Amazon Web Services Review for AI and Machine Learning Workloads

The research in brief

Amazon Web Services is a strong fit for GPU cloud compute for AI training when a team needs H100/H200-class instances, multi-GPU distributed training, and flexible purchasing modes. Three of six platforms named AWS during the ranking stage (50% share), at an average listed rank of 6.67 and a best rank of 5. The strongest reason to consider it is the combination of 8-GPU P5/P5e nodes, EFA networking for multi-node NCCL training, and three purchasing paths (On-Demand, Spot, Capacity Blocks). The main limitation is cost and complexity: AWS carries high on-demand list prices, variable Spot interruption risk, dynamic Capacity Block pricing, and significant non-GPU charges for storage, egress, and orchestration.

Research Snapshot

Why Amazon Web Services Qualified for This Study

Questions explored

• Is Amazon Web Services a good choice for GPU cloud compute for AI training?
• Why did Amazon Web Services qualify for this GPU cloud hosting study?

AWS qualified because it met the study's GPU criteria across every evaluated factor: GPU model availability, multi-GPU cluster support, framework and driver support, and spot or preemptible pricing options. Three of six platforms named AWS during ranking discovery, and all six platforms that produced fit research assessed it as either a strong or good fit for GPU cloud compute for AI training.

The entity's relevant offering is Amazon EC2 GPU instances, particularly P5/P5e/P5en, P4d/P4de, and G5 families, plus EC2 Spot Instances and EC2 Capacity Blocks for ML ^[18]. AWS documents P5.4xlarge with one NVIDIA H100 80-GB GPU, P5.48xlarge with eight H100 GPUs, P5e/P5en with eight H200 GPUs, P4d/P4de with eight A100 GPUs, and G5 instances with NVIDIA A10G GPUs ^[18]. Exact regional availability is capacity-dependent and must be checked before launch ^[18].

This review evaluates AWS only for GPU cloud compute for AI training. It is not a broad company review, and it does not assess AWS services outside that use case. The full field of providers is covered in the GPU cloud compute for AI training consensus index.

The hosting offering in focus

Questions explored

• Which Amazon Web Services GPU instance should a buyer choose for H100 training?
• Does Amazon Web Services offer 8-GPU nodes for distributed AI training?

The most relevant AWS offering for this use case is Amazon EC2 P5 and P5e instances, paired with EC2 Spot Instances and EC2 Capacity Blocks for ML. P5.48xlarge is the flagship 8-GPU H100 node; P5e/P5en add H200 GPUs with 141 GB HBM3e memory per GPU and GPUDirect RDMA ^[24]. P5.4xlarge provides a single-H100 entry point for right-sizing smaller runs ^[38].

For distributed training, AWS documents EFA support for AI/ML workloads, NCCL, RDMA, and selected GPUDirect capabilities ^[22]. AWS also documents multi-node NCCL distributed training using EFA on EC2-backed EKS node groups ^[19]. P5 and P5e/P5en instances support up to 3,200 Gbps of EFA clustering bandwidth ^[25].

For software, AWS Deep Learning AMIs provide preinstalled frameworks, GPU software, model servers, and visualization tools ^[16]. DLAMI Conda environments update CUDA and cuDNN for supported GPU instances and install optimized framework versions ^[17], and AWS publishes GPU DLAMI releases containing NVIDIA drivers, CUDA, cuDNN, PyTorch, NCCL, and EFA components ^[15].

What the AI Platforms Agreed About

Questions explored

• What do AI platforms agree Amazon Web Services does well for AI training?
• Is Amazon Web Services reliable for multi-node GPU training clusters?

All six platforms that produced fit research agreed on several points. First, AWS offers a broad, current NVIDIA GPU portfolio spanning A100, H100, and H200 classes, with 8-GPU nodes available ^[18]. Second, AWS supports multi-GPU and multi-node distributed training through EFA and NCCL ^[22]. Third, AWS provides multiple purchasing modes: On-Demand, Spot, and Capacity Blocks ^[2].

Platforms also agreed that Spot Instances can materially reduce compute cost for checkpointable training, with AWS stating discounts of up to 90% versus On-Demand ^[14]. Agreement here reflects platform-reported consensus, not independently verified product quality.

