Older NVIDIA GPUs Are Renting for More, but a16z's A100 Data Is Easy to Misread

Andreessen Horowitz (a16z), the major US venture capital firm, reported in a newsletter dated October 23, 2026, that rental rates for NVIDIA GPUs have risen since the start of the year. According to a16z's "Charts of the Week", rental rates for all five models tracked by Ornn Data are above their levels at the start of 2026, and that includes the A100, which debuted in 2020. Compared with a baseline that assumes GPU value declines evenly over six years, rates as of October 19 were 9% to 143% higher.

However, it would be premature to read these figures as evidence that GPU asset values have risen. The largest figure, +143%, is for the H200, and it does not mean the H200 has gained 143% since the start of the year. The comparison is against a baseline that assumes each GPU's first observed rental rate falls to zero over six years in a straight line. Working backward with the same method, the A100's +9% implies that its current rate has fallen to about 58% of its level when observation began. Assessing GPU value requires distinguishing the chart's comparison basis, the transactions the rental index covers, and the useful lives companies actually use in their accounting.

A year-to-date rise and the gap to a six-year depreciation line are different numbers

The two charts a16z published deal with the same GPU rental rates but use different comparison bases. The first chart is a seven-day moving average indexed to January 1, 2026 = 100, using Ornn Data through October 19. It covers five models: A100 SXM4, H100 SXM, H200, B200, and RTX PRO 6000 WS. a16z gives approximate launch years of 2020 for the A100, 2022 for the H100, 2024 for the H200 and B200, and 2025 for the RTX PRO 6000 WS.

Reading the chart, the index as of October 19 is roughly 263 for the H200, 222 for the B200, 165 for the RTX PRO 6000 WS, 136 for the H100, and 107 for the A100. In other words, H200 rental rates are about 2.6 times their level at the start of the year, while the A100 is up only about 7%. Prices have not risen steadily all year, though. The H100 and B200 indexes fell to about 77 in February and March before rebounding in April and May.

The second chart, titled "GPU Rental Rates Aren't Depreciating On Schedule," sets each model's first rental rate observed by Ornn at 100%. It compares actual rates with a baseline that assumes rates fall linearly to zero over six years, starting from the first complete seven-day average available. According to a16z, as of October 19 actual rates exceeded the baseline by 9% for the A100, 54% for the H100, 143% for the H200, 111% for the B200, and 73% for the RTX PRO 6000 WS.

{
  "type": "bar",
  "title": "GPU rental rates versus a six-year straight-line baseline (October 19, 2026)",
  "unit": "%",
  "categories": ["A100 SXM4", "H100 SXM", "H200", "B200", "RTX PRO 6000 WS"],
  "series": [
    { "name": "Gap to baseline", "values": [9, 54, 143, 111, 73] }
  ],
  "caption": "Gap versus a baseline that sets each model's first observed rental rate at 100% and assumes a straight-line decline to zero over six years. Formula: actual rate ÷ baseline value − 1. This is not a year-to-date price increase.",
  "source": "a16z Charts of the Week (October 23, 2026), Ornn Data"
}

The premium used in this chart is calculated as "actual rate ÷ baseline value − 1." For example, the H200's +143% means the October 19 rate was 2.43 times the baseline. Because the baseline falls over time, the premium grows steadily larger even if the actual rate stays flat. Conversely, for models with a short observation history, the baseline is close to 100%, so increases in the rental rate itself are strongly reflected in the premium.

Importantly, the five models are not compared over the same period. According to the chart's horizontal axis, observation began in early 2024 for the A100, mid-2024 for the H100, early 2025 for the H200, and autumn 2025 for the B200 and RTX PRO 6000 WS. The earlier observation began, the further the baseline has dropped, so the premiums cannot be compared directly across models as a ranking or as rates of asset-value appreciation.

In addition, a16z labels the dotted line in the legend "Six-year book value (% of original cost)." But the chart's explanation sets the first observed rental rate at 100%; it does not use actual acquisition cost. The dotted line is not an accounting book value but a baseline representing an assumed pace of decline in rental rates. Moreover, although a16z explains that GPUs depreciate from the time they launch, the chart does not start at launch. For the A100, rates between its 2020 launch and the start of observation in early 2024 are not part of the comparison.

Working backward, the A100's "+9%" implies actual rates have fallen about 42%

Looking closely at the A100 data makes clear the difference between being above the baseline and rising over the long term. In a16z's chart, A100 observation begins in early 2024. Assuming the first complete seven-day average was January 7, 2024, 1,016 days, or about 2.78 years, have elapsed through October 19, 2026.

Using a baseline that falls linearly from 100% to zero over six years, the value after 2.78 years is "1 − 2.78 ÷ 6," or about 0.536, roughly 54% of the level at the start of observation. According to a16z, the actual rate is 9% above this baseline, so multiplying 0.536 by 1.09 gives about 0.58. The A100's rental rate is about 58% of its level at the start of observation, which works out to a decline of roughly 42% over about two years and nine months.

