From Ritu Raj | Product & Market Analysis

H100 Rental Price Index: $3.00 an Hour, and Three Sources Disagree on the Path

On this page

Silicon Data's neocloud index read $3.00 per GPU-hour on 8 November 2026, up 50% from $2.00 on 9 December 2025. Ornn's trade-based index printed $3.37 the same day. Those two H100 rental price readings describe the same chip on the same date and sit 12% apart, which matters more than either number alone.

Key takeaways

  • The H100 rental price has reversed, not collapsed. Silicon Data's neocloud index fell to $2.00 in December 2025 and reached $3.00 by 8 November 2026, a 50% rise in 11 months.
  • Two daily indices gave very different months. From 10 October to 8 November 2026, Silicon Data rose 6.4% and Ornn rose 27.2%, and Ornn's gap to Silicon Data ranged from minus $0.17 to plus $0.57.
  • Hyperscaler list prices held far above a straight-line book value. About $7.10 now is 92% of the $7.76 late-2023 median, while a 6-year schedule would have 49% of cost remaining.
  • A price rise in a shortage does not prove a 6-year life. Each $1.00 per hour is worth $8,760 per GPU-year at full utilization, and that revenue can vanish faster than a depreciation schedule can adjust.
$3.00Silicon Data neocloud H100 index, 8 November 2026. Source: Silicon Data.
$3.37Ornn trade-based H100 SXM index, same day. Source: Ornn.
$7.10Hyperscaler on-demand H100, or 92% of the late-2023 median. Directional, two series. Source: Silicon Data.

The H100 rental price is about $3.00 to $3.37 per GPU-hour on the neocloud market as of 8 November 2026, according to Silicon Data and Ornn. Hyperscaler on-demand runs near $7.10. The last public one-year contract reading, from SemiAnalysis in March 2026, was $2.35.

Last verified 9 November 2026. Every index reading below was retrieved from the publisher's own page on that date, except where the text says otherwise.

The H100 rental price in November 2026 depends on the index

There is no single H100 rental price, and any article quoting one has made a choice for you. A neocloud is a GPU specialist such as CoreWeave, Nebius or Lambda. A hyperscaler is a general cloud such as AWS, and the two tiers price the same chip very differently.

Three publishers now produce a recurring series you can cite. Silicon Data launched the first daily H100 rental index in May 2025. Ornn's Compute Price Index reached the Bloomberg Terminal in April 2026. SemiAnalysis runs a monthly one-year contract index behind a subscription.

Direction is the one thing they share. All three show the H100 getting more expensive through 2026, which is the opposite of what a three-year-old accelerator is supposed to do. The size of the rise is where they part ways, and I would not quote any one of them alone.

How the three H100 rental price indices are built

The three indices differ in what enters the calculation, so they answer different questions. Read the construction before you read the number.

Three public H100 rental price series compared, as retrieved 9 November 2026
FeatureSilicon Data SDH100RTOrnn OCPI-H100SemiAnalysis index
InputNeocloud and marketplace observations, standardized for spec, term and regionTransacted H100 SXM rental pricesSurvey of 100+ market participants, checked against transactions
AggregationWeighted average across providers, weights not statedVolume-weighted and winsorized25th to 75th percentile range
ContractBenchmark-equivalent rentalSpot-style rental trades1-year contract
FrequencyDailySettles once per trading dayMonthly
Latest public reading$3.00, 8 November 2026$3.37, 8 November 2026$2.35, March 2026
Where it losesBlended basket hides variant mixDepends on the trades that print that dayNewer values are paywalled

Silicon Data: the broad neocloud basket

Silicon Data converts each rental record to a per-GPU-hour rate, filters errors and maps it to a benchmark contract. It then aggregates by provider into a weighted average across providers. The launch release cited 3.5 million pricing data points. The weights themselves are not on the page I read, and neither are the denominators behind its coverage claims.

Ornn: printed trades on a single variant

Ornn describes OCPI as a volume-weighted, winsorized average of transacted H100 SXM rentals, explicitly excluding scraped listings, surveys and rate cards. That is the cleanest definition of a price here. It also means the index moves whenever the day's trade mix moves, which the next section shows.

