Selling used GPU servers hands their drives, logs and management controllers to a new owner. NIST's Guidelines for Media Sanitization, SP 800-88 Rev. 2, count a resale as a change of control over storage media [1]. This explainer sets out what to do before you sell used GPU servers, in order, from NIST's guide, the makers' documents, EPA's page on certified recyclers and the Export Administration Regulations (EAR).
Where a GPU server keeps data
Data sits in more places than the drives. Dell's PowerEdge XE9680 is an 8-GPU server [36]. Its statement of volatility, Dell's list of the server's memory, names BIOS settings in CMOS memory and a 32 MB BIOS flash that no utility or application can clear [2]. Its management controller's 4 MB flash and 16 GB eMMC hold firmware, user data and logs and cannot be cleared completely, and a Trusted Platform Module (TPM) stores encryption keys [2]. The statement leaves out add-in options that are not soldered to the motherboard [2].
NVIDIA's DGX H100/H200 carries two 1.92 TB drives in RAID 1 for the operating system and eight 3.84 TB self-encrypting drives in RAID 0 for data [14]. On DGX systems, the keys for those drives can sit in the TPM, where they persist even if the system is reimaged [15]. The baseboard management controller operates separately from the operating system and allows remote management even when the system is shut down [16]. Each GPU has an InfoROM, which NVIDIA describes as "a small, persistent store of configuration and state data for the GPU" [18].
NIST asks owners to count every type of storage in an item and to focus on the non-volatile kind, which keeps its contents after power is removed [1]. The maker's statement of volatility is where that list starts, though NIST notes that "vendors can state volatility details differently" [1].
Save the health records before any erase
A buyer judges a used GPU by its history, kept on the GPU and in the server's logs. In nvidia-smi, aggregate ECC error counts "persist indefinitely and thus act as a lifetime counter", and clearing them "is not supported on Ampere+" [18]. Remapped memory rows are recorded in the InfoROM [19]. For a GPU issue reported to the system vendor, NVIDIA's triage guide lists nvidia-bug-report.sh output, Fabric Manager logs on HGX systems and Data Center GPU Manager diagnostic logs [22]. The full set a buyer works through is in what to check before buying a used GPU server.
Order matters. After Dell's server-wide System Erase, "All previous information is removed" from the Lifecycle Log, which keeps only the erase request and what Dell's management controller logs as it restarts [3]. NIST advises consulting records retention offices before sanitizing [1]. Save serial numbers, counters and logs first, then erase.
Those records also bear on price. What sets the price of a used H100 server ties a used unit's price to its condition, warranty and diagnostic records, and evidence classes in GPU pricing keeps listings, quotes and verified transactions apart when reading any resale figure.
Clear, purge and destroy under NIST SP 800-88 Rev. 2
Revision 2, final in September 2025, superseded the 2014 Revision 1 and, apart from cryptographic erase, replaced its technique details with recommendations to follow IEEE 2883, National Security Agency specifications or an approved organizational standard [1]. It defines three methods [1]. Clear uses logical techniques against simple recovery through the interface available to the user. Purge defeats state-of-the-art laboratory recovery and leaves the media potentially reusable. Destroy defeats the same recovery and leaves the media unusable for storage [1].
"When possible", NIST says, purge should be used instead of clear [1]. In NIST's decision flow, media not meant for reuse is destroyed, and reused media that leaves the organization is purged, unless its data has a high security categorization, in which case the flow sends it to destroy [1]. NIST, which abbreviates information storage media as ISM, names "selling or donating the ISM" among the reasons purge can be more appropriate than destroy [1].
NIST's FAQ says "Revision 2 clarifies that multi-pass overwriting is unnecessary" [11]. Degaussing should not be used on flash, and on an SSD it "can complete successfully, but no sensitive data is sanitized" [1]. Cryptographic erase sanitizes the keys that encrypt the data [1]. NIST says not to trust it on media that have been backed up or escrowed unless the owner has a high level of confidence in how and where the keys were stored and managed outside the media [1].
