TL;DR
GPU decommissioning isn’t a faster version of standard server retirement. It’s a different operation. Six factors separate them:
- Residual value scale. H100s retain 60–70% of new price after three years. Standard servers retain 20–30%.
- Data sanitization complexity. High Bandwidth Memory (HBM) on the GPU package isn’t addressed by standard NIST 800-88 procedures.
- Export controls. NVIDIA H100, H200, and B-series GPUs fall under ECCN 3A090, making cross-border handling an ITAR/EAR matter.
- Physical infrastructure. 50–140 kW racks, liquid cooling, and chassis weights up to 3,000 lbs change removal logistics.
- Per-GPU documentation requirements that standard ITAD contracts don’t anticipate.
- Specialist secondary-market access, which determines whether recovered value is captured or left on the floor.
Most enterprise ITAD contracts written before 2023 don’t address any of the six. If you’re running AI workloads, you need a separate decommissioning workflow specifically for the GPU portion of your fleet.
Why This Matters in 2026
The forced GPU retirement wave is here.
NVIDIA Blackwell B200 reached general availability in Q1 2026. The architectural leap, combined with the platform’s 140 kW rack power profile, is pulling H100 and A100 fleets out of service faster than anyone planned. Hyperscalers are rotating Hopper-generation inventory to make floor space for Blackwell. Enterprise operators with air-cooled facilities are retiring H100 clusters that physically can’t host B200s. Co-location tenants are coordinating GPU fleet retirements in 200 to 500 accelerator batches aligned with fiscal year planning.
This creates a problem most ITAD programs weren’t built for. A five-year contract signed in 2022 was written for CPU-based servers worth $200 to $2,000 each at retirement. Those same contracts are now being asked to handle GPU chassis worth $25,000 to $50,000 each, under export-control regulations, with HBM memory that standard sanitization procedures don’t fully address.
The cost of treating GPU decommissioning as a standard ITAD task isn’t theoretical. It’s the difference between recovering $4 million on a retired 200-unit H100 fleet and recovering $400,000 on the same fleet through a generalist vendor.
The 6 Structural Differences
1. Residual Value at Scale
Standard servers depreciate sharply. A four-year-old enterprise rack server with two Xeon Scalable CPUs and 512 GB of DDR4 trades at 15 to 25% of new value. The labor cost of ITAD processing often consumes most of that recovery.
GPUs don’t behave this way.
The NVIDIA H100 80GB held a new-purchase price of $25,000 to $40,000 from mid-2024 into early 2026. Two-year-old used H100s traded at a median of roughly 61% of contemporaneous new price; three-year-old units at roughly 69%. Even at the lower bound, a single retired H100 represents $15,000 to $20,000 of recoverable capital. A 200-GPU fleet retirement is a $3 to $4 million asset recovery event, not an e-waste line item.
This changes the entire ITAD economic model:
| Asset Class | Typical New Price | 3-Year Resale Value | Ratio | Implication |
|---|---|---|---|---|
| Standard rack server | $8,000–$15,000 | $1,500–$3,500 | ~20–25% | Recycling-led; recovery is incidental |
| Enterprise storage array | $30,000–$80,000 | $4,000–$10,000 | ~12–15% | Component harvest model |
| NVIDIA A100 80GB | $25,000+ | $12,000–$18,000 | ~50–60% | Remarketing-led; speed matters |
| NVIDIA H100 80GB | $25,000–$40,000 | $15,000–$25,000+ | ~60–70% | Remarketing-led; specialist buyer access required |
| NVIDIA H200 141GB | $32,000–$45,000 | (early secondary market) | Higher | Window-sensitive |
Standard ITAD treats your fleet as recycling input with some salvage upside. GPU ITAD has to treat it as a depreciating financial asset, where time-to-market determines outcome.