ICE plans an OCPI H100 future under the symbol HPR. As of a TradeInformer report on 30 September 2026, ICE had planned to release hypothetical settlement prices, and listing remained subject to a CFTC comment period. I found no verified first trade.

SemiAnalysis: the contract market

SemiAnalysis surveys more than 100 providers, buyers and sellers monthly. Its one-year H100 contract price rose from a low of $1.70 in October 2025 to $2.35 in March 2026, a 38.2% move. This is the series most relevant to anyone signing a term deal, and it is the one whose recent values I could not read.

The same chip showed a 6.4% rise and a 27.2% rise in 30 days

Silicon Data and Ornn both publish 30 consecutive daily values for 10 October to 8 November 2026, which allows a direct comparison. Silicon Data went from $2.82 to $3.00, up 6.4%. Ornn went from $2.65 to $3.37, up 27.2%.

H100 rental price, two daily indices, same 30 days $ per GPU-hour. Dark line is Silicon Data SDH100RT. Red line is Ornn OCPI-H100 (SXM). $2.50 $2.75 $3.00 $3.25 $3.50 10 Oct 20 Oct 31 Oct 8 Nov $3.53 Source: Silicon Data and Ornn daily series, retrieved 9 November 2026. Chart: Zan Digital.
Look at where the red line starts. Ornn sat below Silicon Data on 5 of the 30 days, so the gap is not a fixed premium you can subtract.

The daily spread is the more useful finding. Across the 30 days, Ornn averaged $0.20 above Silicon Data, with the gap running from minus $0.17 on 10 October to plus $0.57 on 3 November. The largest one-day move was $0.05 for Silicon Data and $0.39 for Ornn.

My read is that this is a construction difference, not an error in either index. A blended, standardized basket should be smooth. An average of whatever traded that day should not. Using one to check the other is sound, and treating them as interchangeable is not.

H100 rental prices since 2024: a long slide, then a reversal

H100 rental prices fell through 2025 and rose through 2026, and only Silicon Data's retrospective reaches back to 2024. The two tiers tell related but separate stories.

Hyperscaler on-demand fell by a third

Silicon Data's December 2025 retrospective gives hyperscaler medians of $7.76 for August to December 2023, $9.34 for June to December 2024 and $6.26 for July to December 2025. That is a 33.0% decline from the 2024 median to the late-2025 median. The same retrospective attributes a step down to an AWS H100 price reduction of roughly 30% in June 2025.

The neocloud index bottomed at $2.00

The dedicated index ran between 2.22 and 2.26 in August 2025 and hit 2.13 in mid-September. It reached $2.00 on 9 December 2025, then $2.20 by 6 January 2026. From there it climbed to $2.57 on 4 May, $2.79 on 5 June, $2.75 on 27 July and $2.81 on 15 October, before reaching $3.00.

From a $2.00 floor to $3.00 in 11 months Line: Silicon Data neocloud index, dated points. Red dots: SemiAnalysis 1-year contract, a different market. $1.50 $2.00 $2.50 $3.00 $2.00 $3.00 $1.70 $2.35 Sep 25 Jan 26 May 26 Sep 26 Source: Silicon Data and SemiAnalysis. Late-June point placed at 28 June. Chart: Zan Digital.
The contract series and the spot-style index both turned within two months of each other in late 2025, despite different inputs. That agreement is the strongest evidence that the reversal is real.

I think the popular 2025 story, that older GPUs were sliding toward worthlessness, was right on direction and wrong on what came next. The cause is reported consistently. SemiAnalysis attributes the rise to rental capacity selling out, and Seoul Economic Daily reported in September 2026 that VESSL AI raised its H100 rate 24.7%, from $2.39 to $2.98.

GPU rental rate cards spread from $1.99 to $6.16 for one chip

Published rate cards in November 2026 show how far apart the list prices remain. These are provider pages read on 9 November 2026. Lambda's instance prices take effect on 16 November.