That 58% figure, however, was not published directly by a16z or Ornn; it is a back-calculation using an observation start date estimated from the chart and the premium a16z reported. If the start date is off by a month, the result shifts by about 1.5 percentage points. Reading the chart line directly also puts the A100's latest value at about 55% to 58% of its starting level, nearly coincident with the baseline. Note, too, that what is being compared here is GPU rental rates, not used-GPU sale prices or accounting book values.

Compared with the start of 2026, A100 rental rates are up about 7%. But over the longer span since early 2024, they have fallen more than 40%, so "higher than at the start of the year," "higher than the six-year depreciation baseline," and "rising in value over the long term" are three different claims.

a16z itself acknowledges that A100 and H100 rates spent most of the observation period below the six-year baseline. It argues that long-term comparisons alone fail to capture market changes since spring 2026, and it emphasizes the expansion in GPU demand that began in April. The newsletter used OpenRouter data to show that token demand surged from spring, slowed temporarily in late summer, and then rose again. A100 rental rates also moved above the baseline from spring to early summer, dipped, and as of October 19 were once again slightly above it.

Still, the token usage OpenRouter can observe is only a portion of the overall market. a16z also cautions that this data does not accurately represent token spending or GPU prices as a whole. It is evidence suggesting expanding GPU demand, but it cannot by itself establish the supply-demand picture of the entire market or changes in asset values.

What does the GPU rental index OCPI measure?

The a16z charts cite "Ornn Data" as the source but do not name a specific index. Ornn publishes a GPU rental index called the Ornn Compute Price Index (OCPI), which offers series for the A100, H100, H200, B200, and RTX PRO 6000. The data a16z used appears to come from these series, but an explicit correspondence cannot be confirmed. To understand the nature of the index, therefore, it is necessary to refer to Ornn's published OCPI methodology.

OCPI is an index calculated from transaction prices actually realized when GPUs are rented. Ornn collects anonymized data from GPU providers around the world whose prices fluctuate with demand, and records the US-dollar rate for using one GPU for one hour among GPUs rented on demand. A distinguishing feature is that it covers prices actually paid, not list prices posted by providers or quotes for which no transaction took place.

However, OCPI covers only on-demand transactions; it does not include reserved contracts, long-term contracts, or forward contracts committing to future use. Ornn explains that such contracts are individually negotiated in terms of usage period, prepayment conditions, and discount rates, making them hard to compare as prices on the same terms. Price information on long-term contracts and the like is handled separately as forward curves.

The index is calculated using a volume-weighted winsorized mean of transaction prices collected over the most recent hour for each GPU type. Winsorizing is a technique that, rather than excluding extreme values outside a certain range, replaces them with the range's upper or lower limit before averaging. This limits the extent to which a handful of extreme prices can move the whole index. No smoothing beyond this treatment is applied to OCPI's hourly index calculation. That is separate from the seven-day moving average a16z used in its charts.

In the confirmed values Ornn published for October 25, 2026, the A100 SXM4 was $1.09 per GPU-hour, the H100 SXM $3.03, the H200 $6.30, and the B200 $9.70. The A100 stood at $1.00 as of October 19. At a reference rate of ¥158 to the dollar, the A100 on October 25 comes to about ¥172 per hour and the H100 to about ¥479. Using an A100 for 30 days, or 720 hours in total, would come to $784.80, about ¥124,000.

This is a simple estimate using index values, however, and does not mean one could actually contract at that rate. Nor does it reflect Japanese cloud pricing or individual contract terms. In fact, among four GPU clouds Ornn compared, published on-demand rates for the H100 SXM as of October 6 ranged from $3.49 to $6.16 per hour. Compared with the OCPI-H100 value of $3.03 on October 25, that is 15.2% to 103.3% higher. The dates being compared differ by 19 days, which warrants caution, but it shows that published prices and actual transaction prices can diverge significantly.

If the a16z charts do use OCPI series, therefore, what they show is the change in rental rates actually realized in the on-demand market. They do not measure long-term contract prices, used-GPU sale prices, or the accounting asset values held by owners. On-demand rental rates are an important indicator for thinking about GPU profitability, but they are insufficient on their own for judging residual value.

Amazon uses five years, Meta 5.5: GPU useful lives differ by company

a16z adopts six-year straight-line depreciation as its baseline for rental rates. But the useful lives used in actual corporate accounting are not uniform, and companies make different judgments about AI-related equipment.

In its 2025 Form 10-K (annual report), Amazon disclosed that from January 1, 2025, it shortened the estimated useful life of certain servers and networking equipment from six years to five. The reason it cited was the rapid pace of technological advance, particularly in AI and machine learning. The change increased 2025 depreciation expense by $1.4 billion. The company had extended the estimated useful life of servers from five years to six in 2024, so it revisited its judgment on some equipment the following year.

