a16z says Nvidia GPU rental rates are up since January, with the H200 about 160 percent higher

Old GPUs are renting for more than they did in January

Renting an NVIDIA GPU costs more now than it did at the start of the year, and that applies even to chips that are several generations old. Venture firm Andreessen Horowitz said in a post on X on 23 October 2026 that rental rates are up across the board since 1 January, with the H200 about 160 percent higher, the B200 about 120 percent higher, the RTX PRO 6000 about 65 percent higher and the H100 about 35 percent higher.

The post summarised a chart from a16z's weekly Charts of the Week newsletter, written by Moses Sternstein, which uses rental price data from Ornn Data. The newsletter says every model in the series, including the oldest A100s, costs more to rent than it did at the beginning of the year. The post had been viewed about 179,000 times by the afternoon of 24 October.

Hardware is supposed to lose value from the day it ships. If the rental price of a 2020-era chip is rising, it means demand for compute is outrunning supply at almost every price point, not only at the top end.

The numbers a16z published

The figures in a16z's post are approximate year-to-date changes. The firm called them a sign that GPUs are aging like fine wine, and added that even the six-year-old A100 rents for more than it did in January, without giving a percentage for it.

The ordering is telling. The biggest jump is in the H200, the memory-upgraded Hopper chip that sits between the H100 and NVIDIA's newer Blackwell parts. The B200, a current Blackwell chip, is next. The RTX PRO 6000, a workstation-class card that many smaller clouds rent for inference, is up by around two thirds. The H100, the workhorse of the 2023 and 2024 build-out, has risen the least of the four named chips but is still well above its January level.

Coverage has not been fully consistent. Wccftech reported different percentages, including 143 percent for the H200, and appears to have mixed the year-to-date chart with a second chart in the newsletter about depreciation. AI Lookout uses the year-to-date figures a16z itself posted.

Beating the depreciation curve

The second chart compares current rental prices with where they would be if each chip lost value in a straight line over a six-year life, a common accounting assumption for data centre servers. On that basis, the newsletter says, prices are anywhere from 9 percent to 143 percent higher than steady depreciation would predict, for every model.

Sternstein is careful about the longer history. He writes that the older A100s and H100s spent more time below the depreciation line than above it, so over their full life their pricing has underperformed straight-line depreciation. The surplus is a recent phenomenon, not a permanent feature.

That distinction matters for the debate over AI infrastructure accounting. Critics of hyperscaler spending have argued that six-year useful lives flatter earnings because GPUs become obsolete faster. Rising rents for old chips cut against that argument for now, but the a16z data also shows those same chips went through a long stretch when they were worth less than the accounting implied.

Why April changed the picture

According to the newsletter, April 2026 is when things really took off. It ties the jump to token demand measured by OpenRouter, a service that routes requests to many AI models and so captures only a slice of the market. Token consumption began a steep climb in the spring, plateaued briefly at the end of summer and then leapt again in the latest reading, for both open and proprietary models.

The timing lines up with what cloud providers have said publicly. This week Amazon chief executive Andy Jassy said AWS capacity is sold out for 2026 and largely for 2027, as covered in our report on Jassy's sold-out comments. When the largest cloud says its order book is full, buyers spill over to smaller GPU clouds and to older chips, which pushes their prices up.

Supply is growing, but slowly. NVIDIA has been signing suppliers to expand packaging and substrates, as in its new deals with Samsung Electro-Mechanics and others, and new capacity is being licensed in places like Kazakhstan, where Firebird won a US export licence for 45,000 GPUs. Those chips take months or years to turn into rentable capacity.

What Silicon Data's indices show

A separate price series points the same way. Silicon Data, which publishes daily GPU rental indices, listed its H200 index at $3.49 per GPU-hour on 23 October 2026, up 5.8 percent in a week. Its index overview showed the H100 at $2.88, up 9.1 percent in a week, the A100 at $1.56, the B200 at $6.45 and the B300 at $7.53.

Silicon Data's H200 history only starts on 4 May 2026, so it cannot confirm a16z's January baseline. But its August review of the series put the H200 at $2.70 per GPU-hour at the start of May and the H100 at $2.57 on 4 May. Against those readings, the October figures imply rises of roughly 29 percent for the H200 and 12 percent for the H100 since early May, on top of whatever happened between January and May.

The two data sets use different methods and samples, so the percentages are not directly comparable. What they agree on is direction: rental prices for both current and older NVIDIA chips have been climbing through the second half of 2026.

The caveats

Sternstein acknowledges that the indices are not perfect proxies for what companies actually pay for compute or tokens. Large buyers sign long-term contracts at negotiated prices that do not show up in public rental quotes, and index levels depend on which providers report data. a16z also did not publish the underlying Ornn Data series or the exact start and end values behind its percentages.

It is also worth remembering who is saying it. a16z is a large investor in AI companies and infrastructure, and its newsletter framed the data as a response to what it called recent fear, uncertainty and doubt about AI spending. The numbers can be checked against independent series like Silicon Data's, but the interpretation comes from a firm with a stake in the outcome.

Why it matters for AI buyers and lenders

For companies renting compute, rising prices for older chips remove an easy escape route. Teams that planned to save money by running inference on H100s or A100s while frontier labs fought over Blackwell are now paying more for that fallback too. Budgets set in January are likely to be short.

For lenders and investors who finance GPU fleets, the data is good news in the short term, because collateral is earning more than expected. The risk is the reverse of the past: if demand cools or a wave of new Blackwell and Rubin capacity arrives at once, rents for older chips could fall back below the depreciation line, as they did earlier in their lives.

What to watch

Daily index readings from providers like Silicon Data will show whether the autumn jump holds, and hyperscaler earnings in the coming weeks will show how much new capacity is landing. AI Lookout will keep tracking compute prices as part of its coverage of the AI infrastructure squeeze.