GPU & AI Hardware Buyback
Reuse-First buyback for AI accelerators and enterprise GPUs — NVIDIA H100, A100, A10, RTX 6000 Ada, AMD Instinct MI-series — quote validity 5 business days due to volatile AI secondary market, settled in MYR against PO.
OEMs covered
NVIDIA H100 · NVIDIA A100 (40GB / 80GB) · NVIDIA A10 / A40 · NVIDIA RTX 6000 Ada · NVIDIA L40 / L40S · AMD Instinct MI200 / MI300
Pricing notes
AI hardware secondary market is volatile; quote validity is 5 business days. Cross-border export classification handled per NVIDIA / AMD export-control regimes. Reuse-First Cryptographic Erase on accelerator memory; per-asset certificate retained.
What we see most of in Malaysia
How we process your engagement
Send your asset list. We respond with a written MYR quote in per engagement SLA (5 business days for AI accelerators). Pickup against signed manifest within Pickup scheduled per engagement, nationwide Malaysia. Wipe + functional test included as standard — drives sanitised to NIST SP 800-88 Rev. 1 Purge for spinning HDDs and IEEE 2883-2022 firmware Sanitize for SSDs/NVMe. Per-asset Certificate of Destruction issued, line-item invoicing per asset, settlement against your PO. Multi-site programme engagements consolidate to a single MYR settlement line; single-event engagements settle within 7 business days of manifest reconciliation.
Condition grading rubric — how we set the offer
AI accelerators grade on a market-volatility-aware rubric. Grade A: powers on at the host, all VRAM detects, ECC clean, no thermal-throttling history (where the host driver retains telemetry), original packaging or equivalent. Typical recovery for current-generation accelerators: 60–88% of original list (NVIDIA H100 80GB clears highest; A100 40GB clears at 35–55%; AMD MI300 / MI250 clears at programme-specific rates depending on geographic demand). Grade B: cosmetic wear on cooling shroud, minor scuffs, no functional defects. Typical recovery: 50–75% of original list for current-gen. Grade C: VRAM ECC errors, thermal-throttling history, partial-function state. Recovered at parts-pulls economics or for refurb where the OEM permits. Grade D: damaged, fire damage, water damage. Material recovery only. Crucial market dynamics: AI accelerator prices reprice weekly on supply, model-release timing, and export-control news; quote validity is 5 business days, not 30. Export-classification: cross-border resale requires per-unit BIS classification and (for top-tier accelerators) export licence in some jurisdictions; this is handled by our AI Hardware Desk before settlement.
Typical recovery ranges — current secondary market
| Accelerator | Age band | Typical Grade A recovery (highly volatile) |
|---|---|---|
| NVIDIA H100 80GB SXM5 | 0–18 months | 60–82% |
| NVIDIA H100 80GB PCIe | 0–18 months | 52–72% |
| NVIDIA H100 NVL (94GB) | 0–18 months | 55–75% |
| NVIDIA A100 80GB SXM4 | 18–36 months | 32–48% |
| NVIDIA A100 40GB PCIe | 24–48 months | 22–35% |
| NVIDIA A40 / A30 / A10 | 24–48 months | 18–32% |
| NVIDIA RTX 6000 Ada | 0–24 months | 40–55% |
| NVIDIA L40S / L40 | 0–18 months | 34–48% |
| NVIDIA RTX A6000 | 24–48 months | 22–34% |
| AMD Instinct MI300X | 0–18 months | 50–70% (geo-dependent) |
| AMD Instinct MI250X | 18–36 months | 28–42% |
| Intel Gaudi3 / Gaudi2 | 0–24 months | 22–38% |
All AI accelerator pricing reprices weekly. Quote validity: 5 business days. Settlement at the validity-window price. Cross-border export classification handled per US BIS rules and equivalent local regimes (UAE Federal Export Control Authority, India DGFT, Singapore Strategic Goods Control, Canada Export Controls List).
Sanitisation specifics — what the certificate cites
AI accelerators sanitise differently from storage because the data is in volatile VRAM, not in persistent storage. VRAM clearing: GPU memory is cleared at host power-down; the sanitisation step is to verify the accelerator is power-cycled and host-side telemetry confirms clean state. NVIDIA Cryptographic Erase: where the engagement requires a documented sanitisation event (BFSI, government, regulated R&D), we issue NVIDIA-supported Cryptographic Erase on the accelerator; the certificate cites the NVIDIA-driver API call and the verification response. AMD Instinct equivalent: AMD's ROCm-side memory-clear command issued and verified. Firmware-resident metadata: NVLink topology, MIG (Multi-Instance GPU) configuration, and any firmware-resident host-pairing data are explicitly cleared. BIOS/host firmware: the host's GPU enumeration cache is cleared; the accelerator is independent of the host on detach. Cooling shroud / thermal-pad inspection: shroud is inspected for damage; thermal pads are re-applied where degraded. Export classification: every accelerator is classified per US BIS rules and equivalent local regimes before refurb routing; the classification and the routing decision are logged.
