building AI capacity, U.S. neoclouds, and domestic research institutions means meaningful recovery value can often be captured without crossing a border at all. An ITAD partner with deep domestic buyer relationships can frequently match covered hardware to compliant domestic buyers, sidestepping the export question while still capturing value.
How This Fits the Broader AI Retirement Wave
GPU export controls are arriving at the same moment as the first large wave of AI infrastructure retirement. The two trends intersect directly.
The AI hardware deployed at scale in 2024 through 2026 runs on compressed refresh cycles, with GPU systems reaching retirement candidacy in as little as 18 to 36 months as each new accelerator generation makes the prior one uneconomic for training. That means the H100 and A100 fleets deployed during the AI build-out are entering their retirement windows now and over the next few years, precisely the covered hardware under 3A090.
The result is a growing volume of high-value, export-controlled equipment flowing into the secondary market, much of it from operators whose export-compliance attention was focused on acquisition rather than disposition. The operators who plan disposition with export controls in mind (classifying early, screening buyers, documenting thoroughly, and leaning on compliant domestic channels where appropriate) will capture the substantial recovery value in these fleets without the compliance exposure. Those who treat retired GPUs as ordinary surplus risk inadvertent violations in a heavily scrutinized area.
This is why export-control awareness has become a genuine ITAD vendor-selection criterion for GPU-heavy operators, not a nice-to-have. The recovery value is real, but it is recoverable compliantly only by a process that treats export controls as integral to disposition.
Frequently Asked Questions
The following is general information, not legal advice. Consult qualified export-control counsel for guidance on specific transactions.
What is ECCN 3A090?
ECCN (Export Control Classification Number) 3A090 is a U.S. export classification for advanced computing integrated circuits. The 3A090.a sub-classification controls chips meeting specific performance thresholds (a “total processing performance” of 4800 or more, or 1600 or more combined with a “performance density” of 5.92 or more). It captures advanced AI accelerators including the NVIDIA H100, H200, A100, B100, B200, and GB200 series and AMD MI300X and MI325X, while related classifications (4A090 and corresponding .z ECCNs) cover systems containing those chips. The thresholds are administered by the Bureau of Industry and Security and are subject to periodic revision.
Do export controls apply when I resell or retire GPUs?
Yes. Export controls attach to the hardware, not just to the original purchase. Reselling, trading, consigning, or shipping a controlled GPU across a border at end of life is an export, reexport, or in-country transfer subject to the Export Administration Regulations, with the same classification that applied when the chip was new. This is a common blind spot: operators that managed export compliance carefully at acquisition sometimes treat the same hardware as ordinary surplus at retirement, which can lead to inadvertent violations. Disposition of covered GPUs should be planned with export controls in mind.
Which GPUs are subject to export controls?
The chips and systems commonly understood to fall under the advanced-computing controls (ECCNs 3A090.a, 4A090.a, and corresponding .z) include NVIDIA’s H100, H200, A100, B100, B200, and GB200 series and AMD’s MI300X and MI325X series, along with complete server systems and assemblies containing them. Whether a specific item is controlled depends on its actual performance characteristics against the current thresholds, which is why a formal export classification is the necessary starting point rather than assuming based on model name alone.
Can I sell retired H100s or H200s internationally?
Sometimes, depending on the destination, the buyer, and the licensing posture. Many destinations are restricted, licensed, or quota-limited under the current framework. Close U.S. allies generally enjoy a presumption of license approval, China and arms-embargoed destinations face a presumption of denial (though a January 2026 rule opened a case-by-case review path for certain chips like the H200 to China/Macau IaaS providers under rigorous conditions), and other destinations fall in between. International resale of covered GPUs is possible but requires classification, denied-party and destination screening, and often a license. The domestic U.S. secondary market frequently offers strong recovery value without the export dimension.
Can NVIDIA H200 chips be exported to China in 2026?
Legally, a narrow path exists, but in practice the trade has not materialized. The January 15, 2026 BIS rule moved H200 and AMD MI325X exports (those below a TPP of 21,000 and DRAM bandwidth of 6,500 GB/s) from presumption of denial to case-by-case review for China and Macau, subject to a 25 percent tariff, a 50 percent volume cap, third-party testing, and strict know-your-customer requirements. A license is still required, and eligibility for case-by-case review does not guarantee approval. Notably, although Commerce cleared roughly ten Chinese firms to buy H200s (each capped at about 75,000 units), as of mid-2026 none had actually shipped, reportedly because China moved to block the imports to favor its domestic industry. For disposition planning, the takeaway is that a destination being legally permissible under a narrow window does not make it a practical or reliable market, and the most advanced chips remain under the tighter posture regardless.
What is the Chip Security Act?
The Chip Security Act (H.R. 3447) is proposed U.S. legislation that advanced on March 26, 2026, when the House Foreign Affairs Committee voted unanimously, 42-0, to move it forward, with a Senate companion bill introduced. If enacted, it would direct Commerce to require covered chips (under ECCNs 3A090, 3A001.z, 4A090, and 4A003.z) to incorporate a hardware-level location-verification mechanism before export, and would impose chip-level, serial-number inventory, periodic on-site audits, and end-user attestation obligations on license holders globally. For ITAD, its most relevant feature is the serialized, chip-level tracking it contemplates, which would reinforce per-asset disposition documentation. As of mid-2026 it had cleared committee but was not yet enacted.
