AI Hardware Desk

Why AI hardware needs its own desk

AI accelerators are the most volatile residual-value asset in our pipeline — H100 prices fell 25-35% from peak through 2025 as supply increased; A100 prices fell substantially as H100 became broadly available; older V100 holds niche value but is no longer competitive at high-end AI workloads; the upcoming B100/B200 and MI400 announcements will set new benchmarks. Standard enterprise-server buyback desks quote per engagement SLA; AI hardware requires 5 business days because the price moves weekly on supply, model-release timing, export-control updates, and geopolitical news. The dedicated AI Hardware Desk maintains live trader-channel pricing across UAE, India, Singapore, Canada and Hong Kong, plus an export-classification function that ensures cross-border resale is licence-compliant before the trade closes.

What we accept and what we route through

NVIDIA H-series — H100 (80GB SXM5, 80GB PCIe), H200 (141GB SXM5). NVIDIA A-series — A100 (80GB SXM4, 80GB PCIe, 40GB SXM4), A40, A30, A10. NVIDIA RTX professional — RTX 6000 Ada, RTX A6000, RTX A5000. NVIDIA L-series — L40, L40S, L4. AMD Instinct — MI300X, MI300A, MI250X, MI210, MI100. Intel Gaudi — Gaudi3, Gaudi2. Plus full-system GPU servers (HGX baseboards, DGX systems, Dell PowerEdge XE, HPE Apollo, Supermicro, Lenovo SR670). Card-level and system-level engagements both accepted; pricing per SKU.

NIST SP 800-88 Cryptographic Erase on accelerator memory

NVIDIA H100 / A100 accelerators have on-board HBM (High Bandwidth Memory) that retains data through power cycles in some failure modes. NIST SP 800-88 Rev. 1 directs to Cryptographic Erase where the accelerator supports it. Maxicom executes the NVIDIA Confidential Computing key destruction (where SEV-SNP / TEE keys are in use), then runs the on-board memory test pattern that overwrites HBM content. For H100, the Confidential Computing key destruction is via the NVIDIA NVSwitch + driver stack. AMD Instinct MI-series: similar protocol via the AMD ROCm stack. Per-accelerator certificate documents the form factor, on-board memory state at retirement, and key-destruction method.

US BIS export controls and cross-border routing

NVIDIA H100, H200, A100 and similar high-end accelerators are subject to U.S. Bureau of Industry and Security (BIS) export controls. The October 2022 rules introduced thresholds; the October 2023 rules tightened them further; subsequent BIS notices have refined coverage. Cross-border resale of these parts requires export classification (ECCN), end-user verification, and in some cases licensing. The AI Hardware Desk handles classification for the routing path, requests end-user statements where required, and declines transactions where the export licence path is not clear. We do not route AI hardware to entities on the BIS Entity List or comparable restricted-party lists. This protects the customer from downstream-buyer compliance issues.

Engagement profiles we serve through the AI Hardware Desk

AI startup refresh cycles — typically 9-12 month cycles, single-vendor (NVIDIA) clusters, growth-stage cash recovery to extend runway. BFSI ML platform refreshes — typically 18-month cycles, mixed asset class (GPUs + CPUs + storage), regulator-aligned destruction trail. Sovereign-fund-backed lab refreshes — multi-vendor, longer cycles, NDA-bound. Hyperscale tenant exits — multi-hall AI cluster pulls under access-window scheduling, paired with InfiniBand and storage decommissioning. Cross-border refurb supply — feeding our trader-channel relationships across MENA, ASEAN, North America, South Asia.

Buyback Index — how we set quote pricing

The AI Hardware Desk maintains an internal Buyback Index updated weekly: NVIDIA H100 80GB SXM5 reference price, A100 80GB SXM4 reference price, A100 40GB SXM4, A40, A30, A10, RTX 6000 Ada, L40S, AMD MI300X, MI250X. Reference prices are derived from active trader-channel pricing across our five operating regions, not from published guidance (which lags 2-4 weeks). Quote validity is 5 business days because that is the typical window before reference prices move enough to require re-quoting. Where the customer needs pricing certainty over a longer window, we offer locked-rate contracts at a 10-15% discount to reference.