A fourth shared finding: AWS Deep Learning AMIs and container support reduce initial driver, CUDA, and framework setup effort ^[16]. Fifth, platforms agreed that AWS is strongest for teams already operating workloads, storage, and data on AWS (ChatGPT, DeepSeek, Kimi).

Where the AI Platforms Disagreed or Were Uncertain

Questions explored

• How much does Amazon Web Services charge per H100 GPU-hour, and why do sources disagree?
• Is Amazon Web Services Spot pricing for P5 instances predictable enough to budget for?

Pricing was the sharpest area of disagreement. Published Capacity Block rates were the most consistent figure: $5.191 per H100 GPU-hour, equivalent to $41.528 per eight-H100 P5.48xlarge instance-hour in several US regions ^[3]. Grok reported a different Capacity Block figure of approximately $5.97 per accelerator-hour effective October 2026 ^[3], and Google reported $5.970 per accelerator-hour for P5, $6.866 for P5e, and $7.895 for P5en as of October 7, 2026 ^[26]. These figures conflict and should be verified directly.

On-demand pricing for p5.48xlarge was also inconsistent. Grok reported a range of roughly $55–98 per hour ^[32]. Kimi reported $55.04 per hour on-demand and approximately $19.691 per hour Spot ^[34]. Google reported approximately $98.32 per hour on-demand for p5.48xlarge ^[38]. Perplexity noted that a stable, directly verified on-demand price was not confirmed from AWS-owned pages, and that third-party trackers show $55.04 per hour ^[31].

Spot pricing for P5 was not identified as a single fixed public rate by any platform. OpenAI stated that current P5 Spot pricing must be checked in the AWS console or API (ChatGPT). DeepSeek marked exact 2026 rates as unverified and reported low pricing confidence (DeepSeek). Kimi reported high pricing confidence but sourced its Spot figure from a third-party tracker ^[34].

Platforms also differed on fit rating: ChatGPT and Grok rated AWS a strong fit; DeepSeek, Gemini, Kimi, and Perplexity rated it good. DeepSeek's research date was 2026-06-12, earlier than the 2026-10-03 run date, so its pricing statements are older provenance.

Features for this workload

Questions explored

• Does Amazon Web Services support multi-node NCCL training with EFA for large AI models?
• Which NVIDIA GPU models does Amazon Web Services offer for AI training in US regions?

AWS provides the core capabilities this use case requires. On GPU availability, the portfolio includes P5.4xlarge (1× H100 80 GB), P5.48xlarge (8× H100), P5e/P5en (8× H200), P4d/P4de (8× A100), and G5 (A10G) ^[18]. G5 is positioned for graphics-intensive applications and machine learning inference, and can train simple to moderately complex models ^[6]. AWS claims G5 can offer up to 15% lower cost-to-train than EC2 P3 instances ^[6].

On multi-GPU clustering, P5.48xlarge supplies eight H100 GPUs in one instance, and EFA supports distributed AI/ML networking integrated with NCCL ^[2]. Capacity Blocks can reserve colocated UltraCluster capacity for one to 64 instances, subject to supported offerings and availability ^[2]. Grok reported Capacity Blocks up to 64 instances, or 512 GPUs ^[11].

On frameworks and drivers, DLAMIs ship preinstalled deep-learning frameworks, GPU software, drivers, CUDA, and cuDNN ^[16]. Teams using custom CUDA, kernels, or framework versions may need to manage compatibility themselves (ChatGPT).

On capacity predictability, Capacity Blocks allow buyers to reserve supported accelerated instances for a future start date, with reservations of up to six months ^[2]. Reservations are charged upfront and are not generally modifiable or cancellable after purchase ^[21].

Pricing, renewal terms, and ongoing costs

Questions explored

• How much does Amazon Web Services cost per month for GPU training, and what fees are added?
• Are Amazon Web Services Capacity Blocks refundable or cancellable after purchase?

