Set a neocloud vs hyperscaler GPU price side by side and the gap looks plain. On 7 November 2026, AWS posted its 8-GPU NVIDIA H100 instance in US East at $55.04 an hour, or $6.88 per GPU-hour, and Crusoe posted its 8-GPU H100 at $3.90 per GPU-hour [9, 33, 34]. The two prices describe different products sold on different terms. This explainer puts posted H100 prices from four hyperscalers and five neoclouds on one basis and shows what each price includes.
The basis is one product: an 8-GPU H100 node with 80 GB per GPU and NVLink, on-demand, Linux, in a U.S. region where the provider names one, posted on 7 November 2026 and divided by eight for U.S. dollars per GPU-hour [8]. On that basis the hyperscalers ranged from $6.88 to $12.29 per GPU-hour and the neoclouds from $3.90 to $6.16 (computed for this report from [9, 15, 18, 22, 25, 29, 30, 33, 35]). On multi-year commitment, hyperscaler prices fall inside that neocloud range. Priced as InfiniBand clusters, one neocloud's rates rise to near AWS's.
Neocloud vs hyperscaler: what each one sells
No regulator in the sources reviewed defines either word. The UK Competition and Markets Authority (CMA) found in July 2025 that Microsoft, AWS, Google, Oracle and IBM all supply accelerated compute alongside their standard compute services [4]. It named CoreWeave, Lambda Labs, Crusoe Cloud, Nebius and Nscale as smaller providers that specialise in accelerated compute [4]. Their offers tend to target a narrower group of customers, with fewer cluster management options, and they generally do not serve the wider cloud services market [4].
The companies draw the same line. CoreWeave's 10-K says it competes mainly with hyperscalers that offer general-purpose cloud computing as part of a broader portfolio, several of which are also its customers and partners [1]. NVIDIA's 10-K lists cloud service providers and neocloud builders as separate types of indirect customer [2].
This article follows the CMA's grouping and adds Together AI to the neoclouds by its own rule. AWS, Microsoft Azure, Google Cloud and Oracle Cloud Infrastructure are the hyperscalers, and CoreWeave, Lambda, Nebius, Crusoe and Together AI are the neoclouds. Oracle's place is a classification choice: the CMA reports Oracle's share as smaller than Google's, with a narrower range of services [3].
Posted H100 prices on one basis
| Provider and product | Segment | Posted price | Per GPU-hour | Region |
|---|---|---|---|---|
| AWS p5.48xlarge | Hyperscaler | $55.04 an hour [9] | $6.88 | US East (N. Virginia) |
| Azure ND96isr H100 v5 | Hyperscaler | $98.32 an hour [15] | $12.29 | East US |
| Google a3-highgpu-8g | Hyperscaler | $88.49 an hour [18] | $11.06 | Iowa (us-central1) |
| Oracle BM.GPU.H100.8 | Hyperscaler | $10 per GPU-hour [22, 23] | $10.00 | None in the record |
| CoreWeave HGX H100 | Neocloud | $49.24 an hour [25] | $6.155 | North America |
| Lambda 8x H100 | Neocloud | $3.99 on 25 October, $4.59 from 16 November [27, 29] | $3.99 or $4.59 | Not stated |
| Nebius HGX H100 | Neocloud | $4.50 per GPU-hour [30] | $4.50 | eu-north1 [32] |
| Crusoe H100 HGX | Neocloud | $3.90, $4.50 from 1 December [33] | $3.90 or $4.50 | U.S. and Iceland zones [34] |
| Together AI HGX H100 | Neocloud | $3.99 per GPU-hour [35] | $3.99 | Not stated |
Per GPU-hour figures are the posted price divided by eight, computed for this report. Lambda's $3.99 comes from a 25 October capture, because on 7 November its page showed only the prices taking effect on 16 November [27, 29]. Together AI's page names the HGX H100 without a node size [35].
The spread inside each segment is wide. AWS sat 12% above CoreWeave, while Azure was 3.2 times Crusoe (computed for this report). Nebius documents its H100 in a European region only, so its row fails the region test against the U.S. hyperscaler rows [32]. The conversion from each provider's units follows the method in Reading GPU prices from cloud pricing APIs.
What the hourly price includes
Per GPU, the nodes compared carry from 12 vCPUs at Azure to 28 at Oracle, and from 120 GB of host memory at Crusoe to 256 GiB at AWS and 256 GB at Oracle and CoreWeave [10, 17, 23, 26, 34]. Local storage runs from none bundled at Nebius, which bills local SSD at $0.065 per GiB a month, to 7.68 TB of NVMe per GPU at Oracle [23, 30]. CoreWeave's pricing page lists 61.44 TB of local storage per node, and its documentation gives 30.72 TB usable after RAID 10 [25, 26].
Adding AWS's 3.84 TB per GPU at Nebius's posted local SSD rate would cost about $0.32 per GPU-hour, about 7% of Nebius's on-demand price (computed for this report from [9, 30], 730 hours a month).