2. HBM and GPU-Specific Data Sanitization
NIST Special Publication 800-88 Rev. 1 was published in 2014, before HBM was widely deployed. Its sanitization guidance addresses HDDs, conventional SSDs, magnetic tape, and standard DRAM. It does not provide explicit procedures for High Bandwidth Memory stacks integrated on the GPU package.
This matters because modern AI accelerators carry substantial on-package memory:
| Accelerator | On-Package Memory | Type |
|---|---|---|
| NVIDIA A100 80GB | 80 GB | HBM2e |
| NVIDIA H100 80GB | 80 GB | HBM3 |
| NVIDIA H200 | 141 GB | HBM3e |
| NVIDIA B200 | 192 GB | HBM3e |
| NVIDIA GB200 (per GPU) | 192 GB | HBM3e |
A retired H100 carries 80 GB of HBM3 that may have processed training data subject to HIPAA, GLBA, ITAR, or sovereign data residency requirements. Sanitization of this memory requires GPU-specific firmware procedures (reset commands, GPU memory clear routines invoked through the driver stack) that most ITAD intake workflows don’t perform. Physical destruction is the only universally compliant fallback, but physical destruction forfeits the $15,000 to $25,000 of residual value the GPU still carries.
The newer IEEE 2883-2022 standard addresses modern storage technologies more comprehensively. Enterprise-grade GPU sanitization in 2026 typically requires a layered procedure:
- Firmware-level memory clear invoked through the GPU driver while the system is operational.
- NVMe sanitization of any local SSDs in the GPU chassis to NIST 800-88 Purge level (or IEEE 2883 Purge).
- Server BMC reset and firmware reflash to eliminate persistent administrative state.
- Physical removal and serialized chain of custody before any third party handles the asset.
- Vendor-attested certificate specifying which sanitization methods were applied per GPU, by serial number.
A standard ITAD vendor handling your CPU fleet doesn’t have a procedure for steps 1, 3, or 5. Specifying these in your decommissioning SOP isn’t optional for AI workloads handling regulated data.
3. Export Controls Apply
Standard server CPUs aren’t export-controlled. NVIDIA H100, H200, B100, B200, GB200, and several variants are.
The U.S. Department of Commerce’s Export Administration Regulations (EAR) classify advanced AI accelerators under ECCN 3A090 (integrated circuits with specified performance thresholds) and ECCN 4A090 (computers containing 3A090 chips). The October 2023 amendments expanded the controlled list, and subsequent updates have continued to tighten cross-border restrictions on AI hardware.
For GPU decommissioning, this means:
- Any cross-border handling requires export-license consideration. An ITAD vendor that processes equipment in Mexico or Canada (common for cost reasons in standard ITAD) cannot ship retired H100s across the border without compliance review.
- End-use certification matters. Retired GPUs can’t be sold to entities on the U.S. Entity List or to addresses in restricted jurisdictions, even through second- or third-party buyers.
- Documentation must include export classification. Your asset recovery vendor must be able to demonstrate, on audit, where each unit ended up and to whom it was sold.
- ITAR-compliant providers are the practical baseline. Vendors without formal export-compliance programs shouldn’t be handling AI hardware retirement.
This is the single most-overlooked aspect of GPU decommissioning in 2026. A vendor that delivers strong economics but can’t produce export-compliance documentation creates a liability that survives the transaction by years.
4. Physical Infrastructure Complexity
Standard rack servers ship at 30 to 80 lbs each in 1U or 2U form factors and consume 5 to 15 kW per rack. They’re designed for two-person lift, standard air cooling, and conventional cable management.
GPU systems are different objects:
| System | Approximate Weight | Power Density | Cooling |
|---|---|---|---|
| DGX A100 (8× A100) | ~287 lbs | ~6.5 kW | Air |
| DGX H100 (8× H100) | ~287 lbs | ~10.2 kW | Air |
| HGX H100 8-GPU baseboard | Varies by OEM | Up to 12 kW | Air or hybrid |
| GB200 NVL72 | ~3,000 lbs (rack-scale) | ~140 kW per rack | Liquid (direct-to-chip) |
The physical decommissioning implications:
- Liquid cooling requires drainage before any GB200 system can be moved. This is plumbing work, not server work, and it must precede de-racking. Coolant disposal carries its own environmental compliance.
