RTX 4090 Cloud Rentals: SDXL Performance vs. Price

The NVIDIA RTX 4090, once the darling of high-end consumer gaming, has found a second life as the prosumer workhorse for local AI inference. Its generous VRAM (24GB) and impressive raw compute make it ideal for tasks like Stable Diffusion XL. Many of us can’t justify the upfront cost, or simply need burst capacity, which is where cloud rentals come in. We decided to see which providers offered the most bang for your buck when it came to generating images on a 4090.

We rented RTX 4090 instances from three prominent providers: Runpod, Vast.ai, and Vultr. Our goal was simple: generate a consistent set of SDXL images and measure both throughput and cost. For each provider, we spun up a machine, installed our standard Stable Diffusion XL environment (Automatic1111 with specific extensions, sd_xl_base_1.0.safetensors, sd_xl_refiner_1.0.safetensors), and ran a script to generate 1,000 images using a common prompt, 20 steps, DPM++ 2M Karras sampler, and 512x512 resolution, followed by a refiner pass. Our full methodology is detailed in our benchmark playbook. We ran each test for a minimum of 24 hours to account for any transient performance dips or network variability.

Runpod: Consistent, but Not Always Cheapest

Runpod offers 4090s across both their Community Cloud and Secure Cloud. We focused primarily on Community Cloud for its typically lower pricing, though we did briefly test Secure Cloud for comparison. Finding an available 4090 on Community Cloud was usually straightforward, though popular regions could see brief shortages during peak hours. The setup process was relatively painless; Runpod provides pre-built Docker templates for various ML frameworks, and we were able to get our Automatic1111 container running with minimal fuss. For those who prefer a more managed experience, their serverless offering is also an option, though we’ve previously noted its cold start peculiarities.

Performance on Runpod was remarkably consistent. Across several instances, we observed an average generation rate of 12.8 images per minute for our SDXL benchmark. This included the refiner step, which can often be a bottleneck. Pricing hovered around $0.38 per hour for Community Cloud, occasionally dipping to $0.35 or peaking at $0.42 depending on demand. This translates to an effective cost of approximately $0.0049 per image.

Support was responsive, and the overall experience felt polished. Billing was straightforward and granular. If you’re looking for a reliable, no-nonsense experience without the need to constantly monitor a marketplace, Runpod is a solid contender. We’ve previously covered their strengths in our Runpod review. For those considering it, you can find them at https://runpod.io/?ref=8vbo5oc9.

Vast.ai: The Wild West of GPU Rentals

Vast.ai is a different beast entirely. It’s a marketplace where individuals rent out their hardware, leading to significant variability in price, performance, and sometimes, stability. Finding a 4090 was never an issue; there were always dozens, if not hundreds, available. The challenge lies in sifting through them. We found instances ranging from $0.25 to $0.45 per hour, often with significant differences in CPU, RAM, and network attached to the GPU.

Setting up our environment on Vast.ai required a bit more elbow grease. While Docker images are supported, the experience is less guided than Runpod. We often had to debug host-specific issues or deal with less-than-optimally configured base images. Performance varied wildly. Our benchmark ran as fast as 13.5 images per minute on some well-maintained hosts, but dropped to 11.0 images per minute on others, likely due to CPU bottlenecks or slower storage. Averaging it out, we’d say 12.2 images per minute was a realistic expectation for a decently priced instance.

At its lowest advertised price of $0.25/hour, this could mean an astounding $0.0034 per image. However, this often came with caveats: slower CPU, less RAM, or inconsistent uptime. More realistically, for a stable host at $0.30/hour, you’re looking at $0.0041 per image. The cost savings are tangible, but so is the increased management overhead. Vast.ai definitely caters to the more technically inclined, as we’ve noted in our guide for hobbyist ML.

Vultr: Predictability at a Premium

Vultr’s Cloud GPU offering is a more traditional cloud experience. We’ve previously reviewed their A100 offerings, noting their predictability but also their pricing structure in our Vultr Cloud GPU review. For the 4090, availability was the biggest hurdle. There were times when no 4090 instances were available in our preferred regions, requiring us to check back later or deploy to a less optimal location. When available, spinning up an instance was as simple as launching any other cloud VM.

Setup was standard: a clean Ubuntu server, manual NVIDIA driver installation, Docker, and then our Stable Diffusion environment. This process took significantly longer than on Runpod’s pre-configured templates. Performance, once up and running, was rock solid at 12.9 images per minute. No surprises, no dips, just consistent output. However, this consistency comes at a price: $0.50 per hour. This puts the effective cost per image at around $0.0065.

Vultr is a good option if you prioritize stability, predictable performance, and a familiar cloud VM environment, and if your budget allows for the higher hourly rate. It’s less about finding the absolute cheapest rate and more about reliable access to compute when you need it, assuming it’s in stock.

The Numbers: A Quick Overview

Provider	Average Price/Hour	Average Images/Minute (SDXL)	Effective Cost/Image	Availability	Setup Complexity
Runpod	$0.38	12.8	$0.0049	Good	Low
Vast.ai	$0.30 (avg low)	12.2 (variable)	$0.0041 (avg low)	Excellent	Medium-High
Vultr	$0.50	12.9	$0.0065	Spotty	Medium

Note: Vast.ai prices and performance are highly variable. The ‘average low’ here represents what we considered a good balance of cost and performance after some searching.

The Verdict

For most users focused on Stable Diffusion XL inference, the choice comes down to how much you value consistency and ease of use versus absolute rock-bottom pricing. Vast.ai, at its best, offers the lowest cost per image. If you’re comfortable navigating a marketplace, troubleshooting occasional host quirks, and don’t mind a bit of a hunt for a good deal, it’s undeniably powerful for the price-conscious hobbyist or power user. Its highly variable nature, however, isn’t for everyone.

Runpod strikes an excellent balance. Its performance is consistent, setup is streamlined with good templates, and the pricing is competitive. For developers and teams who want reliable access to 4090s without the overhead of constantly monitoring a volatile marketplace, Runpod is probably the sweet spot. It’s where we’d recommend most users start.

Vultr, while predictable and stable, sits at the higher end of the pricing spectrum for 4090 instances. Its appeal is largely for those already in the Vultr ecosystem, or who require a more traditional cloud VM experience with dedicated resources and are less sensitive to cost. The availability issues we encountered also temper enthusiasm somewhat.

Ultimately, if you’re chasing the absolute cheapest Stable Diffusion output, Vast.ai can deliver, but you earn those savings. For a more polished, predictable, and only slightly more expensive experience, Runpod remains our recommendation for serious SDXL work.