Same H100. AWS charges $6.88. We charge $3.582.

Competitor rates are each vendor's own published on-demand price — click a figure to check the source. Whole-node prices are divided by GPU count.

• Price

  NexGPU · H100 SXM 80GB at $3.582, A100 SXM4 80GB at $1.088, H200 141GB at $6.660 — all list, all per GPU-hour, metered per second and priced per hour with no minimum and no setup fee. Consumer parts go lower still: RTX 4090 $0.540, RTX 3090 $0.193. Storage runs $0.414/GB-month median, egress $0.0081/GB median.

  AWS · AWS lists p5.48xlarge at $55.04/hr in us-east-1 — $6.88 per H100 GPU-hour. The identical machine in Tokyo is $68.80/hr, or $8.60/GPU-hr: 25% more for the same silicon. A100 80GB on p4de.24xlarge is $27.4471/hr, or $3.4309/GPU-hr. And Capacity Blocks, the channel H100/H200 capacity actually moves through, has no list price at all — AWS says the price depends on supply and demand at the time of purchase, visible only after you log in.

• Time to first boot

  NexGPU · Sign up, top up, boot. Three steps. 2,000+ prebuilt images — PyTorch, vLLM, ComfyUI, Stable Diffusion, AUTOMATIC1111, Whisper ASR, Ubuntu CLI — reachable over SSH, Jupyter or a web terminal, with a REST API and CLI ready the same minute. No quota request, no approval queue, no prepayment, no booking window. Shut it down whenever you like; billing stops by the second.

  AWS · A new AWS account's GPU quota is zero. Running On-Demand P instances defaults to 0 vCPUs; Running On-Demand G and VT instances is 0 too, so even a single-GPU g6.xlarge is blocked. Spot is 0 as well. Quotas count vCPUs, and one p5.48xlarge is 192 of them, so you are asking AWS to move you from 0 to 192 — a request routed to AWS Support, which can approve, partially approve, or deny. For H100/H200 the practical route is Capacity Blocks: its own quotas, also 0 by default, charged in full up front, released if payment does not clear at least 5 minutes before start or within 12 hours, and per AWS, cancellations aren't allowed. Blocks end on AWS's clock at 11:30 UTC, with termination starting at 11:00. Build a custom AMI for P5 and you are pinning NVIDIA driver 530, CUDA 12.1, GDRCopy 2.3, EFA 1.24.1 and NCCL 2.18.3 by hand.

• Global footprint

  NexGPU · 51 countries and regions, 1,175 verified rentable nodes, 2,498 GPUs, 75 GPU models, up to 14 GPUs on a single node and up to 2,152GB of VRAM on one node. One price list, one console — boot wherever the capacity is, with no per-region capacity lottery.

  AWS · AWS's region count is impressive; its GPU capacity is far narrower. AWS's own Capacity Blocks support table lists 18 locations in total: p5 (8x H100) in 11 regions plus one Local Zone, p4d (A100 40GB) in just three — us-east-1, us-east-2, us-west-2 — and p4de (A100 80GB) in exactly two, us-east-1 and us-west-2. The entire A100 80GB fleet is US-only; there is none in Asia or Europe. No mainland China region appears in that table at all. The nearest region to China carrying on-demand H100/H200 is Tokyo, and Tokyo costs 25% more.

• Support

  NexGPU · Bilingual ZH/EN, around the clock, straight to a human on Telegram at t.me/nextipkf. No ticket queue, no severity tiers, no support contract — and support costs 0% of your spend.

  AWS · AWS Basic is free and covers customer service, docs and Trusted Advisor; 24/7 phone, web, chat and email technical support with unlimited cases starts at the paid tiers. Business Support+ is a $29/month minimum plus a percentage of your monthly usage — 9% up to $10K, 7% to $80K, 5% to $250K, 3% above. Support therefore gets more expensive precisely because you rented GPUs: a single p5.48xlarge running a 30-day month is $39,628.80 of EC2 spend, and Business Support+ adds roughly $3,566/month for a sub-30-minute severity-1 target, with everything else falling back to under-24-hour general guidance. Enterprise Support starts at $5,000/month; Unified Operations at $50,000. And that same paid organisation is the gatekeeper on the quota increase you needed before any of it.