{
  "id": 10136880,
  "title": "What it costs to rent an H100, B200 or RTX 4090 in September 2026: live prices from 28 GPU clouds",
  "url": "https://urgent.news/2026/09/27/what-it-costs-to-rent-an-h100-b200-or-rtx-4090-in-september-2026-live",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-09-27T04:11:49.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/fastgpu/what-it-costs-to-rent-an-h100-b200-or-rtx-4090-in-september-2026-live-prices-from-28-gpu-clouds-12n3"
  },
  "original_language": "en",
  "account": "GPU rental prices remain inconsistent, with the same card varying in cost depending on the cloud provider and daily fluctuations. FastGPU compiles prices from 28 GPU clouds, ranking them from cheapest to most expensive. The data, updated on September 27, 2026, at 04:11 UTC, highlights key patterns influencing pricing disparities. Prices are provided hourly and compared to the lowest on-demand price offered by major hyperscalers (AWS, Azure, Google Cloud, Oracle). The comparison reveals significant price differences for flagship GPUs. Hyperscalers charge 3.0x to 3.6x more for the same high-end silicon compared to the cheapest on-demand price. For instance, the H100 SXM 80 GB costs $5.38/hr on AWS, while Cudo Compute offers it for $1.79/hr. The gap widens with newer models like the B200 and B300, where hyperscalers charge 3.1x and 3.6x more respectively. The largest discrepancy is observed for the L4 24 GB, at a rate of $0.71/hr on Google Cloud compared to $0.11/hr on Vast.ai. Hyperscalers charge premium prices due to ecosystem benefits like IAM, networking, managed services, committed-use discounts, and startup credits, which may justify the cost. However, this premium is a deliberate choice rather than an unavoidable requirement. The exception to this trend is the workstation-class RTX PRO 6000, which has a nearly identical price on Vast.ai and Google Cloud, indicating minimal premium for this model. Spot capacity offers further cost savings, but with important caveats. Interruptible capacity reduces prices further: H100 SXM drops to $0.98/hr, H200 to $1.84/hr, and B200 to $1.77/hr. Despite this, spot capacity can be reclaimed at any moment, making it suitable for checkpointed training and offline batch jobs, but not ideal for user-facing endpoints. The cheapest spot listings for whole nodes can be prohibitively expensive when scaled to individual GPU prices. The newest Blackwell GPUs, including the B200, B300, and GB200, are now rentable on-demand. Providers offer rates ranging from $3.69/hr for the B200 to $8.00/hr for the GB200. This expansion into public pricing signals increased accessibility for these cutting-edge models. For smaller models, consumer-grade graphics cards like the RTX 4090 and RTX 5090 offer more competitive pricing. An RTX 4090 with 24 GB VRAM rents for $0.14/hr, while an RTX 5090 with 32 GB VRAM costs $0.21/hr on marketplace hosts. These small models can be trained on a single GPU, making them more cost-effective options for certain workloads. However, small models may face reliability issues with individual consumer hosts, so it's essential to verify host stability before committing to a rental. Marketplaces provide fluctuating prices as hosts come and go, so it's crucial to monitor these rates regularly. Each GPU listing includes a 30-day price strip showing the minimum, typical, and current prices, helping users determine the optimal time to rent. To choose the best GPU rental option, follow these four steps: first, determine the VRAM requirement for your model. This initial step dictates the subsequent decisions. Next, consider the job's tolerance for interruption. Interruptible spot capacity can significantly reduce costs but is best suited for checkpointed training and offline batch jobs. If your job requires continuous, uninterrupted access, opt for non-spot pricing. Third, evaluate the pricing structure and consider the trade-offs between hyperscaler premiums and the ecosystem benefits they provide. While hyperscalers charge a premium, these benefits may justify the cost for certain users. Finally, assess the reliability and stability of consumer-grade rental options. While marketplace hosts offer competitive pricing, ensure the host is reliable and capable of supporting your workload without data loss. By following these steps, you can navigate the complex landscape of GPU rental prices, selecting the most cost-effective solution for your specific needs and use case.",
  "summary": "GPU rental prices have not converged. The same card can cost several times more depending on which cloud you rent it from, and the cheapest option changes through the day. We run FastGPU , which pulls prices from 28 GPU clouds (marketplaces, neoclouds, serverless platforms and the hyperscalers) and ranks them cheapest first. Here is the market as it stood on September 27, 2026 at 04:11 UTC, and…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}