Recent engagement scenarios (anonymised)
Scenario 1 — H100 cluster refresh at AI startup, 64 H100 SXM5. A growth-stage AI startup refreshing from H100 to H200 retired 64 H100 80GB SXM5 accelerators (mounted in 8 DGX H100 systems). 5-business-day quote validity due to volatile market; quote issued and accepted within window. Cross-border export classification completed; routing through trader channel (Hong Kong + UAE) compliant with US BIS rules and local regimes. Per-accelerator certificate citing NVIDIA Crypto Erase. Settlement in MYR against PO at the validity-window price.
Scenario 2 — A100 secondary-market resale, 24 A100 40GB PCIe. A research lab retiring 24 A100 40GB PCIe accelerators on cutover to a managed cloud platform retired the accelerators with chassis-host servers retired separately. AI Hardware Desk quoted on 5-day validity; export-classification confirmed routing eligibility. NVIDIA Crypto Erase applied; per-unit certificate. Settlement in MYR reflecting the validity-window price (which had moved meaningfully over the prior 60 days as H100 supply expanded).
Scenario 3 — MI300 retirement at hyperscale tenant, 16 MI300X. A hyperscale tenant exiting an AI training deployment retired 16 AMD Instinct MI300X accelerators. AMD ROCm memory-clear applied; export-classification completed. Geographic-routing decision: MI-series demand is regionally concentrated, so routing optimised to the highest-demand region. Settlement in MYR at the validity-window price.
Pickup, chain of custody, and logistics
AI accelerator pickups operate on heightened security and tighter timelines than other asset classes. Operators arrive with anti-static, padded transit cases; each accelerator is photographed (top, bottom, connector edge), serial-scanned, and tagged. Accelerators are then placed in anti-static individual sleeves, then in tamper-evident sealed containers. GPS-tracked vehicle to facility. Per-accelerator photograph at receipt; sanitisation in-flight; per-accelerator certificate. Witness destruction option: where the customer requires accelerator destruction rather than reuse (uncommon, but sometimes specified for top-classified AI workloads), particle size is selected against the data classification. Cross-border routing: accelerators that route across borders trigger the export-classification workflow; settlement is gated on classification completion. Validity-window price-lock: the MYR quote is valid for 5 business days; pickup is scheduled inside the validity window so the customer settles at the quoted price, not at the next reprice.
Common pitfalls that erode recovery
Pitfall 1 — Quote expires before pickup. The single most common source of value erosion. AI accelerator prices reprice weekly; a 14-day delay between quote and pickup can move the settlement price by 8–18%. We schedule pickup inside the 5-day validity window. Pitfall 2 — Export classification not started early enough. Top-tier accelerators (H100, MI300X) require export classification before cross-border routing; we start classification at quote-acceptance. Pitfall 3 — Accelerators de-mounted from host but not photographed. Where the customer de-mounts and ships the accelerators separately from the host servers, photographs at de-mount are required for the chain-of-custody record. We supply de-mount procedure documentation if required. Pitfall 4 — NVLink topology not documented. NVLink-paired accelerators (in DGX systems) sometimes ship without the NVLink-pairing record; we re-document at receipt. Pitfall 5 — Customer wants the GPUs destroyed but routes through buyback. Where the customer's engagement requires destruction, that overrides Reuse-First; we destroy at the specified particle size and the certificate cites destruction rather than reuse.
Why customers in Malaysia choose Maxicom for gpu / ai hardware buyback
Continuous operation since 1996 — the Maxicom group was founded in India in 1996 and operates across Malaysia, plus four other Maxicom regions. Per-asset certificate format admissible against , PDPA Malaysia, NIST SP 800-88 Rev. 1, IEEE 2883-2022, and equivalent regulators in every jurisdiction we have served. Reuse-First reuse rate of 65–75% blended across our 2024–2025 cohort — typically 1.5–3× the residual value of destruction-first OEM trade-in programmes. Settlement in MYR against your purchase order, line-item per asset, payment terms 7 business days from manifest reconciliation. Cross-border resale routing under NDA preserves channel-respect — your retired gear never re-appears in your local market unless you specifically authorise local resale. Programme-level pricing for multi-event commitments; quarterly business reviews for ongoing engagements. Single SOW, single ledger, single regulator-facing report — your fixed-asset team reconciles to one document.