AWS pricing is usage-based and varies by instance type, region, operating system, purchase model, storage, data transfer, and ancillary services (ChatGPT). The clearest published GPU figure found across platforms is the Capacity Block rate of $5.191 per H100 GPU-hour for P5 in several US regions, with dynamic reservation pricing ^[3]. Capacity Block pricing is dynamic, charged upfront, and includes applicable operating-system charges ^[21].

Spot pricing varies with supply and demand, can be interrupted, and provides a two-minute interruption notice ^[14]. AWS states Spot Instances can be discounted up to 90% versus On-Demand ^[8]. Kimi reported p5.48xlarge Spot at approximately $19.691 per hour, roughly 64% below its reported on-demand rate, but this figure came from a third-party tracker ^[34].

Known additional costs include EBS or other attached storage, S3 or FSx durable storage, data transfer and networking charges, and charges for orchestration or managed services such as EKS, SageMaker, or monitoring (ChatGPT). Kimi reported egress fees of $0.08–$0.09 per GB and persistent storage of $0.02–$0.10 per GB per month ^[34]. Google reported data egress fees, EBS and S3 storage fees, and Capacity Block operating-system fees billed in addition to the upfront reservation fee (Gemini).

On contracts, On-Demand and Spot EC2 are generally usage-based rather than requiring a long-term contract (ChatGPT). Capacity Block payment is charged upfront, and AWS documentation states that Capacity Blocks cannot be modified or canceled after purchase ^[21]. Google reported that Savings Plans and Reserved Instances require a 1-year or 3-year commitment with penalties for early cancellation (Gemini). AWS's own pricing page describes Savings Plans as offering savings of up to 72% compared to On-Demand in exchange for a one- or three-year commitment (official:C3).

Pricing confidence varied by platform: high for Kimi and Gemini, moderate for ChatGPT, Grok, and Perplexity, and low for DeepSeek. Buyers should treat all figures as platform-reported and verify current rates in the AWS console or pricing pages.

Best Suited For

Questions explored

• Who gets the most value from Amazon Web Services for GPU cloud compute for AI training?

AWS is best suited for distributed training requiring single-node 8-GPU servers or multi-node clusters (ChatGPT). It fits teams already using AWS storage, networking, IAM, EKS, S3, FSx, or SageMaker (ChatGPT, DeepSeek). It suits workloads that can checkpoint and tolerate Spot interruptions (ChatGPT, DeepSeek, Kimi). It fits short, scheduled high-performance training runs using EC2 Capacity Blocks (ChatGPT, DeepSeek). It also fits teams needing H100, H200, or A100 GPU choices in US regions (ChatGPT, Gemini).

Kimi added that AWS suits enterprises needing SOC 2, ISO 27001, and FedRAMP compliance certifications, and teams requiring multi-region deployment or integration with existing AWS cloud spend (Kimi). Grok highlighted large-scale distributed ML training requiring EFA networking and UltraClusters ^[11].

Probably Not Best Suited For

Questions explored

• Who should not choose Amazon Web Services for GPU cloud compute for AI training?

AWS is probably not best for small teams wanting a turnkey GPU notebook or highly simplified managed training experience (ChatGPT). It is a weaker fit for jobs that cannot tolerate interruption or do not implement checkpointing (ChatGPT, DeepSeek). It is not ideal for buyers optimizing solely for the lowest publicly advertised hourly GPU price (ChatGPT, DeepSeek, Kimi, Gemini). It is also a poor fit for teams without AWS operations expertise that want minimal cluster and driver administration (ChatGPT, Kimi).

Kimi added that AWS is not best for teams needing the newest NVIDIA silicon such as GB200/GB300 NVL72 immediately without lead time, or teams without Kubernetes/Slurm expertise who want training-optimized orchestration out of the box (Kimi). Google noted AWS is not best for users expecting stable, flat-rate pricing for reserved GPU capacity, due to dynamic pricing on EC2 Capacity Blocks (Gemini).

When Another Option May Be Better

Questions explored

• What is a better alternative to Amazon Web Services for a buyer who needs the lowest H100 GPU-hour price?
• When should a buyer choose a specialized GPU cloud over Amazon Web Services for AI training?