H100 published prices per GPU-hour on provider rate cards, retrieved 9 November 2026
ProviderConfigurationPrice per GPU-hour
RunpodH100 PCIe, Community and Secure Cloud$1.99 and $2.89
RunpodH100 SXM, Community and Secure Cloud$2.69 and $3.99
LambdaH100 PCIe instance$3.79
LambdaH100 SXM and HGX instances$4.59 to $4.89
NebiusHGX H100, on-demand$4.50
AWSCapacity Blocks, p5.48xlarge, prepaid$5.97
CoreWeaveHGX H100 8-GPU node, on-demand, divided by 8$6.16

Rate cards are posted prices, not transaction prices. Runpod shows an update date of 4 November 2026. AWS states its Capacity Blocks rates are next updated in January 2027. CoreWeave's on-demand per-GPU figure is $49.24 per node-hour divided by 8.

Ornn's page draws the same line. It lists posted rates of $3.49 to $6.16 across four clouds, read by hand on 6 October 2026, while its traded index stood at $2.65 four days later. On that evidence, a posted rate is an asking price and a benchmark is closer to what buyers paid.

AWS provides one dated before-and-after from its own page. A worked example dated 20 March 2025 prices a p5.48xlarge Capacity Block at $31.464 per instance-hour, against $47.757 today. That is a 51.8% rise, though the older figure comes from an example rather than a rate table.

The premium is compressing too. Silicon Data put hyperscaler H100 at roughly 3 times the neocloud tier in April 2026. At $7.10 against $3.00, the multiple is now 2.37. If you are buying at the hyperscaler tier for the convenience, the rising neocloud price has narrowed what that convenience costs you. A B200-class comparison sits in the MI355X and B200 rental price comparison.

What the 10-Ks assume about server useful life

Most of the companies buying these chips now book servers over 5 to 6 years, a span that H100 rental pricing cannot confirm on its own. Here is what each latest annual report states.

Stated useful life for servers or technology equipment, latest annual filings
CompanyStated lifeNoted change or effect
Alphabet (FY2025)Generally 6 years, servers and network equipmentNone in this filing
Oracle (FY2026)1 to 6 years, mostly servers and networking at 6None in the pages read
CoreWeave (FY2025)6 years, technology equipmentMoved from 5 to 6 years, effective January 2023
Microsoft (FY2026)2 to 6 years, servers and network equipment4 to 6 years in fiscal 2023
Meta (FY2025)5 to 5.5 years, servers and network assetsMost moved to 5.5 years in January 2025, cutting depreciation $2.92 billion
Amazon (FY2025)5 to 6 years, servers and networkingSubset cut from 6 to 5 years in January 2025, adding $1.4 billion of expense

The profit sensitivity of those choices is already computed in the filing-by-filing server useful life analysis, and the argument over the overstatement figure is in the review of the GPU depreciation debate. This post asks a narrower question. Given what the market actually paid for the H100, does the price path contradict those assumptions?

One disclosure limit shapes the answer. None of these filings gives a separate life for a GPU, except that CoreWeave's category is built around computing equipment in data centres. Each figure applies to a class of assets that includes older CPU servers.

Testing the H100 price path against straight-line book value

The H100 price path does not contradict a 6-year schedule, and it does not validate one either. Straight-line depreciation removes the same share of cost every year, so the remaining book value at any age is simple arithmetic.

List prices have held above every book line

Take a chip deployed in October 2023, the midpoint of Silicon Data's first hyperscaler period. By October 2025, the on-demand median had fallen to 80.7% of its starting level, against book values of 66.7% on 6 years, 60.0% on 5 and 50.0% on 4. By November 2026, list price stood at 91.5% of the starting median, with book values of 48.6%, 38.3% and 22.9%.

H100 list price retained, against straight-line book value October 2023 = 100. Solid line to October 2025 is one series; dashed section joins a different Silicon Data series. 0 25 50 75 100 125 91.5 6 years: 48.6 5 years: 38.3 4 years: 22.9 Oct 23 Sep 24 Oct 25 Nov 26 Illustrative comparison. List price is not realised revenue. Source: Silicon Data, Zan Digital arithmetic.
Illustrative, not measured book value. Read the red line against the lines beneath it: even the October 2025 low never crossed a 4, 5 or 6 year schedule.