Egress differs as well. AWS gives 100 GB a month free across its services, then charges $0.09 per GB for the next 10 TB [9]. Google's default Premium Tier charges $0.12 per GiB up to 1 TiB, then $0.11 [21]. Oracle's first 10 TB a month are free, and Lambda, Nebius and Crusoe post no egress fees [24, 27, 30, 33]. Sending 10 TB to the internet from one AWS node running all month costs about $891, about $0.15 per GPU-hour or 2% of its on-demand price (computed for this report from [9]).
Commitments move hyperscaler prices into the neocloud range
Hyperscalers post many prices for one node. Microsoft's retail API lists three- and five-year reservations for its H100 VM that work out to $5.40 and $4.92 per GPU-hour, with the one-year term at $7.87 (computed for this report at 8,760 hours a year) [15]. Google posts a three-year committed use rate of $4.86, a Flex-start rate of $4.79 for runs of up to seven days and a calendar-mode rate of $5.20 [18, 20]. AWS posts Capacity Blocks in US East at $5.970 per GPU-hour, paid up front [11].
The multi-year, Flex-start, calendar-mode and Capacity Block rates all fall between $3.90 and $6.16, the neocloud on-demand range on the same day (computed for this report from [11, 15, 18, 25, 33]). A three-year rate beside an on-demand rate still compares two different offers. None of the five neoclouds posts a price for a term longer than one year [25, 28, 30, 33, 35]. CoreWeave offers up to 60% off on-demand for committed usage, and Lambda lists its "1 year+" cluster term with no price [25, 28]. Together AI posts reserved rates from $3.69 down to $3.19 per GPU-hour for terms of 3 to 180 days [35]. How neocloud contracts are written and financed is the subject of Neocloud economics: contract first, hardware second.
AWS's Spot rate in us-east-1 worked out to $2.66 per GPU-hour, Azure's to $2.27 and CoreWeave's to $2.46 (computed for this report from [12, 15, 25]). Google's Spot rate of $6.59 sat above its own Flex-start and three-year rates [18]. What interruptions cost a training run is covered in What research shows about training on spot GPUs.
Interconnect and cluster size
Lambda's InfiniBand-connected 1-Click Clusters, for two weeks to a year, posted $6.16 per GPU-hour at 16 GPUs, $5.85 at 64 and $5.54 at 256 [28]. That is 21% to 54% above its single 8-GPU instance, depending on cluster size and instance price, and close to CoreWeave's InfiniBand node at $6.155 and AWS's EFA node at $6.88 (computed for this report from [9, 10, 25, 26, 27, 28, 29]).
AWS lists 3,200 Gbps of EFA per instance with GPUDirect RDMA [10], Azure a 400 Gbps InfiniBand link per GPU [17], Oracle 8 x 2 x 200 Gbps of RDMA [23], Crusoe 3,200 Gbps of InfiniBand [34], Nebius InfiniBand at "400 Gbps (8× ConnectX-7)" [32] and Google up to 1,000 Gbps per VM with GPUDirect-TCPX [19]. No benchmark reviewed shows that these nodes deliver the same training or inference throughput.
Three neocloud price rises, October to December 2026
Nebius raised its H100 on-demand price from $3.85 to $4.50 on 1 October [30, 31]. Lambda posts $4.59 from 16 November, against $3.99 in a 25 October capture [27, 29]. Crusoe posts $4.50 from 1 December, against $3.90 on 7 November [33]. Each is a rise of 15% to 17% (computed for this report). AWS announced a 44% cut to P5 on-demand prices from 1 June 2025 [14], and says its current Capacity Block prices are scheduled to be updated next in January 2027 [11].
One index publisher reports that the hyperscaler premium in its H100 data narrowed from about 250% to about 120% between the first quarter of 2025 and the third quarter of 2026 [7]. It adds that its quarterly medians cannot separate list-price changes from shifts in the mix of contracts and configurations behind them [7].
What a like-for-like comparison holds constant
A fair comparison of neoclouds and hyperscalers fixes the GPU variant, since an H100 SXM and an H100 NVL differ in memory and NVLink bandwidth [8]. It fixes GPUs per node, interconnect, region, pricing type, term and capture date, then sets the bundles and egress side by side. What posted GPU cloud prices tell you, and what they do not sets out seven such adjustments.
Posted prices also stop short of contracts. CoreWeave says its customers generally buy multi-year commitments on a take-or-pay basis [1], the CMA found committed spend agreements widespread [3], and SemiAnalysis writes that smaller labs get stuck with large prepayments and worse prices than profitable ones [6].
Rillor's GPU and neocloud pricing method keeps hyperscalers, the large general-purpose clouds, and neoclouds, the clouds built around GPU capacity, in separate segments, and puts two prices on one basis before they are compared.
Read the full report
The full report adds the per-GPU bundle table, the posted commitment and interruptible rates on one basis, the method, its limits and all 35 sources.