Two months ago, I wrote about the competition concerns with the GenAI infrastructure boom. One of my provocative claims was that the lifespan of the chips may be significantly shorter than the accounting treatment given to them. Others like David Rosenthal, Ed Zitron, Michael Burry and Olga Usvyatsky have raised similar concerns. NVIDIA has a response. When asked about chip lifespan, a spokesperson pointed to the secondary market where chips get redeployed for inference, general HPC, and other workloads across different kinds of data centers. In other words, older chips have buyers and the secondary market is robust enough to support resale prices that will justify the accounting treatment.
This appears to be the industry’s best argument. And the logic is sound in principle. Training is concentrated among a few frontier labs, while inference is distributed across the entire economy. Every enterprise deploying AI applications needs inference capacity. The “value cascade” that has training chips becoming inference chips, becoming bulk High-Performance-Computing (HPC) chips could be a reasonable model for how markets absorb generational transitions. If it holds, then a 5-6 year depreciation schedule might be justified, even if the chip’s competitive life at the frontier is only 1-2 years.
Let’s analyze whether that model holds up. For the secondary market thesis to justify current depreciation schedules, three things would need to be true:
1. The buyer pool must be large enough to absorb supply at meaningful prices.
NVIDIA’s data center revenue now exceeds $115 billion annually. The downstream buyers for older chips represent a market a fraction of that size. These buyers exist, but they didn’t scale with NVIDIA’s AI business and would not have the capacity to absorb all the supply. Moreover, most enterprises are not building their own inference capabilities. They are renting those services from hyperscalers.
2. New supply must not overwhelm secondary demand.
NVIDIA’s relentless pursuit of new chips on roughly annual cycles means each generation offers substantially better price-performance. The cascade model assumes orderly absorption at each tier: frontier buyers move to new chips, mid-tier buyers absorb their old ones, budget buyers absorb the generation before that. Each tier ideally clears before the next wave arrives.
But supply gluts break cascades. When new supply floods the market faster than downstream demand absorbs it, you don’t get orderly price discovery. Tesla is instructive here. When Tesla slashed new vehicle prices to maintain volume, used values didn’t gently adjust. They cratered 25-30% annually. Competitors explicitly refused to match the cuts because they understood the damage to their resale markets. The mechanism was simple: why buy used when you can buy new at the same price? The secondary market didn’t find a new equilibrium. It fell until it hit buyers with fundamentally different use cases; people who couldn’t afford new at any price.
NVIDIA isn’t cutting prices, but each new generation has a similar effect. Why rent a three-year-old H100 when Blackwell offers better price-performance? The generational improvement compresses what anyone will pay for old chips. And unlike cars, GPUs face this compression every 12-18 months. The cascade has to clear faster than NVIDIA releases new generations.
There’s a further problem. Each new generation doesn’t just offer more compute, it offers more compute per watt. In a power-constrained data center, a provider faces a choice: run older chips that generate $X in revenue per kilowatt, or replace them with newer chips that generate multiples of that on the same power budget. Once the performance-per-watt gap reaches a certain threshold, the older chip isn’t just worth less. It becomes uneconomical to run. If the electricity and cooling required to operate an H100 costs more than the market rate for the inference it produces, the chip’s residual value approaches zero regardless of what the depreciation schedule says.
3. The rental market should reflect robust secondary demand.
Another way to test NVIDIA’s argument is to look at pricing in the rental market. For most buyers, renting GPU capacity is functionally equivalent to purchasing into the secondary market. You get access to hardware without building data center infrastructure. If secondary demand were robust at current supply levels, rental prices would reflect that. Providers would be able to charge rates that justify their capital costs. Older chips would command prices proportional to their remaining useful life.
Pricing is notoriously opaque, but the indicators are that prices do not support the industry argument. H100 rental rates have fallen 70% from peak—from over $8/hour to around $2.50. According to Silicon Data as of December 17, 2025, H100 and A100 rental rates have converged to nearly identical levels—both around $2.50-2.60/hour. If the value cascade worked as described by NVIDIA, you’d expect tiered pricing: H100s at a premium and A100s at a discount reflecting lower capability and 3x lower cost. Instead, both generations have collapsed to the same floor. The hardware is effectively being given away as renters are just paying to keep the lights on. The efficiency gap makes this worse. Silicon Data index shows that the newer Blackwell chips are roughly 25x more efficient. Once Blackwell scales, the economics of running H100s in power-constrained data centers become untenable at any rental price.
One further constraint: chips need data centers. Buying a used H100 means you also need power, cooling, networking, and expertise to operate it. This limits secondary buyers to entities that already have that infrastructure, which overlaps heavily with the players contributing to rental oversupply. In other words, the potential sellers and potential buyers are substantially the same people. There is no overflow capacity sitting idle.
***
Putting these factors together suggests that relying on a robust secondary market for chips is illusory. None of this means older chips become worthless. They will find buyers. But they find them at prices that don’t support 5-6 year useful life assumptions. The gap between accounting depreciation and economic depreciation is the competitive subsidy I described in the original post. The secondary market argument doesn’t close that gap. It assumes it away.
Mihir Kshirsagar directs Princeton CITP’s technology policy clinic, where he focuses on how to shape a digital economy that serves the public interest. Drawing on his background as an antitrust and consumer protection litigator, his research examines the consumer impact of digital markets and explores how digital public infrastructure can be designed for public benefit.
Leave a Reply