RTX 4090 Price: New, Used, and Cloud Rental Costs

The RTX 4090 launched at a $1,599 MSRP, but it's out of production and now sells well above that. Used cards have recently sold for about $2,300 on eBay, while sellers ask around $3,000. New cards are leftover stock and rarely sit near MSRP. Renting one in the cloud costs about $0.31 to $0.89 an hour, and Beam's RTX 4090 machine is $0.44 an hour.

RTX 4090 price at a glance

Why the RTX 4090 costs more than its MSRP

Board partners reported that RTX 4090 production wound down in late 2024 to make room for the RTX 50 series. Nvidia never announced an end date, but no new supply is coming, so every card for sale is either leftover retail stock or used. Nvidia's own product page no longer lists a price.

Demand didn't fall with supply. The 4090's 24 GB of VRAM makes it one of the few consumer cards that can run image models and mid-size LLMs locally, so AI users compete with gamers for the same used cards. The RTX 5090 hasn't relieved the pressure either: it launched at $1,999, and its retail stock is scarce and priced far above that MSRP.

New vs used RTX 4090 prices

The gap between sold prices and asking prices matters more than usual for this card. A tracker of completed eBay sales put the recent used value at $2,271, down from a 90-day average of $2,435. Active listings tell a different story: a cross-marketplace index puts the median asking price at about $3,000, with most listings between $2,000 and $4,500. Sellers ask high; buyers mostly pay in the low $2,000s.

If you buy used, check three things before paying:

• The power connector. The 4090 uses the 16-pin 12VHPWR connector, which has a history of melting when it isn't fully seated. Ask for photos of the connector and the card's own socket.
• Its past workload. Cards that ran AI training or mining around the clock can have worn fans and dried thermal paste. Neither is fatal, but price it in.
• Warranty transfer. Some board partners tie the warranty to the original buyer. Check the brand's policy and ask for the receipt.

Where to buy an RTX 4090

New stock, where it exists, sits with retailers such as Amazon, Best Buy, Newegg, and B&H, usually as partner models from ASUS, MSI, Gigabyte, and others. Most cards changing hands today are used, and eBay carries the bulk of those listings, followed by Facebook Marketplace and OfferUp. eBay's sold-listings filter shows what buyers actually paid, which is a better guide than the asking price on any single listing.

RTX 4090 cloud rental prices by provider

Renting an RTX 4090 costs about $0.31 to $0.89 per GPU-hour on demand. Marketplaces hold the floor, and managed clouds charge more for steadier capacity. Beam's machine is $0.44 an hour with vCPU, RAM, and NVMe included.

Each figure is a single-GPU, on-demand list price. Vast.ai prices change by the hour, because each host sets its own rate. The sources:

• Vast.ai's RTX 4090 page lists live marketplace offers.
• RunPod's pricing page has both cloud tiers.
• Salad's pricing page lists each priority tier.
• Beam's pricing page has the machine and serverless rates.

Cheapest ways to rent an RTX 4090

Vast.ai's marketplace and RunPod Community Cloud undercut Beam's $0.44, and Salad's lower priority tiers go down to $0.16 if your job can be interrupted. Vast.ai hosts set their own prices, so check the host's reliability score along with the rate. Salad runs on consumer PCs: a High priority node isn't preempted by other workloads, but it can still disconnect, which suits retry-safe inference better than long training runs.

Beam's machine price buys a fixed rate with NVMe included. The serverless option costs more per GPU-hour, so it only saves money when the GPU would otherwise sit idle for long stretches.

Should you buy an RTX 4090 or rent one?

Rent unless you'll keep the card busy for most of a year. The break-even point is simple to estimate:

break-even hours = local build cost / hourly cloud price

Take a used 4090 at $2,300 plus about $1,000 for the rest of the machine. Against Beam's $0.44 machine, that's 7,500 GPU-hours, or about 10 months of round-the-clock use. Electricity pushes it further. A 4090 is rated at 450 W, so a full system under load draws around 600 W, which is about 432 kWh a month running nonstop. At $0.18 per kWh, that adds roughly $78 a month, and break-even moves out to about 14 months.

Usage patterns change the answer more than the price does. At two hours a day, the same $3,300 build takes over ten years to pay off against a $0.44 rental. Buying makes sense for steady, heavy use where you want the hardware in the room. Renting wins for bursty work, experiments, and anything that sometimes needs more than one GPU.