On NVMe drives, the sanitize command offers block erase, overwrite and crypto erase, and its Purge Required option makes a supporting controller fail any operation that does not purge as IEEE 2883 defines [20]. Other stores need resets. The DMTF Redfish standard defines a factory reset for management controllers, with one type that keeps network settings and local accounts [17]. For devices that offer only a factory reset, NIST counts it as clear "as long as the device interface available to the user does not facilitate retrieval of the original data" [1]. NVIDIA's DGX guide gives the BIOS path to clear the TPM [15].
The certificate of sanitization
NIST's assurance step has two parts [1]. Verification checks the outcome, which for logical techniques means "checking the completion status of the tools and identifying errors, anomalies, and the health of the ISM" [1]. Validation decides whether the target data was effectively sanitized, and a rejection means repeating with a different technique or escalating to a more secure method [1]. On NVMe drives, the sanitize log reports success or failure, and its Purged bit shows that the operation purged user data as IEEE 2883 defines [21].
NIST says "a certificate of sanitization (see Appendix C) should be completed for each ISM that has been sanitized" [1]. It records the manufacturer, model, serial number, media type, method, technique, the tool and its version, the verification method and the people who verified and validated the result [1]. The sample form's destination field offers "Internal Reuse, External Reuse, Recycling Facility, Manufacturer, Other" [1]. A drive sold with its server is external reuse.
Warranty and software terms on resale
Each maker publishes its own warranty terms [4, 5]. DGX, HGX and MGX explained sets out how NVIDIA's own DGX systems differ from other makers' HGX servers. Supermicro's published warranty applies only to the original purchaser [4]. Dell's system warranties may be transferred when the current owner records the transfer with Dell, at Dell's sole discretion, and support applies only in the country where Dell originally sold the system [5]. NVIDIA support entitlements are non-transferable and may be terminated when the product changes hands [6]. NVIDIA ties the software subscription bundled with some of its PCIe and NVL cards to the specific card, with a start date that cannot be changed [23]. Record which of these terms apply before the sale, so the buyer knows what cover comes with the unit.
Recycle what is not reused
EPA names two accredited certification standards for electronics recyclers, R2 and e-Stewards, and both "require destruction of all data on used electronics" [7]. EPA recommends certified recyclers for unwanted used electronics [7]. Sustainable Electronics Recycling International describes R2v3's data sanitization appendix as applying to facilities that perform logical data sanitization [24]. A drive sent to a recycler gets a certificate like any other, with the destination marked Recycling Facility [1].
Export rules when you sell used GPU servers
This section is general information on published rules, not legal advice. The EAR covers all items in the United States and U.S.-origin items wherever located, and a transfer (in-country) is a change in end use or end user within the same foreign country [10]. Export Control Classification Number (ECCN) 4A090 covers computers containing chips that meet the thresholds of ECCN 3A090 [29]. The Commerce Department's Bureau of Industry and Security (BIS) names "servers classified as ECCN 5A992.z" among the items its guidance covers [9].
General Prohibition Ten bars selling an item subject to the EAR with knowledge that a violation has occurred, is about to occur or is intended to occur, and it has no license exceptions [8]. Knowledge includes awareness of a high probability, which can be inferred from willful avoidance of facts [28]. BIS's Know Your Customer guidance says "Do not self-blind" [32].
BIS's guidance of 13 May 2025 lists 11 red flags for advanced computing chips [9]. Two name domestic as well as foreign customers, one of them a residential address with quantities inconsistent with personal use, two more apply to parties wherever located, and another is an unknown delivery or installation address [9]. Among the information to ask a customer, BIS lists whether the items are to be resold and to whom [9]. Buyers can be checked against the Consolidated Screening List, which combines export screening lists of the Departments of Commerce, State and the Treasury [33]. A seller that exports also puts a destination control statement on the commercial invoice and keeps records for five years [34, 35]. The license text in the EAR and BIS's 31 May 2026 enforcement guidance differ, and the full report sets out both, dated [30, 31].
Rillor's GPU systems service resells GPU servers their owners no longer need, and tests, grades and documents each unit to the same standard before offering it.
Read the full report
The full report adds the statement of volatility table, NIST's certificate fields, the warranty terms in detail, the EAR entries that name GPU servers, a seller's checklist and all 36 sources.