The NVIDIA H100 lost roughly 85% of its value in under three years. That statistic gets quoted often — but the headline number is the wrong thing to fixate on. The real story is what drives the curve: a compressed, forced refresh cycle that is flooding the secondary market while export controls, specialized sanitization requirements, and power-reallocation strategy reshape what “responsible disposition” even means at hyperscale. For data center operators managing an AI hardware refresh and GPU disposition pipeline right now, the window to capture value is small, the penalty for mishandling is seven to eight figures per engagement, and the generic ITAD playbook does not apply.
This guide walks through the four dynamics that make AI-era data center IT asset disposition — the lifecycle discipline of retiring, sanitizing, and routing end-of-life infrastructure, commonly shortened to ITAD — structurally different from the enterprise refresh cycles that preceded it: lifecycle compression, GPU secondary market behavior, export control exposure, and sanitization at the AI edge. Then it closes with what a defensible disposition program looks like when the hardware depreciates daily.
Why Are AI Hardware Refresh Cycles Compressing So Aggressively?
Because the economics of running last-generation silicon at hyperscale are brutal, and the facilities infrastructure can’t absorb both.
For most of the last decade, hyperscalers were extending server lifecycles, not compressing them — AWS publicly stretched its server useful life from three years to five and then to six across 2022-2024. Then AI broke that trend. In late 2024 and early 2025 earnings disclosures, AWS reversed course, cutting the useful life of certain AI-focused servers and networking gear back to five years and retiring some equipment early. That reversal triggered roughly $920 million in accelerated depreciation charges for AWS alone. The industry is now modeling useful life for high-performance AI infrastructure at a highly compressed 18 to 36 months, fracturing the unified lifecycle strategy of the previous decade.
The physical driver is power, not performance. Traditional enterprise racks drew 5 to 10 kW. Modern AI compute racks, particularly those running dense NVLink topologies, routinely demand 40 to 120 kW, with frontier systems projected to push past 150 kW toward 600 kW in the next few years. Because data centers are bounded by utility power allocations — and new grid interconnections can take up to seven years — operators cannot simply add AI clusters to existing halls. They have to displace legacy CPU infrastructure to free up the power footprint, whether or not that equipment is at its failure point or scheduled retirement date.
The downstream effect on data center ITAD is a flood of mid-lifecycle hardware — functional servers, storage arrays, top-of-rack switches, miles of fiber — that was never supposed to be retired this early. The operational implication: your decommissioning pipeline is running against an inbound-hardware clock that doesn’t care about your existing refresh cadence.
What Is Happening to H100 Secondary Market Value Right Now?
Two things, simultaneously: a steep cliff for raw used pulls and a stable premium for properly refurbished units. The gap between them is the entire case for specialized ITAD handling.
The H100 originally retailed between $30,000 and $40,000 at peak — more in scarcity periods. By late 2025, the secondary market had bifurcated sharply. Certified refurbished H100 units at three years post-launch retain roughly 84-85% of their contemporaneous new price, clearing between $30,000 and $34,000. Raw used H100 pulls — divested without reconditioning, diagnostic testing, or warranty backing — clear at 45-69% of new retail, roughly $21,000 to $28,000 per unit. On a rack-scale engagement, that delta is seven figures of capital recovery preserved or destroyed based on the vendor’s refurbishment capability alone.
The cloud rental market tells the same story from the demand side. Peak H100 rentals hit $8.00 to $10.00 per hour in late 2023 and early 2024. By end of 2025, with over 300 new tier-two “neocloud” providers absorbing supply, rental rates had crashed up to 75% — stabilizing between $2.10 and $3.50 per hour. The floor under secondary hardware pricing is the cloud rental rate, and the cloud rental rate is being set by neocloud operators competing on price with last-generation silicon.