- Specialized lift gear is needed for chassis above ~150 lbs. Standard two-person ITAD lift protocols don’t apply.
- NVLink and NVSwitch fabrics carry high-value optical interconnects that shouldn’t be cut or scrapped. A typical 8-GPU H100 system carries several thousand dollars of optical modules alone, distinct from the GPU value itself.
- 800G optics in newer GPU systems are themselves recoverable inventory worth $1,500 to $3,000 per port at retirement. Standard shred-and-recycle approaches destroy this value.
A decommissioning team without specific procedures for liquid cooling, NVLink optics, and rack-scale chassis handling is the wrong team for AI infrastructure retirement.
5. Per-GPU Serialization and Documentation
A standard ITAD contract typically tracks retired equipment at the chassis or rack level. A Certificate of Destruction lists 40 servers by manufacturer, model, and chassis serial number.
GPU ITAD requires deeper documentation. Each GPU is itself a tracked asset:
- Per-GPU serial numbers captured at de-rack, separate from server chassis serials
- Sanitization attestation per GPU specifying which procedures (firmware clear, BMC reset, physical destruction) were applied
- Resale or destruction disposition per GPU with buyer or refiner reference, for export-compliance audit
- HBM data residue attestation stating that on-package memory was sanitized through a validated GPU-specific procedure
- Warranty status documentation at point of retirement, which materially affects secondary-market value
This level of documentation isn’t optional for AI workloads. The same compliance teams that demand serialized HDD destruction certificates for CPU servers will demand equivalent or stronger documentation for GPU retirement, particularly for workloads handling regulated data.
A standard ITAD contract that ships a single chassis-level Certificate of Destruction for a retired 8-GPU DGX system hasn’t met the audit standard for AI infrastructure retirement. Insist on per-accelerator documentation.
6. Specialist Secondary-Market Access
The secondary GPU market is structurally different from the secondary server market.
Standard server resale is a liquid, well-indexed market. Sites like ServerMonkey, IT Creations, and a hundred others publish pricing and accept consignment. Recovery is mechanical: list, sell, ship, settle.
The GPU secondary market is opaque, relationship-driven, and time-sensitive. Real recovery requires:
- Direct relationships with neoclouds and AI infrastructure operators. CoreWeave, Lambda, Crusoe, Together AI, and dozens of similar operators are active buyers of retired hyperscaler H100 fleets.
- Direct relationships with hyperscalers for inter-company transfers when fleets meet specific configuration and burn-in thresholds.
- Access to international buyers in jurisdictions where export compliance has been verified.
- Pricing intelligence on warranty transferability. A unit with 12+ months of transferable third-party warranty trades 8 to 12% above an equivalent unit with expired coverage.
- Speed-to-market discipline. The value depreciation curve for H100s is non-linear and event-driven. Each NVIDIA generation release creates a step-down.
A standard ITAD vendor selling your retired H100s through general broker channels typically clears at 60 to 70% of what a specialist with direct buyer relationships can recover. On a $4M fleet retirement, that gap is $1.2M to $1.6M of capital left on the floor.