What a rental price has to cover

A second test needs no price history at all. Divide a GPU's installed cost by its useful life and by 8,760 hours, and you have the depreciation charge per hour. Every $10,000 of installed cost is $0.285 per hour over 4 years, $0.228 over 5 and $0.190 over 6, at full utilization.

Installed cost per GPU that rental revenue covers, depreciation only, at 100% utilization
Rental priceAnnual revenue per GPU4-year life5-year life6-year life
$1.70 (contract low, October 2025)$14,892$59,568$74,460$89,352
$2.00 (index low, December 2025)$17,520$70,080$87,600$105,120
$3.00 (Silicon Data, November 2026)$26,280$105,120$131,400$157,680
$3.37 (Ornn, November 2026)$29,521$118,085$147,606$177,127

Zan Digital arithmetic: price times 8,760 hours times useful life. Excludes power, networking, staff, financing and downtime, so real ceilings are much lower. At 50% utilization, halve every cell.

The table shows why rental price matters more than the accounting debate. A $1.00 move in the hourly price is worth $8,760 per GPU-year at full utilization. Moving a life from 6 years to 4 costs one-twelfth of installed cost per year. The two effects are equal only at an installed cost of $105,120 per GPU, so at any lower cost the rental price is the larger swing.

Utilization and financing decide who actually survives that swing. The owner-versus-renter arithmetic is in the GPU utilization break-even model. For a borrower, the timing of debt matters as much as the price, as the CoreWeave debt maturity ladder sets out.

Where this argument is weakest

This post leans on series that cannot be fully audited from the outside, and three problems deserve to be stated plainly.

A shortage is not a useful life

Rising rents for an old chip show that demand exceeded supply this year. They do not show the chip will earn anything in year six. SemiAnalysis reported H100 contracts renewing at their original rates, some running through 2028, which is real evidence of extended earning life. It is also evidence from a period when buyers had few alternatives.

There is survivorship in this too. The chips still listed are the ones still working. Neither index reports failure rates, downtime or the share of the installed base that has already been retired, and those facts matter more to a schedule than a price does.

The Silicon Data series do not splice

I had to join two different Silicon Data products to run the book-value test. The retrospective puts the late-2025 hyperscaler median at $6.26, and its hyperscaler index opened at $7.50 on 1 March 2026. I cannot tell you how much of that $1.24 step is price and how much is construction, which is why the chart breaks the line there.

The retrospective's neocloud column is the bigger warning. It shows medians of $2.99 for late 2024 and $3.33 for late 2025, while the dedicated index sat near $2.00 to $2.26 in the second half of 2025. I have not spliced those series anywhere in this post, and you should not either.

Every publisher here has a conflict or a gap

Silicon Data's own page carried a stale-looking FAQ answer of $2.53 next to a $3.00 headline. Ornn's page states 18.2% over thirty days from $2.85 while its own daily table starts at $2.65. Seoul Economic Daily reported an Ornn figure of $3.28 in September, above the $3.15 September high in Ornn's current table, and I could not reconcile the two.

Both companies also sell into the futures market their indices would settle. Silicon Data's site advertises a partnership with CME on compute futures. Ornn's H100 index is the planned reference for the ICE contract. That does not make their data wrong. It does mean that neither publisher is a disinterested observer, and the SemiAnalysis index is the only one of the three with no visible contract tied to it.

Frequently asked questions

What is the H100 rental price per hour right now?

On 8 November 2026, Silicon Data's neocloud index read $3.00 per GPU-hour and Ornn's trade-based index read $3.37. Hyperscaler on-demand sat near $7.10 on Silicon Data's separate series. Published rate cards run from $1.99 for a community PCIe unit on Runpod to $6.16 for a CoreWeave on-demand HGX node, so the answer depends on the tier you buy.

Why do H100 rental price indices disagree?