Another option may be better in several specific situations. Choose a specialized GPU cloud when the primary objective is a simpler user experience or a lower transparent hourly price (ChatGPT). Choose a managed training platform when the team wants less responsibility for cluster orchestration, job recovery, and environment management (ChatGPT). Choose another provider when AWS cannot offer the required GPU count, region, or start time at acceptable pricing (ChatGPT).

Kimi named specific alternatives: CoreWeave, Lambda Labs, and Nebius for 40–70% lower effective rates; Crusoe, CoreWeave, or Nebius for earlier GB200/GB300 NVL72 availability; Lambda Labs or CoreWeave for zero egress fees; CoreWeave, Lambda, or Together AI for training-optimized orchestration; and Nebius for EU data residency (Kimi). These are platform-reported comparisons, not independently verified.

Choose On-Demand or reserved capacity rather than Spot when training cannot safely resume after interruption (ChatGPT). Choose a different provider if uninterrupted access to large H100 clusters is critical and the buyer cannot tolerate Spot risk (Perplexity). Choose a different option if the buyer mainly needs inexpensive inference or small training jobs rather than H100-scale training (Perplexity).

Questions to ask before choosing a provider

Questions explored

• What should a buyer confirm with Amazon Web Services before signing a GPU compute contract?

Before committing, verify the following with AWS directly. Can AWS provide the required number of P5, P5e, P5en, or P4d instances in the intended US region and Availability Zone at the required start time (ChatGPT)? What are the current Spot prices, interruption history, and capacity-availability signals for the exact GPU instance and region (ChatGPT, DeepSeek)? Does the selected instance and DLAMI support the required CUDA, NVIDIA driver, PyTorch, TensorFlow, NCCL, and custom-kernel versions (ChatGPT, Perplexity)?

Also confirm: will the workload use EFA, and has multi-node NCCL performance been benchmarked with the intended topology (ChatGPT)? Would Capacity Blocks or On-Demand be cheaper than Spot after including interruption-related restart time and storage costs (ChatGPT)? What are the exact Capacity Block start, end, cancellation, modification, and refund terms for the proposed reservation (ChatGPT, Perplexity)? What quotas, account approvals, service limits, or purchase restrictions apply to the required GPU quantity ^[12]?

Finally, verify expected EBS, FSx, S3, data-transfer, EKS, SageMaker, monitoring, and support-plan costs (ChatGPT, Kimi); how checkpoints, logs, datasets, secrets, and recovery state persist if a Spot instance is interrupted (ChatGPT); and whether the buyer needs AWS enterprise pricing, Savings Plans, Reserved Instances, or a negotiated commitment, and what minimum commitment would apply (ChatGPT, Kimi).

The consensus assessment

AWS is a strong fit for technically capable AI startups and data-science teams that need high-end NVIDIA GPUs, distributed training, AWS integration, and multiple purchasing options. The main tradeoffs are infrastructure complexity, uncertain GPU availability, variable Spot economics, and potentially significant non-GPU charges. AWS is strongest when the buyer values ecosystem depth and cluster capability over the simplest or lowest headline price.

Platform ratings split: ChatGPT and Grok rated AWS a strong fit; DeepSeek, Gemini, Kimi, and Perplexity rated it good. No platform rated it a poor fit. The consensus is that AWS is a credible, capable option for this use case, with cost and operational complexity as the recurring caveats. This verdict reflects platform-reported assessments, not independently verified performance.

How This Review Was Produced

This review was produced from six platform fit-research responses collected for the study date 2026-10-03. Each platform evaluated AWS against the study's GPU criteria: GPU model availability, GPU pricing per hour, multi-GPU cluster support, framework and driver support, and spot or preemptible pricing options. Three of six platforms named AWS during the ranking stage. All six produced fit research. Platform fit ratings, pricing figures, and capability claims are reproduced as reported. Company-owned citations materially outnumber independent citations in the supplied evidence, so AWS claims should not be described as independently verified. The broader category is covered in the vps cloud and dedicated server hosting directory.