What to Specify in a GPU Decommissioning SOP
A defensible GPU decommissioning specification should include:
| Section | Minimum Requirement |
|---|---|
| Vendor certifications | R2v3, RIOS, NAID AAA, ITAR-registered, documented export-compliance program |
| Sanitization protocol | NIST 800-88 Purge or IEEE 2883 Purge for local NVMe; GPU-specific firmware memory clear; BMC reset; per-GPU attestation |
| Physical handling | Liquid cooling drainage procedure; chassis-specific lift protocols; NVLink optics preservation; serialized chain of custody at de-rack |
| Documentation deliverables | Per-GPU serial inventory; per-GPU sanitization certificate; per-GPU disposition record (resale buyer or destruction refiner); HBM data residue attestation; export-classification confirmation |
| Asset recovery model | Specialist remarketing channel with direct neocloud/hyperscaler relationships; documented warranty-transfer process; settlement timeline aligned with depreciation curve (60–90 days max) |
| Liability transfer | Loading dock transfer with serialized chain-of-custody documentation; carry of export-compliance liability through final disposition |
| ESG reporting | Scope 3 Category 5 and Category 12 inputs; mass-balance recovery data; on-request precious-metal assay |
Specifying this in your decommissioning contract, and verifying the vendor can actually deliver against it, is the difference between an AI hardware retirement that recovers capital and one that creates a multi-year liability.
When to Engage a GPU Decommissioning Specialist vs. Your Existing ITAD
| Scenario | Recommendation |
|---|---|
| Retiring 1–4 GPU servers, no regulated data | Standard ITAD vendor with documented sanitization can suffice |
| Retiring a fleet of 50+ GPUs | Specialist with secondary-market buyer access |
| Mixed environment retirement (CPU servers + GPU systems) | Specialist for the GPU portion; standard ITAD acceptable for CPU servers |
| Regulated data workloads (HIPAA, GLBA, ITAR-controlled training data) | Specialist with documented GPU sanitization SOP; standard ITAD is insufficient |
| Cross-border decommissioning | Specialist with formal export-compliance program; standard ITAD without ECCN handling is non-compliant |
| Liquid-cooled rack-scale systems (GB200 NVL72, GB300) | Specialist with documented liquid-cooling and rack-scale handling procedures |
| Compressed timeline (less than 30 days) | Specialist with hyperscale-grade field operations and 48-hour mobilization |
The threshold isn’t arbitrary. If any one of these conditions applies, the cost-benefit math favors a specialist over a generalist by a material margin.
Frequently Asked Questions
Why can’t I use my existing ITAD vendor for GPU retirement?
Most enterprise ITAD contracts written before 2023 don’t address GPU-specific requirements: HBM sanitization procedures, export-control classification (ECCN 3A090), per-accelerator documentation, liquid-cooling handling, or specialist buyer access for secondary-market value recovery. Your existing vendor may still be the right partner for the CPU server portion of a mixed retirement, but the GPU portion typically benefits from a specialist with documented procedures for AI hardware.
What is HBM and why does it matter for data destruction?
HBM (High Bandwidth Memory) is high-speed memory integrated directly onto the GPU package, providing dramatically higher bandwidth than conventional DRAM. NVIDIA H100s carry 80 GB of HBM3; H200s carry 141 GB of HBM3e; B200s carry 192 GB of HBM3e. Because HBM is on-package rather than socketed, standard memory sanitization procedures don’t apply directly. Workloads that processed sensitive training or inference data through HBM require GPU-specific firmware memory-clear procedures, attested per-GPU, as part of a compliant sanitization SOP.
Are NVIDIA H100s and H200s export-controlled?
Yes. NVIDIA H100, H200, B100, B200, GB200, and several variants fall under U.S. Export Administration Regulations classification ECCN 3A090 (the chips) and ECCN 4A090 (computers containing them). This means any cross-border handling, including processing through international ITAD facilities, resale to international buyers, or transit through restricted jurisdictions, requires export-license consideration and formal compliance documentation. ITAD vendors handling these accelerators should maintain an active export-compliance program and ITAR registration.
How much value can I expect to recover from a retired H100 fleet?
As of early 2026, retired H100 80GB units trade at roughly 60 to 70% of contemporaneous new pricing in the specialist secondary market. On a 200-unit fleet retirement, that translates to $3 to $4 million in recoverable capital. Standard ITAD broker channels typically clear at 35 to 50% of new value, capturing perhaps $1.5 to $2 million on the same fleet. The gap is direct buyer access, warranty preservation, and speed-to-market discipline. All three require a specialist partner.
