Where can you rent an RTX 4090 cheapest, and what can its 24GB actually run?

The cheapest place to rent an RTX 4090 is Vast.ai's marketplace. At 12:20 UTC on Nov 3, 2026 the cheapest on-demand 4090 there was $0.27/hour, with the median offer at $0.46 across 39 listings, and interruptible bids started at $0.21. Runpod Secure Cloud lists the same card at $0.89/hour. Its 24GB runs 8B to 14B language models at 8-bit, 20B to 30B models at 4-bit, Flux image generation and QLoRA fine-tunes up to 14B. It does not run a 70B model. Prices move every hour, so check today's live prices before you rent.

Where is the cheapest RTX 4090 rental right now?

These are single-GPU RTX 4090 prices from our hourly price feed, read at 12:20 UTC on Nov 3, 2026. Vast.ai, Novita AI and Lium figures come from their own market APIs; the Runpod figure is its published list price, which we also confirmed on runpod.io/pricing the same day. "From" is the cheapest listing and "median" the middle one. For the price this hour, see our live cloud-GPU table.

The gap is big: the median Vast 4090 is about half Runpod's list price. What you trade for it is choice and checking. Vast is a marketplace of independent hosts, so each offer has its own reliability score, location, disk and bandwidth price, and the cheapest listing may have a low reliability score or already be taken. Runpod's $0.89 is one fixed price in data centers it lists as "T3/T4". Our Vast.ai review and Runpod vs Vast vs Lambda cover that trade in detail.

Is on-demand or interruptible the better deal?

Vast's docs define on-demand as "Fixed pricing, guaranteed resources (high priority)" and interruptible as "Lowest cost, may be paused (low priority)". On Nov 3 the interruptible floor was $0.21 against $0.27 on-demand, and the medians were $0.40 against $0.46. That is only about 13% saved at the median, so for most jobs on-demand is the better buy. Interruptible makes sense for work you can restart without losing much: a batch of images, or a training run that saves checkpoints.

What does a rented 4090 cost for a real job?

Vast and Runpod both bill GPU time by the second, so you pay for the hours the machine is on. At the Nov 3 prices:

• A 4-hour session costs about $1.84 at the Vast median ($0.46) or $3.56 on Runpod Secure Cloud ($0.89).
• A 40-hour month of evenings costs about $18 at the Vast median or $36 on Runpod.

Storage and bandwidth come on top. Runpod lists container disk and running volume disk at $0.10/GB a month, and an idle volume at $0.20/GB a month. Vast's docs warn that storage is "billed continuously while your instance exists, regardless of running state", and data transfer is charged per byte at each host's rate.

What can an RTX 4090's 24GB actually run?

NVIDIA lists the RTX 4090 with 24 GB of GDDR6X memory. The table below uses our VRAM engine, the same one behind the VRAM calculator: weights at the stated quantization, plus the KV cache at 16-bit and about 0.8 GB of overhead. A total under about 21.6 GB (90% of the card) leaves comfortable headroom.

A few more fits, from the model makers' own pages:

• gpt-oss-20b: OpenAI's model card says it runs "within 16GB of memory", so it fits a 4090 with room for context. See can you run gpt-oss-20b locally.
• Flux image generation: Black Forest Labs calls FLUX.1 [dev] "a 12 billion parameter rectified flow transformer". At 16-bit, the weights alone are about 24 GB (12 billion × 2 bytes), which fills the card, so on a 4090 people run an 8-bit (FP8) version. Note FLUX.1 [dev] ships under a non-commercial license. More in cloud GPUs for ComfyUI.
• Fine-tuning: Unsloth's requirements table puts the minimum for QLoRA (4-bit) at 6 GB for an 8B model, 8.5 GB for 14B and 26 GB for 32B, and LoRA (16-bit) at 22 GB for 8B. So a 4090 handles QLoRA up to about 14B and 16-bit LoRA of an 8B model, but not a 32B fine-tune. Our LoRA fine-tuning guide walks through it.

For any other model, plug it into the VRAM calculator or the cost-to-run page for Qwen3 32B, which matches each model to the cheapest live card that fits.

When is a 4090 the wrong card to rent?

• When 24GB is enough but speed isn't critical: an RTX 3090 also has 24GB and was far cheaper on Nov 3: from $0.09/hr on Vast (median $0.16) and $0.50/hr on Runpod. The 4090 is faster, but for a model that fits, the 3090 often costs less per job. See RTX 4090 vs RTX 3090 for local AI.
• When you need a bit more memory: the 32GB RTX 5090 rented from $0.46/hr on Vast (median $0.65) and $1.19/hr on Runpod on Nov 3, enough for a 32B model at 4-bit with real context.
• When you need 70B or full fine-tuning: step up to an 80GB card; see the cheapest A100 and cheapest H100 rentals.

The short version: rent the 4090 on Vast.ai when price matters and you'll pick a reliable host, on Runpod when you want one fixed price, and check your model fits with the VRAM calculator first. Live prices for every card are on our cloud-GPU compare page, and the cloud-GPU pricing pillar covers the big providers.