🔥 Clearance
Previous-generation NVIDIA hardware, still genuinely useful, marked down 20% — plus buy two, get the second one an additional 20% off.
How the discount stacks, shown with math: take 20% off the list price for item one.
Buy a second clearance item in the same order and it's an additional 20% off its already-discounted price —
so two of the same card costs
clearance_price × 1.8, not × 2. Every card below shows this worked out.
Wikimedia Commons
Desktop GPU
-20% CLEARANCE
NVIDIA GeForce RTX 3090
Previous-gen 24GB card — last-gen silicon, still enough memory for plenty of local AI work
☆☆☆☆☆ No ratings yet
List price
$1,488
Clearance price
$1,190
What you'd pay today, new-old-stock, for one card, before the clearance discount. Street price as of July 2026 (launch MSRP was $1,499 in 2020). Now superseded by the 40- and 50-series, hence clearance.
Memory24 GB GDDR6X
Same 24GB as an RTX 4090 — fine for models that fit, just slower to compute once they're loaded.
Power draw350 W
100W less than a 4090 at full tilt — an easier card to power and cool.
Buy 2 (BOGO 20% off #2):
$1,190 + $952 = $2,142 total
($1,071 average per card)
NVIDIA press image
Workstation GPU
-20% CLEARANCE
NVIDIA RTX A6000
Previous-gen 48GB workstation card — half the memory of the RTX PRO 6000 for well under half the price
☆☆☆☆☆ No ratings yet
List price
$5,015
Clearance price
$4,012
Ampere-generation, one step behind the RTX PRO 6000 Blackwell — the discount reflects being a generation old, not being broken. Average street price as of 2026 based on historical price tracking (range $4,100-$7,371 depending on condition/seller).
Memory48 GB GDDR6 (ECC)
48GB is exactly what two of these need to match one RTX PRO 6000 — sometimes two older cards beat one new one on price per GB.
Power draw300 W
Lower power draw than the Blackwell-generation workstation card, at the cost of raw speed.
Buy 2 (BOGO 20% off #2):
$4,012 + $3,210 = $7,222 total
($3,611 average per card)
Wikimedia Commons / NVIDIA asset
Datacenter GPU
-20% CLEARANCE
NVIDIA A100 40GB PCIe
Previous-gen datacenter GPU — the card that trained the last generation of frontier models
☆☆☆☆☆ No ratings yet
List price
$10,000
Clearance price
$8,000
A fraction of an H100's price for real (if slower) datacenter-grade AI compute. List price around $10,000 new; verified refurbished/enterprise-surplus units run $3,500-$7,200. $10,000 used here as the pre-clearance list price.
Memory40 GB HBM2e
40GB HBM2e — less than an H100's 80GB, but still far faster memory than any desktop card.
Power draw250 W
The PCIe version runs at 250W typical, well under the 700W of an H100 SXM.
Buy 2 (BOGO 20% off #2):
$8,000 + $6,400 = $14,400 total
($7,200 average per card)
NVIDIA press image
Datacenter GPU
-20% CLEARANCE
NVIDIA L40S
Datacenter inference card, passively cooled for server racks
☆☆☆☆☆ No ratings yet
List price
$8,500
Clearance price
$6,800
Positioned between workstation cards and full H100s — built for inference racks, not training clusters. Purchase price typically $7,500-$10,000; $8,500 used as a representative mid-point.
Memory48 GB GDDR6 (ECC)
48GB GDDR6 — more memory than an H100 costs per GB, though slower bandwidth than HBM.
Power draw350 W
Passively cooled — this card relies entirely on server chassis airflow, it has no onboard fan.
Buy 2 (BOGO 20% off #2):
$6,800 + $5,440 = $12,240 total
($6,120 average per card)