What happens to secondary market GPU prices when new chips launch

When a new Nvidia generation launches, secondary-market prices for the previous generation drop sharply and quickly rather than declining gradually, with data center GPUs typically losing 30-60% of value within 3-6 months of the successor becoming widely available.

The Pattern: Sudden Resets, Not Gradual Decline

Data shows GPU depreciation doesn't follow a smooth curve but instead moves in step-function drops tied directly to new architecture announcements and availability, with a pronounced "mid-life cliff" around year three. Asset recovery industry analysis finds resale values drop 30-40% within six months of a new architecture being announced, and every 90-day delay in deciding to sell adds another 15-30% loss in secondary-market value. The H100 illustrates this vividly: cards that traded near $50,000 at the peak of 2024's supply shortage fell to a $6,000-$22,000 range once supply normalized and Blackwell-generation chips began shipping.

Real Numbers From the H100-to-Blackwell Transition

Exit Technologies auction data shows Nvidia's B100 launch triggered a 34% depreciation in H100 SXM5 cards by January 2026. UK reseller Servnet reports H100 cards that sold for around $40,000 in late 2023 now trade for roughly $12,000-$22,000 used, with forecasts of a further 10-20% drop as fleets rotate onto Blackwell hardware. Mercatus' residual-value curve shows H100 cards retain 75-85% of a $30,000 reference price at 18-30 months old, but that value collapses to a 25-35% floor by 60 months, driven by exactly this kind of generational cliff.

The A100 Precedent

The A100, now two generations removed from Blackwell, shows how far value can fall over time: the 40GB variant that once sold new for over $15,000 now trades at just $8,000-$12,000, and the 80GB variant sits at $12,000-$18,000. By 2026, A100 cards trade at just 30-50% of original list price, having already lost 50-60% of value by 2024 alone. This step-down pattern strongly suggests H100 and even current Blackwell-generation chips will follow a similar multi-year decay once the next architecture (Rubin) ships in volume.Why Rental Rates and Resale Prices Sometimes Diverge

Cloud rental rates for older GPUs decline at a fairly steady 26% per year with age, but interestingly, resale prices and rental rates don't always move together. Research from American Compute found periods where H100 rental rates fell even as resale prices rose, driven by scarcity of the data-center "shells" (power, cooling, networking) needed to actually deploy the chips rather than chip value itself. Separately, on-demand cloud pricing tends to fall gradually rather than in cliffs—typically 5-15% within weeks whenever a major provider brings a large new cluster online—since providers compete on utilization rather than reacting to hardware generation shifts alone.

Price Trajectory Comparison

Forward Outlook

Analysts expect further downward pressure as B200 and GB300 systems ramp, with forecasts pointing to an additional 10-20% drop in H100 secondary pricing once Blackwell availability broadens further, and specialist cloud providers are projected to offer H100 80GB SXM capacity at $1.49-$1.99 per hour by mid-2027. This dynamic is exactly why Nvidia's $500 billion financing plan is contentious: lenders extending multi-year credit against GPU collateral are betting against a well-documented historical pattern of sudden, generation-triggered price cliffs rather than smooth depreciation.exittechnologies+3