مراجع موثوقة
Primary sources for the standards and frameworks referenced on this page. Maxicom maps every engagement to these recognised authorities.
Frequently asked questions
How is the buyback offer priced?
Against the current secondary market for the specific make, model, configuration and condition. Settlement in MYR, against your purchase order, line-item per asset.
Do you take partial racks or partial fleets?
Yes — single units to multi-rack programmes. We do not require whole-cage commitments. Mixed-OEM fleets are accepted under a single SOW with line-item per-OEM pricing.
What about data destruction?
Wipe to NIST 800-88 Rev. 1 (for spinning HDDs) and IEEE 2883-2022 (for SSD/NVMe) is included as standard. Per-asset Certificate of Destruction issued. Witness destruction available where the engagement requires.
How long is the quote valid?
5 business days for AI accelerators (volatile market repricing weekly); 30 days for steady-state enterprise hardware. We re-quote on request.
Will my surplus appear in the local market?
Where local demand cannot absorb at fair price, working units route cross-border through our trader-channel network — never back into your own market unless you specifically authorise local resale. Channel-respect is part of the engagement contract.
Do you handle export documentation?
Yes. Cross-border resale routing includes export classification and documentation; particularly relevant for AI hardware subject to US BIS export controls and equivalent local regimes (UAE FECA, India DGFT, Singapore Strategic Goods Control, Canada Export Controls List).
Why is AI hardware quote validity only 5 business days?
The AI accelerator secondary market reprices weekly on supply, model-release timing, and export-control news. H100 prices fell 25–35% from peak through 2025 as supply increased; A100 prices fell substantially as H100 became broadly available. A 30-day quote validity would expose either us or the customer to price moves that misalign the settlement to the validity-window price.
How do US BIS export controls affect AI accelerator buyback?
BIS export controls regulate cross-border resale of certain accelerators (currently including H100 and equivalents). Resale to specific jurisdictions requires export licences; resale to other jurisdictions requires only classification. Our AI Hardware Desk handles the classification workflow and routes accordingly. Settlement is gated on classification completion to ensure the resale is licence-compliant.
What about consumer-grade GPUs (RTX 4090, RTX 5090)?
We accept enterprise-grade accelerators primarily; consumer-grade GPUs are accepted programme-level where the customer is retiring a workstation fleet that includes them. The secondary market for consumer GPUs is more liquid but at lower per-unit value than enterprise.
Do you handle PCIe-form vs SXM-form differently?
Yes. SXM-form accelerators are tied to specific OEM motherboards (DGX, HGX) and clear at higher per-unit recovery because they are paired with high-value host hardware. PCIe-form accelerators are host-agnostic and clear in a wider secondary channel.
How are NVLink-paired accelerators in DGX systems handled?
DGX systems are typically retired as a unit (chassis + 8 accelerators) and quoted at system-level pricing. Standalone accelerator pulls (where the customer has retired the chassis separately) are quoted at per-accelerator pricing.
Can you handle accelerators from sovereign-data engagements?
Yes. Sovereign-data engagements with strict data-residency constraints can run with on-site sanitisation by our operators under customer-witness, with the accelerators routing through a sovereignty-compliant trader channel only. The AI Hardware Desk has handled engagements of this profile in Malaysia.
How fast does the GPU secondary market move?
Quote validity is 5 business days because the AI accelerator secondary market re-prices weekly. H100 and A100 demand is currently high, but supply is increasing, and prices are softening from peak.
Do you handle export documentation?
Yes. AI hardware export is increasingly subject to local export controls (US BIS rules, UAE and Singapore alignment, India SCOMET). We handle export classification and documentation for cross-border resale.
Related practices, regulators & markets
IT Asset Disposal (ITAD)
ITAD
→Data Destruction
Data destruction
→Dell Server Buyback
Dell server buyback
→HPE Server Buyback
HPE server buyback
→Banking & Finance
Banking
→Government & Public Sector
Government
→NIST SP 800-88 Rev. 1
NIST 800-88
→IEEE 2883-2022
IEEE 2883
→IT disposal in Kuala Lumpur
Kuala Lumpur
→Send the asset list. We will send the number.
A photograph of the rack works. A spreadsheet works better. MYR settlement, against PO.