They measure different things. Silicon Data blends neocloud and marketplace observations across providers and standardizes them. Ornn uses a volume-weighted, winsorized average of transacted H100 SXM rentals. SemiAnalysis surveys more than 100 market participants on one-year contracts. Over the 30 days to 8 November 2026, Ornn stood $0.20 above Silicon Data on average, and the daily gap ranged from minus $0.17 to plus $0.57.

Did H100 rental prices fall or rise since 2024?

Both, in sequence. Silicon Data's retrospective shows hyperscaler on-demand medians falling from $9.34 in the second half of 2024 to $6.26 in the second half of 2025. Its neocloud index bottomed at $2.00 on 9 December 2025, then rose 50% to $3.00 by 8 November 2026. SemiAnalysis recorded one-year contracts rising from $1.70 in October 2025 to $2.35 in March 2026.

Do rising H100 rental prices justify six-year GPU depreciation?

Not on their own. Alphabet, Oracle and CoreWeave state six years for servers or technology equipment, and the H100 has so far held price better than a straight line would imply. But a shortage can lift older chips for a while, and rental prices say nothing about utilization, power or what happens when newer supply arrives. Treat the evidence as supportive and incomplete, not as proof.

What is the difference between neocloud and hyperscaler H100 pricing?

A neocloud is a GPU specialist such as CoreWeave, Nebius or Lambda, while a hyperscaler is a general cloud such as AWS, Azure or Google Cloud. In early November 2026, Silicon Data's neocloud index read $3.00 against about $7.10 for hyperscaler on-demand, a multiple of 2.37. Silicon Data describes the hyperscaler rate as a posted ceiling above capacity that is largely already reserved.

Where to start

Replace the single H100 price in your model with a band this week. Use Ornn's range since 10 August 2026 as the working band, which is a low of $2.46, a median of $2.90 and a high of $3.53. If your plan works only at the top of that band, you are underwriting a shortage rather than a business.

Then read both public indices on the same day each Monday and record the gap. A gap that stays positive and widens tells you trades are clearing above the broad basket. A gap that closes while both fall is the signal I would act on first, because it would mean the reversal has started to unwind.

References

  1. Silicon Data, H100 Rental Price Index, daily series to 8 November 2026, and Silicon Index methodology summary. Retrieved 9 November 2026. Used for SDH100RT and hyperscaler readings.
  2. Silicon Data, H100 Rental Price Over Time, 21 December 2025. Used for 2023 to 2025 medians and the June 2025 AWS reduction. The AWS percentage is Silicon Data's statement, not verified on AWS's own pages.
  3. Silicon Data blog: September 2025 update (20 October 2025), H100 price spike (9 January 2026), hyperscaler index (27 April 2026), H200 vs H100 (4 August 2026). Used for dated index levels.
  4. Ornn, NVIDIA H100 SXM Price Index, settled 8 November 2026. Retrieved 9 November 2026. Used for daily OCPI values, monthly ranges and methodology.
  5. SemiAnalysis, The Great GPU Shortage, 2 April 2026. Used for the one-year contract series, methodology and renewals. Later index values are behind a subscription.
  6. Press: TradeInformer, ICE to release GPU futures reference data, 30 September 2026. Seoul Economic Daily, H100 rental prices jump 22%, 22 September 2026, AI-translated from Korean. Secondary coverage; upgrade to ICE notices where possible.
  7. Annual filings: Alphabet FY2025 10-K; Oracle FY2026 10-K; CoreWeave FY2025 10-K; Meta FY2025 10-K; Amazon FY2025 10-K; Microsoft FY2026 annual report. Retrieved 9 November 2026.
  8. Rate cards retrieved 9 November 2026: AWS Capacity Blocks; Runpod; Lambda; Nebius; CoreWeave.

The weakest thing about this source base is that all the price data comes from publishers with a commercial stake in it, and no index discloses the contributing trades or provider weights. Figures are dated and move daily, so re-verify any figure older than 90 days.

RR
Ritu Raj
Founding Member, Zan Digital. Writes about AI product economics, B2B software markets and what the numbers behind vendor claims actually say.

Related reading