AI hardware buyback in Canada: what your retired H100, A100, L40S are worth.
The NVIDIA GPU secondary market is volatile, and AI-hardware buyback is newer territory. H100, A100, and L40S residual values fluctuate weekly based on used-equipment supply, training workload demand, and cloud-provider inventory. This explainer covers residual-value drivers, how hourly pricing works, what NVMe-storage destruction discipline means for AI servers, and where AI hardware re-exports to in North America.
Why AI hardware prices move fast
AI hardware residual values are driven by supply and demand in real time, not historical pricing tables. When a new generation of NVIDIA GPU is announced, H100 prices drop. When data-centre hyperscalers reduce their GPU order books (due to market saturation or capex constraints), used H100 supply floods the market and prices compress further.
Three factors move AI-hardware prices hourly:
- New-unit announcements. H200, H300, or a surprise AMD MI300X release can swing used H100 pricing 5–15% overnight.
- Cloud-provider capex cycles. When AWS, Azure, or Google announce GPU provisioning reductions, the used market floods with off-lease H100s. Prices compress by 10–20% within days.
- Regional demand imbalances. High demand in Canada + Malaysia pulls prices up. Oversupply in Hong Kong pushes prices down. Cross-regional pricing spreads create arbitrage opportunities.
For your 50 retired H100s, pricing is current-market, not 'book value.' We quote hourly, not weekly. A quote locked on Monday may be stale by Friday.
What factors into an H100 / A100 buyback price
Here's what we assess when quoting your retired AI hardware.
- GPU model & generation · H100 80GB > H100 40GB > A100 80GB > A100 40GB > L40S. H-series (newest) commands higher residual than A-series (2022). L40S is lower-demand (specialist graphics).
- Hours-on-clock · A used H100 with 2,000 operating hours vs 20,000 hours affects resale desirability. We verify via logs or OEM service records.
- Condition & thermals · Cosmetic damage doesn't affect value much, but thermal throttling or VRAM errors reduce residual by 10–30%.
- Stock availability · Rare units (e.g. early H100 80GB NVLink units) command premiums. Common units (H100 40GB, bulk supply) see margin compression.
- Regional demand · Canada + Malaysia buyer base is strong (data-centre expansion). Indonesia + Vietnam is weaker (budget constraints). We adjust price for regional destination.
- Seasonal capex cycles · Q4 and Q1 capex budgets drive demand up. Q3 typically sees pricing pressure as enterprises defer purchases.
Our pricing methodology
Real-time market feed
We track 3 secondary-market price indices (Asian resellers, US brokers, OEM used-equipment programs) hourly. Prices are weighted by volume and geography.
Device-level assessment
You send asset list + condition report (cosmetics, thermals, hours-on-clock). We apply condition factors to market price. H100 pristine with 1k hours gets baseline. H100 with 20k hours and cosmetic dings gets 15–25% haircut.
Quote window: 4 hours
Quote is locked for 4 hours. If you accept within that window, price is guaranteed. After 4 hours, we re-quote (market may have moved). If you accept, we schedule pickup within 5 business days.
No consignment
Price is firm at pickup. We take delivery risk, not you. If the device fails post-pickup (undisclosed VRAM error), we cover it. You're protected.
NVMe destruction discipline for AI-hardware return
AI servers (DGX, H100 nodes, Lambda Labs units) often have NVMe boots or attached storage with training data, model weights, or API keys. Here's how we handle them.
- ♦ Pre-assessment: confirm if the server has NVMe storage and whether it's encrypted or unencrypted.
- ♦ Owner data inventory: identify any training data, model weights, or API credentials stored on the NVMe. Some may need to stay with you (licensed model weights); others are junk data that needs destruction.
- ♦ Method decision: wipe (TRIM/secure-erase) or shred the NVMe before the H100 GPU is re-exported.
- ♦ For refurbished re-export: wipe the NVMe, wipe any boot media, leave the GPU clean. Device is ready for re-buyer's own OS install.
- ♦ Per-device log: serial number of GPU, serial of NVMe, method used (wipe / shred), timestamp, operator sign-off.
- ♦ Certificate inclusion: CoD lists GPU and storage separately, methods per device.
Where AI hardware re-exports to in North America
Canada is the hub, but the re-markets are spread across North America. Here's where your retired H100s and A100s typically flow.
- Canada (40–50%): Re-marketers, GPU-rental startups, research institutes. Fast turnover, high demand. Best prices.
- Malaysia (20–25%): Data centres in KL and Petaling Jaya. Growing AI workload. Decent demand, stable pricing.
- Indonesia (10–15%): Jakarta data centres + university research. Lower budget, slower turnover. Prices 5–10% lower than SG.
- Thailand (5–10%): Bangkok tech hubs, export to other regions. Niche demand. Prices pressure.
- Vietnam (5–10%): Hanoi/Ho Chi Minh City data centres. Growing but still lower budget than Malaysia. Prices comparable to Indonesia.
We route devices to the buyer offering the best price, unless you have a preferred destination (e.g., you want them to stay in SG). Most sellers accept best-price routing; it maximizes residual value.
Maxicom Canada — frequently asked
How often do you re-quote? If prices drop overnight, does our quote still hold?
Your quote is locked for 4 hours from issuance. After 4 hours, if you haven't accepted, prices may have changed and we'll re-quote. Once you accept and we schedule pickup, the price is firm — locked until delivery. Market-price movement after your acceptance doesn't change the price you locked.
We have 50 H100s with different hours-on-clock and cosmetic conditions. Do we get one price or a per-unit price?
Per-unit price is the fairest. Send us a detailed condition report (serial, hours-on-clock, any thermal issues, cosmetic damage) for each unit. We quote each one individually. You get a bulk price (e.g., 'Units 1–30 at CAD X; Units 31–50 at CAD Y–5% due to age'). That's more work upfront for you, but you get the most accurate value.
If our H100 has a VRAM error or fails the thermals test after pickup, can we negotiate price down?
No negotiation after pickup. The quote assumes normal condition within the parameters you provided. If you tell us 'pristine, 2k hours' and there's a hidden VRAM error, that's a disclosure issue. We'll work with you on a solution (repair credit, return), but the price doesn't renegotiate post-pickup. If you're unsure about condition, request a pre-pickup third-party assessment — we can arrange that.