TL;DR — Straight-to-the-point answer：RTX 4090 generates one SDXL image in ~3–5 seconds; at ¥2/hour rental ratePer-GPU cost under ¥0.002Moderate usage (100 images/day) costs ~¥60/month—just a fraction of a Midjourney subscription—and supports any open-source model or LoRA.
Benchmark data (SDXL + ComfyUI)
- SDXL 1024×1024: 3–5 seconds per image (on RTX 4090); ~6–9 seconds on RTX 3090.
- SD1.5 512×512: 1–2 seconds per image; batch processing 8 images concurrently is faster.
- Flux.1 dev (FP8): 15–25 seconds per image, runs on 24GB VRAM
Cost Comparison of Three Usage Modes (Monthly)
| Solution | Monthly cost | Who it’s for |
|---|---|---|
| Rent an RTX 4090 on a pay-as-you-go basis (1 hour per day) | Approx. ¥60 | Moderate creators wanting to experiment with any model or LoRA |
| Midjourney Standard Membership | Approx. ¥210 | Just want it to look good—no hassle |
| Buy an RTX 4090 | $14,000+ one-time | Heavy user generating 500+ images daily (~2-year ROI) |
Current RTX 4090 hourly rental rate
| Platform | Price ⇅ | 24 hours | Inventory | Billing Information | |
|---|---|---|---|---|---|
| Vast.ai (USA)Global lowest priceNorth America
Stability 100·A |
¥0.87/hour▼40.4%Currently Monitored Low PriceReal-timeSpot | ¥21/day | sufficient | Vast.ai real-time spot price converted (RTX 4090, USD × 7.2); subject to change at any time | Rent Now |
| Salad (USA)Bulk lowest pricesNorth America
Stability 74·C |
¥1.15/hourReal-timeSpot | ¥28/day | sufficient | Salad Distributed Cloud lowest priority tier: $0.160/hr (24GB VRAM; 4 tiers total—higher tiers offer greater stability) | Rent Now |
| PPIO Paiou CloudLow-cost domestic spot pricingChina
Stability 71·C |
¥1.16/hourReal-timeSpot | ¥28/day | sufficient | PPIO Paiou Cloud GPU spot price: ¥1.16/hr (RTX 4090; default 1-hour protection period, 1-hour termination notice) | Rent Now |
| SuanjiayunCurrent Low PriceChina
Stability 100·A |
¥1.24/hourReal-timePay-as-you-go | ¥30/day | sufficient | Suanjiayun official website lists ¥1.24/hour (original ¥2.08; additional discounts available for long-term rentals) | Rent Now |
| MornAILong-Term Rental Offers Better ValueChina
Stability 100·A |
¥1.38/hourReal-timePay-as-you-go | ¥33/day | sufficient | MornAI official website daily rate: ¥33.01/day ÷24 (NVIDIA GeForce RTX 4090 24G); pay-as-you-go hourly rate may be slightly higher. | Rent Now |
| AIStationChina
Data collection in progress: 5/30 days |
¥1.65/hourReal-timePay-as-you-go | ¥40/day | sufficient | AIStation AI Galaxy (China) Low-Price Zone listed pay-as-you-go price: ¥1.65/hour (RTX 4090 24G) | Rent Now |
| CompshareCash RebateChina
Stability 100·A |
¥1.92/hourReal-timePay-as-you-go | ¥46/day | sufficient | Compshare pay-as-you-go minimum configuration starts at ¥1.92/hour (lower rates for daily/monthly/spot instances available on official website; 5% discount for universities and enterprises) | Rent Now |
| AutoDLMost Comprehensive EcosystemChina
Stability 98·A |
¥1.98/hourReal-timePay-as-you-go | ¥48/day | sufficient | AutoDL pay-as-you-go lowest price: ¥1.98/hour (official marketplace RTX 4090, idle GPU count: 62) | Rent Now |
| MatPoolListed-price competitivenessChina
Stability 100·A |
¥2.20/hourReal-timePay-as-you-go | ¥53/day | sufficient | MatPool pay-as-you-go pricing: ¥2.20/hour (Zone 11, clearly marked on official site) | Rent Now |
| RunPod (USA)Top overseas favoriteNorth America
Stability 73·C |
¥2.45/hourReal-timeSpot | ¥59/day | sufficient | RunPod Community Cloud lowest price: $0.34/hr (NVIDIA GeForce RTX 4090; Secure Cloud: $0.89/hr) | Rent Now |
| Cloudrift (United States)Structured, transparent pricingNorth America
Data accumulation in progress: 2/30 days |
¥2.81/hourReal-timePay-as-you-go | ¥67/day | sufficient | Cloudrift official website JSON-LD RTX 4090 On-Demand Hourly Rate converted from $0.39/hr | Rent Now |
| ParateraRTX 3090 Low PriceChina
Data Accumulation: 1/30 Days |
¥3.60/hourReal-timePay-as-you-go | ¥86/day | sufficient | Paratera pay-as-you-go real-time price ¥3.60/hour | Rent Now |
| LanyunChina
Manually maintained |
¥2.00/hourReference price · Verification pendingPay-as-you-go | ¥48/day | sufficient | Real-world stability is solid; response time is fast ⚠ Note on pricing methodology: official prices are dynamically loaded via frontend APIs, with no publicly scrapable source available; this site conducts manual verification once per quarter. |
Rent Now |
| TensorDock (USA)Global Low PricesNorth America
Manually maintained |
¥2.52/hourReference price · Verification pendingPay-as-you-go | ¥60/day | sufficient | Official starting price: $0.35×7.2 (verified as of Oct 2026); budget-tier machines start at $0.25/hour but frequently run out of stock; third-party data centers ⚠ Note on pricing methodology: this site uses the official “minimum 8-GPU rental” pricing tier; these figures are not fully comparable to single-GPU pay-as-you-go rates from other platforms—provided for reference only. |
Rent Now |
Pro tips to save even more
- Deploy the environment using the platform’s “No-GPU Mode”Model loading and LoRA download don’t consume GPU time—power on only after deployment is complete.
- Mount data disk for model library：Download multi-GB models only once; mount them directly on subsequent startups
- Start jobs during off-peak hours：Spot instance availability—and lower prices—are common during off-peak hours (e.g., overnight or weekday daytime)
Prices update hourly—check the live table above before placing an order. Unsure whether renting or buying suits your usage? Plug your daily image output volume into the formula above. Some links include affiliate commissions, which do not affect ranking.