Research limitations

Several limitations apply. Platform-reported research dates differ from the authoritative run date: DeepSeek reported 2026-06-12, while ChatGPT, Gemini, Grok, Kimi, and Perplexity reported 2026-10-03. Platform-reported dates are provenance metadata and do not independently prove freshness. All included platforms evaluated fit, but platform_mentions counts only platforms that named the entity during ranking discovery.

Conflicting product names, pricing, and capabilities were not resolved by guessing; conflicts are described above and buyers should verify directly. The supplied URLs were collected from platform responses and were not independently validated by the writer stage. Citations are platform-reported evidence, not independently verified facts. DeepSeek's research ran without search enabled, so its claims require explicit verification before being treated as current facts. No-search model claims are labeled platform-reported.

See the broader AI and Machine Learning Workloads consensus index for comparisons across qualified options.

Sources

Company-Owned Sources

1. Announcing Amazon EC2 Capacity Blocks for ML: https://aws.amazon.com/blogs/aws/announcing-amazon-ec2-capacity-blocks-for-ml-to-reserve-gpu-capacity-for-your-machine-learning-workloads/
2. Amazon EC2 Capacity Blocks for ML: https://aws.amazon.com/ec2/capacityblocks/
3. Amazon EC2 Capacity Blocks for ML Pricing: https://aws.amazon.com/ec2/capacityblocks/pricing/
4. Amazon EC2 Accelerated Computing Instances: https://aws.amazon.com/ec2/instance-types/#Accelerated_Computing
5. Accelerated computing Amazon EC2 instance types - AWS: https://aws.amazon.com/ec2/instance-types/accelerated-computing/
6. Amazon EC2 G5 Instances: https://aws.amazon.com/ec2/instance-types/g5/
7. Amazon EC2 P5 Instances: https://aws.amazon.com/ec2/instance-types/p5/
8. Pricing: https://aws.amazon.com/ec2/pricing/
9. Amazon EC2 Spot Instances: https://aws.amazon.com/ec2/spot/
10. 適用於ML 的Amazon EC2 容量區塊定價: https://aws.amazon.com/tw/ec2/capacityblocks/pricing/
11. Capacity Blocks for ML - Amazon Elastic Compute Cloud: https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/ec2-capacity-blocks.html
12. Amazon EC2 Service Quotas: https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/ec2-resource-limits.html
13. Spot Instance interruptions: https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/spot-interruptions.html
14. Spot Instances: https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/using-spot-instances.html
15. AWS Deep Learning OSS Nvidia Driver AMI GPU PyTorch: https://docs.aws.amazon.com/dlami/latest/devguide/aws-deep-learning-x86-gpu-pytorch-2.14-ubuntu-26-04.html
16. Features of DLAMI: https://docs.aws.amazon.com/dlami/latest/devguide/features.html
17. Deep Learning AMI with Conda: https://docs.aws.amazon.com/dlami/latest/devguide/overview-conda.html
18. Specifications for Amazon EC2 accelerated computing instances: https://docs.aws.amazon.com/ec2/latest/instancetypes/ac.html
19. Run machine learning training on Amazon EKS with Elastic Fabric Adapter: https://docs.aws.amazon.com/eks/latest/userguide/node-efa.html
20. Recommended GPU Instances: https://docs.aws.amazon.com/en_us/dlami/latest/devguide/gpu.html
21. Capacity Blocks pricing and billing: https://docs.aws.amazon.com/us_en/AWSEC2/latest/UserGuide/capacity-blocks-pricing-billing.html
22. Elastic Fabric Adapter for AI/ML and HPC workloads on Amazon EC2: https://docs.aws.amazon.com/us_en/AWSEC2/latest/UserGuide/efa.html
23. AWS Announces New Amazon EC2 P5 Instance Size: https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF0TSkZ_CCkb0zGgYLvmwMUTWm7vN2PeiL7Rpl735o95x8wNsWFTV3LsviOdBC8--b9ka-fkvIw5oR-BoDW6u3QwXbsmWube3RNxdLWG0oUOKwbD8TrbAVC2wa7YtEAmcdnkibB2u9-gqk-tbJxdXQ-PWG_aDb6YHGTjNcBz5qpnNfpcQRnqfwcEpkFwygKQGyGLaXc1ND-SYH0
24. EC2 P5e Instances Overview: https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF3PsJ-ljPxZaK2F5gups2XapDo8jzbL-e6wWzuwVukrVLajwXfz84BtORExzET_FIDgTKY3ZkAkqPaRGKrBMN9t9gXmR17GzxXqkUx3ZC6T3c46i2dE4OLXGmFHcW8ukSy0_4s2Eu98Wim2P-UVpD0OHFih2ezx_wWIVzi-tkBiTkEh6oPiLS3DlwrMH773p0M5nmRMuplImQ=
25. Amazon EC2 P5 Instances Product Page: https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGNMn9RQ1uYjrytDnWN2hO45xIG8FxstZg8jgkvmjo9P1maU5TdT3cpUUo43xp0CFf1XQKZoXTDdNB2v4Oc-JtfOSGXNPnYlWnKVU25yAv73_tx4yuOlbV6xHOZIlhPfQsk6q-8nZ8v
26. Amazon EC2 Capacity Blocks for ML Pricing: https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHSLmPIW3iWB4avxB22HQ-qoBeSSDe438-lbAdrX4AnuYz1JM79u78W-iVNC_2uR3iitI8UAaLuIFZK2jJukThE-CKdOsYWZzas3affLr7ptCbwyp2SIOBh1p-j39eM36yDO3I0jRCEgkDd5xE=
27. Official pricing and terms source: https://aws.amazon.com/
28. Official pricing and terms source: https://aws.amazon.com/pricing/?nc2=h_pr_hub
29. Official pricing and terms source: https://aws.amazon.com/savingsplans/?nc2=h_pr_sp