The A100 is instructive as a forward-looking reference point. Originally $15,000 to $20,000 at retail, used A100s in 2025 cleared at $7,000 to $12,000 depending on form factor. The A100 found a durable second life among tier-two cloud providers, VFX studios, and edge inferencing deployments, establishing a floor under residual value. The H100 will likely follow the same pattern — but only for units that reach the secondary market in a condition that supports it.
The practical implication: routing GPU assets through a general ITAD vendor that shreds storage by default and has no refurbishment capability for accelerators is the single most common way operators leave seven or eight figures on the table per disposition event.
Curious how your GPU disposition readiness benchmarks against operators capturing the refurbishment premium? Take the free ITAD Readiness Assessment →
How Do Export Controls Shape GPU Disposition?
Heavily — and with liability that persists long after the hardware leaves your loading dock.
Advanced AI accelerators are regulated under the Export Administration Regulations (EAR), administered by the Bureau of Industry and Security (BIS). The U.S. Framework for Artificial Intelligence Diffusion uses compute performance thresholds to restrict where chips can go, applying licensing requirements across more than 140 countries — including critical transit hubs in the Middle East and Southeast Asia where enforcement gaps historically allowed diversion.
Two things matter operationally. First, the controls apply to the hardware in perpetuity — there is no exit from EAR obligations just because you sold the equipment. Second, if your ITAD vendor liquidates a decommissioned A100 or H100 batch to a secondary broker who then routes units into a sanctioned jurisdiction, federal liability flows back to the original operator and the hyperscaler, not just the broker. BIS has been explicit that end-user verification and post-sale diversion prevention are compliance obligations that sit with the seller.
What this means for a data center ITAD program retiring GPU assets:
- End-user verification is mandatory, not optional. Your vendor needs a compliance department — not just a sales team — that screens buyers against the BIS Entity List and denied parties lists.
- Chain-of-custody extends past the first sale. Elite vendors track the chain of custody — the documented paper trail of who had the equipment, when, and what happened to it — well beyond the loading dock handoff, and contractually obligate downstream buyers to the same visibility.
- Documentation survives audit. BIS enforcement actions typically arrive two to five years after the initial transaction. If your vendor can’t produce the end-user verification and sale documentation on demand, you are exposed.
Political administrations adjust specific waivers — the H20 export debate being one example — but the underlying reality is stable: high-performance AI silicon is a regulated geopolitical asset. The secondary market for these chips will continue to operate under the shadow of trade law for the foreseeable future.
Not sure whether your current ITAD vendor meets BIS end-user verification standards? Take the free ITAD Readiness Assessment →
What Makes Data Sanitization on AI Hardware Different?
Two things: the storage architecture of the GPUs themselves, and the scale at which NVMe sanitization has to run.
Traditional enterprise storage — HDDs and standard SATA SSDs — can be sanitized with well-established protocols targeting known user sectors. AI accelerators complicate that materially. The NVIDIA H200 and B200 carry 141GB to 192GB of High-Bandwidth Memory (HBM3e) integrated directly onto the chip package. That memory is woven into NVLink switch fabrics that pool petabytes of data across entire server racks at terabytes per second. Erasing proprietary model weights, training data, and user prompts from those volatile memory pools and deep caching layers is non-trivial — standard wiping tools cannot reliably access, let alone verify sanitization of, embedded GPU memory or hidden over-provisioned sectors in high-density NVMe checkpointing arrays.
The standards framework has shifted to accommodate this. NIST SP 800-88 Revision 2, finalized September 26, 2025, superseded the 2014 Rev 1 and defers technology-specific sanitization execution to IEEE 2883:2022. For AI hardware refresh and GPU disposition workflows, the practical requirements are specific:
- Cryptographic Erase (Purge-level) via firmware commands, not software overwrite. Multi-pass overwrites don’t work on modern NVMe. The drive’s own controller destroys the media encryption key; the ciphertext becomes mathematically unrecoverable in milliseconds. Automated platforms like Blancco Drive Eraser execute the IEEE 2883 Purge command and generate a per-serial-number Certificate of Erasure.
