NVIDIA A100 and H100 estates are coming out of racks faster than at any point since the AI buildout began. Most of that hardware still has years of economic life and a live secondary market waiting for it — but only if you move before the depreciation curve does. Here is what is in your rack, what it is worth, and how to recover it without creating a data security problem.

There is a specific kind of pallet showing up in data centers right now. Eight-GPU HGX baseboards pulled during a Blackwell migration. DGX A100 nodes displaced by newer racks. InfiniBand switches and ConnectX adapters left over when the fabric was rebuilt. NVMe shelves and DDR5 RDIMMs stripped from decommissioned hosts.

None of it is broken. Most of it is three to five years old, fully functional, and sitting in a corner because the migration team was measured on getting the new cluster up — not on what happened to the old one.

That pallet is likely the single most valuable idle asset your organisation owns. It is also depreciating faster than almost anything else on your books.

The Blackwell effect: why A100 and H100 estates are coming out now

The AI hardware refresh cycle has compressed dramatically. When B200 systems began shipping in volume, the economics of the previous generation reset almost overnight — AWS cut H100 on-demand pricing by 44% in June 2025, and the broader market followed.

The secondary market absorbed that shift rather than collapsing under it. Used H100 cards peaked near $50,000 during the 2024 scarcity window and have come down substantially as supply loosened. As of mid-2026, a used 8-GPU H100 server trades in roughly the $150,000–$180,000 range, against approximately $500,000 for a new B300 system — a 40–70% saving that keeps demand strong. New H100 80GB cards sit around $31,000, with new 8x H100 HGX systems between $250,000 and $320,000.

Older silicon has held up better than most people expect. Used A100s still trade between roughly $7,800 and $18,900 depending on SKU and condition, despite the platform being around six years old.

The reason is straightforward: roughly a third of AI workloads now run on neocloud operators rather than hyperscalers, and those operators build capacity on price-performance, not on having the newest part number. A great many inference workloads run perfectly well on A100, L40S and H100 hardware. Your decommissioned cluster is somebody else’s cost-effective capacity — and they can deploy it this month instead of waiting out a lead time.

The question is not whether your retired AI hardware has value. It is whether you capture that value or watch it decay in a storage room.

What is actually in the rack — and what carries recovery value

Decommissioning teams routinely undervalue everything that is not a GPU. In a full AI cluster teardown, the accelerators are usually 50–70% of recoverable value; the rest is in components that often get scrapped by default. A complete inventory should cover all of the following.

GPUs and AI accelerators

The core of the recovery. NVIDIA H100, H200, A100, L40S, L4, A40, V100 and T4; Blackwell-generation B200, B300, GB200 NVL72 and GB300 as those estates begin to turn over; AMD Instinct MI300X and MI325X; Intel Gaudi 2 and Gaudi 3. SXM modules, OAM baseboards and PCIe cards all trade, though on different curves — SXM and HGX assemblies are worth grading as complete units rather than parting out.

GPU servers and complete systems

NVIDIA DGX A100, DGX H100 and DGX H200; HGX H100 and HGX A100 baseboards; Dell PowerEdge XE9680, XE8640, XE9640 and R760xa; Supermicro GPU SuperServer and SYS-821GE class 8-GPU systems; HPE ProLiant DL380a and DL385 Gen11, Cray XD670; Lenovo ThinkSystem SR675 V3 and SR680a; Gigabyte G-series and ASUS ESC platforms. Complete, tested systems consistently recover more than the sum of their components.

Networking and interconnect

Frequently the most overlooked category. NVIDIA Quantum-2 InfiniBand NDR and Quantum HDR switches; Mellanox ConnectX-6, ConnectX-7 and ConnectX-8 adapters; BlueField-2 and BlueField-3 DPUs; Spectrum-X Ethernet; Arista 7800R, 7060X and 7050X series; Cisco Nexus 9000; Juniper QFX. Add 400G and 800G optical transceivers — QSFP-DD, QSFP112 and OSFP — plus DAC and AOC cabling, which carry meaningful per-unit value in volume and are almost always thrown away.