Does the export control apply to the GPU chip or the whole server?
Both, through different classifications. The advanced chip itself is controlled under 3A090.a, while computers, electronic assemblies, and components containing a 3A090.a-level chip are controlled under 4A090.a, and various items incorporating these controls fall under corresponding .z classifications. A retired AI server is not automatically outside the control framework simply because it is a complete system rather than a loose chip. The configuration matters, which is another reason formal classification is the appropriate starting point for disposition.
What happens if I accidentally export a controlled GPU without a license?
Exporting, reexporting, or transferring a controlled item without required authorization is a violation of the Export Administration Regulations, with potentially serious civil and criminal penalties. Importantly, it is a violation even for non-U.S. persons to proceed with a transaction knowing U.S. export controls have been or will be violated, and “knowing” can include awareness of red flags that were not investigated. This is why compliant disposition screens parties and destinations, documents the compliance basis, and does not ignore gaps or intermediaries who refuse to provide end-user information. Enforcement in this area has been active, and BIS has sought to extend the statute of limitations for violations.
How should ITAD handle export-controlled GPU hardware?
Compliant ITAD for covered GPUs treats export controls as integral to disposition: establishing the export classification of the hardware first, screening all buyers and destinations against restricted-party lists and the country-group framework, documenting the classification and screening basis per asset, matching covered hardware to compliant buyers and destinations (often favoring the domestic U.S. secondary market where strong demand exists without the export dimension), and pursuing licenses where the recovery value justifies the process. This dovetails with the serialized, per-asset chain-of-custody documentation that data-security-driven ITAD already practices. Export-control awareness has become a genuine vendor-selection criterion for GPU-heavy operators.
Is export-control compliance the ITAD vendor’s responsibility or mine?
Responsibility under the EAR generally rests with the parties to the export transaction, which can include the equipment owner, the exporter, and intermediaries, and liability is not necessarily shielded by relying on an intermediary, particularly where red flags exist. In practice, a knowledgeable ITAD partner contributes classification awareness, denied-party screening, destination analysis, and compliance documentation, but the equipment owner retains obligations and should confirm how compliance responsibilities are allocated in any disposition arrangement. Because allocation of legal responsibility is fact-specific, this is a point to work through with qualified export-control counsel rather than assume.
The Bottom Line
GPU export controls do not end when a chip is deployed. They follow the hardware into retirement, which means reselling, trading, or shipping a covered GPU at end of life carries the same export obligations as the original import, under ECCN 3A090.a and its related classifications. The hardware filling AI data centers in 2026 (H100s, H200s, A100s, and the rest) is exactly the controlled hardware now entering its retirement window on compressed AI refresh cycles.
The intersection of these two trends, tightening controls and a rising retirement wave, creates real exposure for operators who treat retired GPUs as ordinary surplus. It also creates real recoverable value for those who handle disposition compliantly: classifying early, screening buyers and destinations, documenting thoroughly, and leaning on the strong domestic U.S. secondary market where the export dimension does not arise. The recovery value in retired AI fleets is substantial. Capturing it without compliance exposure is a process discipline, and increasingly a reason to choose an ITAD partner with genuine export-control awareness.
Export law in this area is complex and changing. The most important practical step for any operator retiring covered GPU hardware is to involve qualified export-control counsel in the disposition plan, and to work with an ITAD partner that treats export compliance as integral to recovery rather than an afterthought.
How ROC Telecom Helps
ROC Telecom is an R2v3, RIOS, NIST 800-88, and ITAR-compliant ITAD specialist with export-control awareness built into GPU and AI infrastructure disposition:
- Export-classification awareness integrated into GPU disposition planning, so covered hardware is identified before it is marketed for resale
- Denied-party and destination screening as part of the resale process, matching covered hardware to compliant buyers
- Deep domestic U.S. buyer network across enterprises, neoclouds, and research institutions, frequently capturing strong recovery value without the export dimension
- ITAR-compliant handling and the documentation discipline that export-controlled equipment requires
- Per-asset serialized documentation with chain-of-custody, aligned with the serialized tracking that emerging chip-security requirements contemplate
- NIST 800-88 data destruction on all data-bearing components before any disposition
- Specialist GPU and AI infrastructure recovery with speed-to-remarketing that protects value against generational decay
15+ years of ITAD experience, $25M+ in client capital recovered, 45M+ pounds diverted from landfill.
ROC Telecom is not a law firm and does not provide legal or export-classification opinions. For covered hardware, we work alongside your export-control counsel to support compliant disposition.
Request a Free GPU Disposition Assessment
Tell us about the AI infrastructure you are retiring. A specialist will discuss recovery value, secure data destruction, and compliant disposition pathways, including domestic options that capture value without the export dimension. No commitment, no spam. Prefer to talk directly? Call 585-406-1249 or email info@roctelecom.com.
"*" indicates required fields
Related reading:
- How GPU Decommissioning Differs from Standard Server Retirement
- Hyperscale ITAD vs Enterprise ITAD: What’s Actually Different in 2026
- ITAD for Cloud Providers: The 2026 Strategic Guide
- Data Center Asset Recovery: The 2026 Strategic Guide
- Why AI Is Shortening Your Hardware Lifecycles (and What to Do About It)
- Why Wiping Isn’t Destroying: A 2026 Guide to NIST 800-88 Data Sanitization
- The Texas Data Center Market: A 2026 Guide