Independent Sources

30. GPU Prices 2026: H100 From $3.49/hr, RTX 5070 at +58%: https://axis-intelligence.com/gpu-prices/
31. p5.48xlarge specs and pricing | AWS | CloudPrice: https://cloudprice.net/aws/ec2/instances/p5.48xlarge
32. AWS GPU Instance Pricing and Availability: P5, P4, G5, and Trn1 in 2026: https://clusterbid.com/blog/aws-gpu-instances-p5-p4-g5-trini-2026-pricing
33. Compare GPU Cloud Providers - Side-by-Side Pricing: https://computecomparison.com/compare-gpu-cloud-providers
34. How to Choose a GPU Cloud Provider for AI Model Training: https://gpusmith.com/articles/en/choose-gpu-cloud-provider-ai-training
35. p5.48xlarge pricing and specs - Amazon EC2 Instance Comparison: https://instances.vantage.sh/aws/ec2/p5.48xlarge?selected=t3.small
36. ClusterMAX 3.0: The Industry Standard GPU Cloud Rating System Returns: https://newsletter.semianalysis.com/p/clustermax-30-the-industry-standard
37. Understanding AWS Spot Instances and Interruption: https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEpHvZ-iZe6S01debRswNgkpFmigvvH72QnyytfT1EM39L_qk4q0Mp1plz1KtZmW7-CLO8nUIxK-1f7XuQGbrVz7WC6Us8jqMdh1_sDnhQeE7lWu7xQwnYuO9X-_TutlPzBx80RMjarJn-lBo5Y
38. EC2 p5.4xlarge Specifications and Pricing: https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEqPpwGgAMOB9DYp60CFuzTBFYyes3fKZWbtspu4Tti9DsFWHDwVRXxT-M6GDGVLf6Ts-JupDeP_f7Sn1f3vN3xEwkpEYKfCnHuKe-l1qvlhMGE6C75_aL0yse-oLNVbulFM18zqDzs658=
39. AWS GPU Instance Pricing Compared: https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFq5jBHF1cG1fBtJtoYkoN88exyOe7DYxlY0IldUoJ2Mh4W322jxuDNpsVb41z5XVrOgck_Hs1c7IzEr1Z5n6qTHyIVy4XokXCDFBvVNJ9QmQfDaw7TLrYLCoVqGkQ7Xq10ZfgIicsTQkITqqKCRg==
40. AWS GPU Instance Pricing: P5, P4d, G5, Inf2 - Wring: https://wring.co/blog/aws-gpu-instance-pricing-guide