Storage, memory and drives

Pure Storage FlashBlade and FlashArray; NetApp AFF and ASA; Dell PowerScale, PowerStore and PowerMax; VAST Data and WEKA nodes; enterprise NVMe SSDs from Samsung, Micron, Kioxia and Solidigm; DDR5 and DDR4 RDIMM in volume. This is also the category with the highest data-security exposure, which is covered below.

Power, cooling and rack infrastructure

AI density changed what is in the room. Liquid cooling CDUs and rear-door heat exchangers, cold plate loops and manifolds; Vertiv, Schneider Electric APC and Eaton PDUs and UPS units; busway and high-amperage whips; 42U, 48U and 52U racks and cabinets. Bulky, unglamorous, and routinely written off at zero when much of it is redeployable.

Timing is most of the recovery

Useful life for AI hardware is genuinely contested. Amazon shortened its assumed server life to five years in early 2025; Meta extended to five and a half in the same period; some analysts argue for an aggressive two to three years, while real deployment evidence points to five to seven years or more of productive use.

For disposition purposes the accounting debate matters less than the market one. Secondary values move with supply, and supply arrives in waves as each new generation ships. Equipment released into the market ahead of a wave recovers materially more than identical equipment released into the middle of one. The practical consequence: the decision to decommission and the decision to remarket should be made at the same time, not eighteen months apart.

The data security problem specific to AI hardware

Standard ITAD checklists were written for general-purpose servers and miss several things that matter in a GPU cluster.

• GPU memory. Accelerators retain state in HBM until power is removed, and multi-tenant or checkpointed workloads can leave recoverable residue. Sanitisation should be explicit, not assumed because the node was rebooted.
• Out-of-band management. iDRAC, iLO, XClarity and BMC controllers hold credentials, network configuration and firmware that are rarely covered by a drive-wipe procedure and frequently ship out intact.
• Local scratch and checkpoint storage. Training checkpoints, embeddings and model weights are proprietary in a way that ordinary business data often is not, and they live on the NVMe tier that gets pulled and resold first.
• Encryption keys. Self-encrypting drives are only as secure as their key handling. A cryptographic erase that was never verified is not a sanitisation record.

NIST Special Publication 800-88 defines the three sanitisation levels — Clear, Purge and Destroy — that should be matched to media type and data sensitivity. R2v3, e-Stewards and NAID AAA certifications exist to verify that a vendor actually implements those methods, controls its downstream chain, and does not export non-working equipment. Ask which standard applies to which component, and require serial-level certificates of destruction.

Three outcomes worth measuring

A disposition programme should return results you can report to finance, facilities and compliance — not just an empty room.

• Recovered capital. GPUs, systems, fabric and storage are tested, graded and remarketed into the live secondary market rather than written off twice.
• Cleared space. Racks, cage space and power envelope come back for the next cluster, with removal scheduled around your operations rather than through them.
• Diverted waste. Whatever cannot be reused is routed to responsible materials recovery with tonnage documented — reportable for ESG rather than assumed. Globally, 62 million tonnes of e-waste was generated in 2022 with only 22.3% formally collected and recycled, and roughly US $91 billion in recoverable metals embedded in it.

Idle assets, back in orbit.

Orbital turns surplus AI and data center hardware into recovered capital, cleared space, and diverted waste — with serial-level chain of custody, certified data destruction and documented diversion reporting at every step.

You do not have to commit to anything to find out what your cluster is worth. Book a free consultation and we will walk your inventory with you — no charge, no obligation, no sales pressure. In about 30 minutes you will leave with:

• A current market recovery estimate for your GPUs, systems, fabric and storage
• A sanitisation approach covering HBM, BMC controllers and NVMe media, mapped to your compliance requirements
• A de-rack and removal plan that works around your operations, with timing and logistics
• A clear answer on what should be remarketed now, what should be held, and what should be recycled

If the numbers do not justify moving forward, we will tell you that too.

Book your free consultation  →  Call (833)255-7225 or email info@orbitalasset.com