GPU rents didn't follow the depreciation schedule—even the A100 costs more than at the start of the year

A six-year depreciation schedule has been defied by GPU rental prices.

a16z included two charts in Charts of the Week, with data from Ornn Data. The first compares rental prices for various GPU models with a standard six-year depreciation line. The conclusion is straightforward: rents have not depreciated as planned.

According to the gaps a16z lists, H200 rent is 143% above the six-year depreciation line, B200 111% above, RTX PRO 6000 73% above, H100 54% above, and A100 9% above. The chart is as of October 19, 2026. The A100 is a 2020 card; under six-year depreciation, it should long ago have been close to its book residual value, yet its rental price is still above the depreciation line.

The second chart looks at year-to-date changes: H200 roughly +160%, B200 roughly +120%, RTX PRO 6000 roughly +65%, and H100 roughly +35%. In a16z's words, even the rent for the six-year-old A100 is higher than at the start of the year.

a16z points to April as the inflection point. The article notes that OpenRouter token demand began climbing steeply in the spring, briefly flattened in late summer, and then jumped forward again. Tokens can be roughly understood as model invocation volume; as demand rises, GPUs used for inference become tighter.

But rising rents do not mean the chips themselves have appreciated. Commenter Sunny Gill notes that part of the increase comes from operating costs: power and financing costs are both rising, and even old cards need power to stay online. Another commenter, Gandalfo, puts it more directly: the premium may come from access to data center capacity with enough power and cooling; the silicon itself may not be the main point. Chips depreciate; scarce access does not.

For developers, commenter Abhishek Pan offers a concrete action: since A100 rents are still rising, rerun the cost per million tokens for your own inference stack every quarter. For small-model traffic, switching between A100 and H100, calculated at that month's rental rates, could be cheaper or more expensive.

This chart also has limits. Commenter frin cautions that Ornn Data cannot be used to characterize the entire GPU market, let alone directly infer hyperscalers' depreciation arrangements. The original a16z article also acknowledges that, over a longer history, A100 and H100 were below the depreciation line most of the time, and only clearly diverged after April 2026. Commenter Winston B's view is that if current rents persist, it would support the assumption of a longer useful life, and depreciation expenses would first show up in hyperscalers' financial statements.

One commenter joked that the depreciation schedule got the notice too late. Another said that people who used to collect Magic: The Gathering cards now collect compute—including himself.

The video is from commenter icefrog.◎, with the caption “My 3090 is paying its own rent.” The image contains only the words YOU UP?.

So now this rental curve looks more like an asset than a consumable. The question is how long it can last. One commenter says to check again in six months. That suggestion is not flashy, and it may be more direct than any opinion.

Source: a16z tweet; Charts of the Week: Attention is All You Need.