- On-site sanitization at scale. At hundreds or thousands of drives per engagement, transit risk exceeds the operational cost of on-site work. Mobile erasure arrays parked at the data center, with failed-wipe drives routed immediately to an on-site shredder before exiting the security perimeter, is the pattern that scales.
- Physical destruction requirements specific to flash. IEEE 2883 explicitly holds that standard shredding is insufficient for high-density flash — NAND chips pass through conventional shredder grates intact. Pulverization, smelting, or incineration is required for true Destruct-level sanitization.
The Iron Mountain engagement that retired 297 server racks and 122,000 data-bearing drives for a hyperscale provider achieved 100% sector-verified elimination on site in four days. That velocity is only possible with fleet-level automated erasure, not manual wiping stations. Ask prospective vendors for comparable references at your scale.
How Do You Capture Value Before the Depreciation Window Closes?
Speed and specialization. In that order.
The AI supercycle is being driven by hardware displacement, not additive expansion — which means the volume of decommissioned servers, switches, and storage flooding the ITAD market is growing faster than specialized remarketing capacity can absorb. Operators who move first and work with vendors built for this category capture pricing that erodes weekly for operators who wait.
Four operational levers:
- Compress discovery-to-extraction. Every week in staging is value lost. Asset discovery and CMDB reconciliation should run in parallel with physical decommissioning prep, not sequentially before it.
- Triage by residual value curve, not uniform policy. H100s and H200s are in the premium refurbishment category — pulling them on Day 1 of extraction, sanitizing via IEEE 2883 Purge, and routing to a certified refurbishment pipeline is materially different from handling the legacy CPU racks in the same hall. Triage has to be deliberate.
- Component harvesting on non-remarketable systems. HBM and DRAM prices have surged on AI demand. Chassis that can’t be remarketed whole can still return meaningful value through component harvesting — memory modules, CPUs, NVMe drives — executed by vendors with the testing infrastructure to sell components, not just whole systems.
- Revenue share contract structure over flat-fee. At scale, the most common hyperscale arrangement is consignment revenue share: the ITAD vendor absorbs logistics, sanitization, and remarketing costs and splits gross proceeds with the client, typically 70-80% to the data center operator. Revenue share aligns incentives; flat-fee structures don’t.
The data center ITAD market was valued at roughly $12.4 billion in 2024, with projections exceeding $28.7 billion by the early 2030s, driven almost entirely by hyperscale turnover. The pipeline is expanding. Vendor capacity at the specialized end of the market is not expanding as fast.
Where You Stand Matters
The organizations capturing value during the AI supercycle are not running a decommissioning process — they are running a specialized disposition program with compliance, sanitization, and remarketing workstreams built for a specific category of hardware moving on a specific timeline. The organizations losing value during the same cycle are applying generic ITAD playbooks to assets that don’t tolerate generic handling.
If you’re managing an AI infrastructure refresh right now — whether that’s displacing CPU halls to free up power, cycling Hopper-class GPUs into the secondary market, or planning for B300 inbound — the diagnostic questions are the same: Does your vendor have refurbishment capability, or just destruction capability? Does your vendor have a BIS-compliant end-user verification program, or just a sales desk? Does your vendor execute IEEE 2883 Purge commands at scale on-site, or ship drives to a facility for wiping? The answers separate the programs that recover capital from the programs that generate regulatory exposure.
Ready to assess your organization’s ITAD readiness? SureDispose’s free assessment evaluates your AI hardware refresh and GPU disposition program across compliance, sanitization rigor, and value recovery sophistication — then connects you with vetted providers built for hyperscale and AI-era infrastructure. Take the Assessment →
SureDispose is an independent advisory platform. We connect organizations with vetted ITAD service providers but do not perform disposition services directly. Providers compensate us for qualified introductions.