When is the right time to retire an H100 fleet?
The H100 depreciation curve is non-linear. Major value step-downs are triggered by NVIDIA generation releases (B200 GA in Q1 2026; expected Rubin announcements in 2026; expected Rubin Ultra in 2027). For value recovery, retirement coordinated with the start of a generation transition, but completed before that generation reaches volume deployment, typically captures the strongest pricing. Holding through a step-down event can cost 10 to 20% of recovered value with no operational upside.
Can GPU systems be sanitized to NIST 800-88 standards?
The local NVMe storage in a GPU server can be sanitized to NIST 800-88 Purge or IEEE 2883 Purge standards using standard procedures. The HBM memory on the GPU package itself isn’t explicitly addressed by NIST 800-88 Rev. 1 (published 2014) and requires GPU-specific firmware procedures, attested per accelerator. The forthcoming NIST SP 800-88 Revision 2 (draft released July 2025) is expected to provide more contemporary guidance on modern storage technologies.
How are Blackwell B200 systems different to decommission from H100s?
B200 systems, particularly rack-scale GB200 NVL72 deployments, introduce direct-to-chip liquid cooling at 140 kW rack power densities. Decommissioning requires liquid cooling drainage as a precursor to any de-racking work, specialized lift gear for chassis weights approaching 3,000 lbs, and coolant disposal under environmental compliance. The GPU sanitization requirements are similar to H100 procedures. The physical handling profile is fundamentally different.
What documentation should I require from a GPU ITAD vendor?
At a minimum: serialized per-GPU inventory at de-rack; per-GPU sanitization certificate specifying methods applied; per-GPU disposition record (resale buyer reference or destruction refiner); HBM data residue attestation; export-classification confirmation; warranty status documentation at retirement; Scope 3 Category 5 and Category 12 reporting inputs; and a serialized Certificate of Destruction for any unit not remarketed. A chassis-level CoD alone doesn’t meet the audit standard for AI hardware retirement.
The Bottom Line
GPU decommissioning shares a name with standard server retirement but has structurally different requirements at every stage of the workflow. Residual value is 3 to 4 times higher, sanitization complexity is substantially greater, export controls apply, physical infrastructure is denser and heavier, documentation must be per-accelerator, and value recovery depends on specialist secondary-market access that generalist ITAD vendors don’t have.
For organizations operating mixed CPU/GPU environments (which is essentially every enterprise running AI workloads in 2026), the right pattern is usually a dual-vendor approach: the existing ITAD relationship for standard server retirement, and a specialist for GPU and AI fabric retirement. The economics of treating GPU decommissioning as a standard ITAD task no longer work.
The fleets coming out of service in 2026, including H100s displaced by B200s, A100s aging out of production workloads, and Hopper inventory rotating in hyperscale facilities, represent the largest concentrated AI hardware retirement wave in the industry’s history. Treating it as e-waste leaves millions of dollars of recoverable capital on the floor and creates compliance liabilities that survive the transaction.
About ROC Telecom
ROC Telecom delivers specialist GPU and AI infrastructure decommissioning for hyperscale and enterprise data centers. R2v3, RIOS, NIST 800-88, and ITAR-compliant. Direct carrier and hyperscaler resale network for NVIDIA H100, H200, A100, and Blackwell-generation infrastructure. Serialized per-GPU chain of custody, validated sanitization procedures, and specialist secondary-market access that captures 60 to 70% of new-price residual value on retired GPU fleets.
Request a GPU Decommissioning Project Assessment
Tell us about your GPU retirement project. A specialist will reach out to discuss liquid cooling drainage, per-accelerator handling, export compliance, and the right recovery channel for your fleet. No commitment, no spam. Prefer to talk directly? Call 585-406-1249 or email info@roctelecom.com.
"*" indicates required fields
Related reading: Top 10 ITAD Companies for Data Center Decommissioning & Asset Recovery in 2026
