There is no single H100 price history. For NVIDIA's H100 GPU, dated public sources record providers' list prices, daily rental indexes and surveyed contract prices, and each moved by its own amount. On 10 November 2026 an H100 read $3.06 per GPU-hour on Silicon Data's daily index of GPU-focused clouds, or neoclouds [22], and $3.10 on Ornn's index of executed rentals [34]. AWS's on-demand list price, viewed on 12 November, came to $6.88 and Azure's to $12.29, each an 8-GPU rate divided by eight for this report [4, 5, 10, 11].
SemiAnalysis's contract series fell from 2023 to October 2025 and Silicon Data's daily index through 2025, and both have risen since. SemiAnalysis's surveyed one-year contract price went from $2.70 to $3.40 per GPU-hour in the first half of 2023 to what it calls "a low of $1.70/hr/GPU in October 2025", then to $2.35 by March 2026 [31, 32]. Silicon Data's H100 index rose from $2.00 on 9 December 2025 to $3.06, or 53.0% (computed for this report) [22, 25]. List prices kept their own schedules, so this article follows each kind of price on its own.
What an H100 price can mean
NVIDIA's product page lists two variants, the 80GB H100 SXM and the 94GB H100 NVL on a PCIe card [3]. AWS, Azure and Google Cloud, the hyperscalers, sell 8-GPU H100 instances by the hour [4, 5, 10, 11, 12]. Neoclouds such as Lambda and Nebius post a price per GPU-hour [13, 19]. An 8-GPU price divided by eight still includes the host's processors, memory and storage [5].
The indexes count different prices. Silicon Data's SDH100RT is a "daily benchmark for NVIDIA H100 hourly rental pricing, with separate neo-cloud and hyperscaler readings" [22]. Ornn's index comes "from executed trades rather than listings or surveys" [34]. SemiAnalysis builds its one-year index from "direct survey data across a pool of 100+ market participants" [32]. The report on GPU rental price indexes compared sets out what each publisher counts and how each governs its index.
H100 price history at a glance
| Date | Price per GPU-hour | Kind of price | Source |
|---|---|---|---|
| First half 2023 | $2.70 to $3.40 | One-year contract survey | [31] |
| 23 Aug 2023 | $12.29, AWS p5.48xlarge (computed) | Hyperscaler list price | [5] |
| 9 Sep 2023 | $2.59, Lambda 8x H100 SXM | Neocloud list price | [13] |
| 1 Jan 2024 | $3.49, Lambda 8x H100 SXM (computed) | Neocloud list price | [13] |
| By 15 Aug 2024 | $2.99, Lambda 8x H100 SXM | Neocloud list price | [13] |
| 27 May 2025 | $2.37 average | Daily rental index | [23] |
| 1 Jun 2025 | AWS P5 on demand cut 44% | Hyperscaler list price | [7] |
| Oct 2025 | $1.45 to $1.95 | One-year contract survey | [31] |
| 9 Dec 2025 | $2.00 | Daily rental index | [25] |
| Mar 2026 | $2.35 | One-year contract survey | [32] |
| By 1 Apr 2026 | $3.99, Lambda 8x H100 SXM | Neocloud list price | [14] |
| 1 Oct 2026 | $4.50, Nebius HGX H100 | Neocloud list price | [19] |
| 10 Nov 2026 | $3.06 | Daily rental index | [22] |
| 16 Nov 2026 | $4.59, Lambda, posted to take effect | Neocloud list price | [15] |
2023: launch-year prices
AWS made its P5 instances available on 26 July 2023 [4]. Its price file of 23 August 2023 listed the p5.48xlarge at $98.32 an hour, or $12.29 per GPU-hour (computed) [5]. Microsoft announced general availability of its ND H100 v5 virtual machines on 7 August 2023 [10], and the current pay-as-you-go rate for the 8-GPU size, also $98.32 an hour, carries a start date of 1 December 2023 [11].
Neocloud rates sat far lower. Lambda priced one H100 PCIe at $1.99 an hour by 3 July 2023 and 8x H100 SXM at $2.59 per GPU-hour by 9 September, with three-year reserved clusters at $1.89 [13]. Eugene Cheah, who co-founded an inference provider that rents GPUs, wrote in October 2024 that in the year before, H100s "were $8/hr if you could get them" [33]. That is one participant's account, not a measured series.
