As of 2026-10-26, the lowest price across all platforms is approximately ¥8.28/hour (Paperspace, US); domestic platforms offer pay-as-you-go rates around ¥20–¥30/hour. Full price comparison across 26+ platforms is available at H100 Price Comparison Pageupdated hourly.
A100 80GB (widely used for LoRA fine-tuning and small-to-medium training; lowest rate across platforms ≈ ¥5.47/hour) or A800 (≈ ¥5.15/hour); if budget allows, go straight to H100. Full-parameter training on consumer-grade GPUs is not recommended—insufficient VRAM leads to frequent OOM errors. Unsure which GPU to pick? Use this site.Free GPU Selection ToolAutomatically recommends GPUs by task.
For large-model inference, RTX 4090 delivers ~60–80% higher throughput than 3090 (both with 24 GB VRAM; 4090 memory bandwidth nearly doubles). Lowest available rates: ¥0.87 vs ¥0.63/hour—4090 becomes increasingly cost-effective with longer runtime.
No influence. rentgpu.link always sorts strictly by price and accepts no paid rankings from any platform. Some “Rent Now” links include affiliate commissions, clearly marked on-page—ranking and inclusion remain unaffected.
Choose domestic platforms (AutoDL, Gpushare, Suanjiayun, etc.) for stability and local payment support; choose overseas platforms (Vast.ai, Salad, etc.) for rock-bottom pricing—if you can tolerate spot/preemptible volatility. Add ~200 ms latency budget for cross-border network delays. See detailed pros and cons per platform atPlatform Cross-Comparison。
Full-platform pricing data (26+ platforms) scraped automatically every hour; homepage and per-GPU comparison pages refresh in sync; public price API (/api/prices) updates simultaneously and may be freely referenced.
By GPU model: Entry-level GPUs (RTX 3090/4090) start as low as ¥0.63–¥0.87/hour across platforms; professional GPUs (A100/A800) average ~¥5.47/hour; flagship GPUs (H100) average ~¥8.28/hour (Paperspace (USA)). The same GPU often varies 2–5× in price across platforms — full ranges shown inPrice insights。
For inference and entry-level fine-tuning: RTX 4090 (24 GB VRAM, low unit price); for large-batch training, 80 GB VRAM, or NVLink multi-GPU interconnect: A100. Rule of thumb: use RTX 4090 if your model + optimizer state fits within 24 GB; otherwise, choose A100.
If you use GPU resources for over 6–8 hours per day, opt for monthly subscription (typically 20%–40% cheaper); for sporadic debugging, use hourly rental. For one-off large jobs, first consider spot/preemptible instances (lower cost but subject to interruption—ensure your workload supports checkpoint resumption).
Typically 30%–70% cheaper than pay-as-you-go, but subject to forced termination when platform compute is tight (usually with several minutes’ notice; some platforms provide no warning). Suitable for checkpoint-friendly training, batch inference, and rendering; unsuitable for live services. This site records spot and pay-as-you-go prices separately—spot prices are excluded from the “Lowest Price Recommendation” ranking.
Domestic: AutoDL, Gpushare, MatPool (per-second/per-hour billing, Chinese UI, Alipay & WeChat Pay supported); Overseas: Vast.ai, RunPod (USD pricing, often featuring rock-bottom rates). Students should claim new-user vouchers from each platform first to save further.
No need to configure from scratch. Major platforms provide preconfigured images (PyTorch/CUDA/Stable Diffusion, etc.); simply select an image and launch instantly. Data can be mounted via cloud storage or object storage. Note: For pay-as-you-go instances, the system disk is typically not retained after shutdown—store critical data on data disks or object storage.
Four channel types serve distinct purposes: ① Third-party price comparison sites show the lowest real-time spot price—rentgpu.link (26 platforms, scraped hourly, only in-stock offers included), suanli.ai (largest service provider coverage, but timeliness requires verification), gpus.io (global focus, 32 providers, 2,648 configurations, with API and historical trend charts), GpuPerHour (25+ cloud providers); ② Official pricing pages of cloud vendors show list prices and contractual discounts—Alibaba Cloud, Tencent Cloud, VolcEngine, AWS, Azure, GCP; ③ Industry price indices reflect market trends and negotiation benchmarks—Mysteel Compute Rental Price Index (launched September 2026, updated daily on business days, covering the eight national computing hubs under the East Data West Computing initiative and North America); ④ Government compute scheduling platforms offer matchmaking and subsidies—Hangzhou Municipal Compute Resource Scheduling Service Platform (up to 30% subsidy via compute vouchers). No single source covers all needs; cross-check at least 2–3 sources during actual inquiries; full directory and suitability guidance per platform available in the Comprehensive Guide to GPU Compute Price Lookup Platforms.Comprehensive list of GPU compute price lookup platforms。
Four reasons: ① Different billing types (spot vs. pay-as-you-go vs. monthly); ② GPU variant differences (e.g., 40G/80G VRAM, NVLink availability); ③ Inclusion of CPU/memory/storage/bandwidth; ④ Platform subsidies and time-based discounts. This site only compares identical models, generations, and VRAM capacities, and explicitly notes billing methodologies in annotations to prevent sorting by non-comparable prices.
rentgpu.link
Neutral GPU compute rental price comparison tool: collects publicly listed prices from 26 global platforms every hour. We do not own GPUs nor sell compute.