Running a model on a rented RTX 4090 with Beam

On Beam, a serverless GPU like the RTX 4090 runs your Python function as an HTTPS endpoint, bills by the millisecond while it works, and scales to zero when traffic stops. This endpoint loads an image model once per container and reuses it for every request:

from beam import endpoint, Image

def load_model():
    import torch
    from diffusers import AutoPipelineForText2Image
    pipe = AutoPipelineForText2Image.from_pretrained(
        "stabilityai/sdxl-turbo", torch_dtype=torch.float16
    )
    return pipe.to("cuda")

@endpoint(
    gpu="RTX4090",
    cpu=2,
    memory="16Gi",
    image=Image(python_packages=["diffusers", "transformers", "accelerate", "torch"]),
    on_start=load_model,       # runs once when the container boots
    keep_warm_seconds=60,      # stay up a minute after the last request
)
def generate(context, prompt: str):
    import base64, io
    pipe = context.on_start_value
    image = pipe(prompt, num_inference_steps=1, guidance_scale=0.0).images[0]
    buf = io.BytesIO()
    image.save(buf, format="PNG")
    return {"image": base64.b64encode(buf.getvalue()).decode()}

Deploy it with beam deploy app.py:generate. With 2 CPU cores and 16 GiB of RAM, this endpoint costs about $1.77 per running hour in total, and nothing while it's scaled to zero. If your traffic is steady, the $0.44 machine is the cheaper way to run the same card. Expect a cold start on the serverless GPU after idle stretches; keep_warm_seconds trades a little idle cost for fewer of them.

Is the RTX 4090 still good for AI workloads?

Yes, for models that fit in 24 GB of VRAM. The 4090 has 16,384 CUDA cores and handles Stable Diffusion, ComfyUI, smaller LLM inference, embeddings, and LoRA fine-tuning well. It's a strong pick for experimentation because cloud rates for it are among the lowest of any modern GPU.

Memory is the limit. If a model needs more than 24 GB, the RTX 5090 has 32 GB and costs $0.72 an hour as a Beam machine. Larger models need data center cards; see RTX PRO 6000 pricing, H100 pricing, or the B200, and compare serverless GPU providers before buying hardware.

What else affects the cost of a cloud RTX 4090

The hourly rate is only part of the bill. These factors decide what you actually pay:

1. Cold start time. On serverless platforms, a cold start happens when a new container is loaded onto a GPU: the image is pulled, the container starts, and libraries initialize. You pay for that time, and your users wait through it.
2. Model loading time. Large weights take time to load onto the GPU. Load them once per container, not once per request, as the snippet above does with on_start.
3. Autoscaling. If traffic spikes, your app needs to scale out to more GPUs and back down again, or you pay for idle capacity.
4. Caching model weights. Downloading weights on every boot is slow and sometimes billed as egress. Cache them on storage close to the GPU.
5. CPU and RAM. Some providers bundle them into the GPU price, and others bill them separately. Compare the full container cost, not just the GPU line.

RTX 4090 price FAQs

How much is an RTX 4090?

The RTX 4090 launched at $1,599, but used cards have recently sold for about $2,300 on eBay and median asking prices are around $3,000. New cards are leftover stock and usually cost more than MSRP.

Is the RTX 4090 discontinued?

Production wound down in late 2024, according to board partners, ahead of the RTX 50 series. Nvidia never announced an end date, but no new supply is being made, which is why prices rose after launch instead of falling.

What was the RTX 4090 MSRP?

The Founders Edition launched at $1,599 in 2022.

How much does it cost to rent an RTX 4090?

About $0.31 to $0.89 per GPU-hour on demand. Vast.ai's marketplace starts around $0.31, RunPod Community Cloud is $0.34, and Beam's machine is $0.44 with vCPU, RAM, and NVMe included.

Is it cheaper to buy or rent an RTX 4090?

Renting is cheaper unless you run the card most of the day for a year or more. A $3,300 build pays off after about 10 months of nonstop use against a $0.44 rental, or about 14 months once electricity is counted.

RTX 4090 vs RTX 5090: which should I rent?

Rent the 4090 if your model fits in 24 GB. Pay for the 5090 when you need 32 GB. On Beam, the machines cost $0.44 and $0.72 an hour.

Run an RTX 4090 on Beam

Beam gives you RTX 4090 machines at $0.44 an hour and serverless RTX 4090 endpoints that scale to zero. Sign up for Beam and deploy your first GPU endpoint in a few minutes.