2024 and 2025: contract prices fall
Lambda raised its 8-GPU rate to $3.49 per GPU-hour (computed) on 1 January 2024, then cut it to $2.99 by 15 August 2024 [13]. SemiAnalysis's one-year range eased in every quarter of 2024, to $1.90 to $2.10 in the fourth [31]. AWS cut its Capacity Block reservation rate for P5 from $4.916 to $3.933 per accelerator-hour by March 2025 [8], then its P5 on-demand price by 44% from 1 June 2025 [7]. AWS wrote that it works to lower its costs "so that we can pass those savings back to our customers" [7]. Silicon Data's retrospective medians put the hyperscaler tier at $8.96 from January to May 2025 and $6.26 from July to December [21].
Prices kept falling into the autumn. Silicon Data's daily index dipped as low as 2.13 in mid-September [24], and the one-year contract range reached $1.45 to $1.95 in October 2025 [31]. The Q3 2025 compute market review follows that quarter, when posted H100 rents drifted lower while buyers signed multi-year capacity contracts.
Late 2025 to 2026: the turn
On 10 November 2025, CoreWeave said the customer on its first 10,000-plus H100 contract to near expiration had re-contracted two quarters early, "at a price within 5% of the original agreement" [18]. Silicon Data's H100 index then rose 10% in four weeks, from $2.00 on 9 December 2025 to $2.20 on 6 January 2026 [25]. SemiAnalysis found one-year contracts above $2 by late January and at $2.35 by March, and wrote that "On-Demand GPU rental capacity is sold out across all GPU types" [32]. It also wrote that "H100s are getting renewed at the exact same rate they were signed at 2-3 years ago" [32]. The Q1 2026 compute market review covers that quarter, when rental capacity sold out across GPU generations.
Silicon Data's hyperscaler index held still while its neocloud index climbed. The hyperscaler reading stayed between $7.40 and $7.55 through March, April and May 2026, and the neocloud reading rose 11.7% from a trough of $2.47 on 8 April to $2.76 on 30 May [26, 27]. By 10 November 2026 the neocloud reading was $3.06 [22].
2026 list prices by provider
Neocloud list prices rose in steps. Lambda's 8-GPU H100 rate went from $2.99 to $3.44 per GPU-hour by 10 March 2026 and to $3.99 by 1 April [14]. Its instances page posts $4.59 to take effect on 16 November 2026, 53.5% above February's $2.99 (computed) [14, 15]. Nebius went from $2.95 to $3.85 by 12 July and to $4.50 effective 1 October 2026 [19]. Nebius told shareholders in August 2026 of "more than 30% higher pricing on older-generation GPUs versus Q1" [20].
AWS raised its Capacity Block rate for P5 three times in 2026, to $5.970 per accelerator-hour, and its page names January 2027 for the next update [8]. Azure's and Google Cloud's on-demand rates showed no change: Azure's rates for its 8-GPU H100 virtual machine still carry their 1 December 2023 start date [11], and Google Cloud's a3-highgpu-8g read $88.49 an hour in every capture from July 2024 and on its live page [12]. CoreWeave showed the same $49.24 hourly rate for its 8-GPU HGX H100 instance in January 2025 and November 2026 [16].
The gap between tiers narrowed. Silicon Data's regional data put a hyperscaler H100 at 3 to 4.5 times the non-hyperscaler rate at the start of 2025 and roughly 2 to 2.5 times in the third quarter of 2026 [28]. On list prices alone, AWS's P5 on-demand rate was 4.7 times Lambda's in 2023 and is 1.5 times Lambda's posted 16 November rate (computed for this report) [5, 13, 15].
The Q3 2026 compute market review sets 2026's H100 moves beside a quarter when newer GPUs rented for more and compute futures went under review by the Commodity Futures Trading Commission (CFTC). The review period for NYMEX's proposed futures on Silicon Data's H100 index runs until 8 January 2027, per CME Group's letter of 4 November 2026 [36].
What a posted price history leaves out
Committed contracts made up over 98% of CoreWeave's 2025 revenue [17], and SemiAnalysis writes that "most of the GPU rental market is transacted on a long-term basis with contracts of at least 6mths and longer" [32]. Bandi and Su write that links between futures and current spot prices "fail due to the non-storable nature of compute" [35]. An unused GPU-hour cannot be stored for later sale, so today's rent and a later one are not tied by inventory.
The Rillor Compute Index is a benchmark for what superintelligence (SI) compute costs, to rent and to own, built from the prices cloud providers publish and the prices of GPU systems, under a published method.
Read the full report
The full report gives each dated price with the capture it comes from, the per GPU-hour arithmetic, the series side by side on 10 to 12 November 2026 and the places where the sources disagree.