Meta, by contrast, explained in its 2025 Form 10-K that it extended the estimated useful life of most servers and network assets to 5.5 years from January 1, 2025. The change reduced 2025 depreciation expense by $2.92 billion. Alphabet changed the useful life of servers and certain networking equipment to six years in a January 2023 review. Microsoft also extended the useful life of servers and networking equipment from four years to six in July 2022, applying it from fiscal 2023.

In its 2023 annual report, Microsoft cited technological progress, along with improved operating efficiency of servers and network equipment through software investment, as reasons for extending useful lives. That contrasts with Amazon, which cited the pace of technological progress as the reason for shortening the life of some equipment. Each company makes its judgments based on how equipment is used and operated and on its outlook for future use, not solely on fluctuations in GPU rental rates.

These useful lives are also accounting estimates for asset groups, namely servers and networking equipment, and are not set individually for specific GPU generations such as the A100 or H100. They do not necessarily match operational decisions about how many years a company will actually use a GPU.

The main figures that have appeared so far are summarized below.

The top three items in the table all concern GPU on-demand rental rates, assuming OCPI series are being used. The bottom three are corporate accounting estimates, and both what they measure and how they are derived differ. Just as a year-to-date index and the gap to a six-year depreciation line cannot be compared directly, rates exceeding the baseline cannot by themselves settle whether a five-year or six-year useful life is appropriate.

Furthermore, evaluating the economic value of GPUs to their owners requires considering not only rental rates but also utilization, power costs, maintenance costs, acquisition cost, revenue from long-term contracts, and future resale prices. Even if on-demand rates rise, profitability does not necessarily improve if upkeep costs increase or utilization falls. a16z's charts are evidence of a shift in GPU demand, but they are not a direct test of whether asset depreciation is appropriate.

Ornn is an a16z portfolio company, and the index's transparency is in question ahead of GPU futures

In assessing this data, the relationship between index provider Ornn and a16z also matters. a16z crypto announced on June 24, 2026 that it led Ornn's $33 million seed round. The investment announcement pointed out that while compute resources such as GPUs and data centers are becoming important assets, reliable price benchmarks and means of hedging price volatility are lacking compared with markets such as oil and energy. It presented OCPI, which Ornn is developing, as a foundational benchmark for such a market.

But the October 23 a16z newsletter, while naming Ornn Data as the source of the charts, does not mention in its text that a16z crypto has invested in the company. The end of the article carries a general disclaimer about portfolio companies and a link to the list of investments. The existence of an investment relationship is not in itself a reason to dismiss the data's accuracy, but the fact that the data provider's business value and a market analysis built on its data are connected is important information for readers to weigh.

Ornn itself acknowledges conflicts of interest in its OCPI methodology. It explains that because Ornn, which calculates the index, and affiliates involved in trading and delivering GPU compute resources share common ownership, there is a structural conflict of interest between fair calculation of the index and the commercial interests of the affiliates. At the same time, it states that Ornn and its affiliates do not trade or make markets in financial products linked to OCPI, and that it is still in the process of putting formal procedures in place to manage conflicts of interest.

The transparency of such an index becomes all the more important with the creation of a futures market tied to GPU rental rates. Intercontinental Exchange (ICE) and Ornn announced on May 19, 2026 plans to introduce GPU compute futures settled against OCPI. Ornn co-founder and CEO Kush Bavaria explained that compute has grown into a trillion-dollar market but lacks the price discovery and risk-transfer mechanisms used in other major commodity markets.

According to a notice ICE Futures U.S. issued on September 29, the first products scheduled to list are two contracts: OCPI H100 futures (HPR) and OCPI B200 futures (BKL). Both are monthly and cash-settled, using the average of the index's daily prices over the contract month as the final settlement price. The daily price is calculated by averaging the hourly index values OCPI publishes.

However, the listing of the futures is conditioned on completion of the US Commodity Futures Trading Commission (CFTC) comment-solicitation process, and ICE has not fixed a listing date. ICE announced that from October 5 it would begin providing index data and hypothetical daily settlement prices to market participants, but this is for trading-system development and preparation and does not mean futures trading has started.

Once futures benchmarked to OCPI are listed, the index values will no longer be mere reference material for market analysis but the settlement basis that determines participants' profits and losses. At that point, questions will be raised about how well an index covering only on-demand transactions represents price movements across the whole market, whether the scope of transaction-data collection is sufficient, and whether mechanisms to prevent price manipulation and manage conflicts of interest are working.

What can be read with confidence from a16z's charts is that GPU rental rates tracked by Ornn Data have risen since the start of 2026 and are also above a baseline assuming six-year straight-line depreciation. But A100 rental rates have fallen to about 58% of their level when observation began, so the value of older GPUs has not risen consistently. Useful lives used in corporate accounting also differ, at five to six years, and on-demand rental rates alone cannot establish whether asset values are justified. Going forward, how far the rise in GPU rental rates continues, along with long-term contract prices, the used market, and the development of a futures market using OCPI, will be important for gauging the economic value of GPUs.