A100 Rents Are Still Rising—Don't Rush to Write the GPU Depreciation Table

a16z's chart this week throws cold water on the GPU depreciation thesis. Under six-year straight-line depreciation, GPU rents should have marched steadily downward; in reality, they are broadly above that line.

The biggest gap is for H200, whose rent is 143% above the depreciation curve. B200 is 111% above, RTX PRO 6000 is 73% above, and H100 is 54% above. Even the A100, launched in 2020, is 9% above. a16z says most of the gap opened after April—exactly when token demand took off.

Another year-to-date chart is even more direct: H200 rents are up 160%, B200 up 120%, RTX PRO 6000 up 65%, and H100 up 35%. A100 rents are also higher than at the start of the year. In a16z's words, GPUs are aging like fine wine, with rents rising across the board.

When Elon Musk reposted this set of charts (https://x.com/elonmusk/status/2114205004340854795), he offered a hiring lens: think carefully about how many super-smart, focused people you want to hire—and they will keep getting stronger. One reply added wryly that this is a rare employee in history who gets a raise even after being fully depreciated. Another commenter put it more technically: compute is essentially hiring; every model update makes these 'employees' stronger, and GPU rents above the depreciation curve show that demand is running ahead of supply.

a16z lays out the data sources clearly in Charts of the Week (https://www.a16z.news/p/charts-of-the-week-attention-is-all): the rent index comes from Ornn Data and covers multiple models from the newest B200 to the oldest A100. The article also cautions that these indices are not perfect proxies for actual token spending and GPU prices, but they are not noise. At least from the rent curves, FUD and reality don't line up.

The depreciation comparison chart does the accounting in more detail. A100 and H100 do indeed spend more time below the straight-line depreciation line over longer periods, but the post-April surge changed the picture. For people running inference services, this is very practical: the traffic shift between A100 and H100 may lead to completely opposite cost conclusions depending on where monthly rents fall. A commenter under the a16z article suggested rerunning the cost per million tokens every quarter.

The demand side also offers an explanation. OpenRouter data shows token consumption shooting up steeply starting in spring, briefly flattening late summer, and recently jumping again. Token consumption for both open-source and closed-source models is growing exponentially, and closed-source models even regained a bit of share last week. OpenRouter covers only a small slice of the market, but it at least explains why the April turning point matters.

So the GPU depreciation table now looks awkward: it assumes a smooth downward line, while the rental market is voting with cash flow. Old cards have not become scrap metal; inference demand has pulled them back to the table. Some joke that their 3090 is paying the rent, while others look at neocloud stock prices and ask why, with rents rising like this, the relevant companies are still falling. The market is debating whether this is a short-term squeeze or a long-term trend, and supply concerns after 2028 remain.

Depreciation is a straight line; demand is not. Before writing an obituary for GPUs, look at the rent curve.