Cheapest Cloud GPU for AI in 2026: RunPod vs Vast.ai vs Lambda Prices for RTX 5090, H100, H200 and B200

With an RTX 5090 selling for close to $7,000, renting has never looked better. You can rent the same card for well under a dollar an hour, or an 80 GB H100 for $2-$4 an hour, and stop paying the second your job ends. This guide compares the cheapest cloud GPUs for AI in November 2026 across RunPod, Vast.ai, Lambda and the big clouds, explains the hidden costs (storage, egress, minimum billing), and shows when renting beats buying for local LLMs, ComfyUI image and video work, and fine-tuning.

Cloud GPU Price Table (November 2026)

Per-GPU hourly prices. RunPod and Lambda prices are from their official pricing pages, checked on 5 November 2026. Vast.ai is a marketplace with no fixed price, so I give typical ranges from Spheron’s October 2026 analysis (Spheron is a competitor, so check live listings).

A dash means the provider does not list that GPU on its public price page. Prices change with demand, sometimes daily.

Which GPU Should You Rent?

Rent the cheapest GPU with enough memory for your job. Paying for an H100 to run a model that fits on a 4090 is the most common waste I see.

Not sure how much memory your job needs? Use my VRAM requirements guide. To understand the datacenter cards, read H100 vs H200 vs B200 vs B300 vs Rubin.

RunPod

RunPod is where I would start most people. It has two tiers: Secure Cloud (Tier 3/4 datacenters) and Community Cloud (vetted third-party hosts, cheaper). Billing is per second, there are ready-made templates for ComfyUI, Ollama, vLLM, Text Generation WebUI and Jupyter, and network volumes let you keep your models between sessions.

• Good for: ComfyUI sessions, short LLM tests, LoRA training, serverless endpoints.
• Watch out for: network volume storage is billed while it exists, even when no pod is running. Delete volumes you no longer need.
• Range: from $0.27/hr (RTX A5000 24 GB) to $8.99/hr (B300 288 GB).

Vast.ai

Vast.ai is a marketplace where individual hosts and datacenters list their machines and set their own prices. That makes it the cheapest place to rent consumer GPUs, but quality varies.

• Unverified vs verified hosts. Unverified hosts are cheapest but have no formal vetting. Datacenter-verified hosts cost more and are more reliable.
• Interruptible vs on-demand. Interruptible instances can be stopped at any time, which is fine for batch jobs that save checkpoints, not for a live demo.
• Check the details of each offer: DLPerf score, internet speed (important when downloading 20-60 GB models), disk price and egress rules.
• Storage is billed while paused, at roughly $0.10-$0.15 per GB per month according to Spheron, and many listings bill in full hours.

For bulk image or video generation and cheap LLM experiments, Vast.ai is hard to beat. For anything sensitive, remember you are running on someone else’s machine; do not upload private data to unverified hosts.

Lambda

Lambda focuses on datacenter GPUs only: on-demand instances at $3.99/hr for an 80 GB H100, $2.79 for an 80 GB A100, $1.99 for a 40 GB A100 and $6.69 per GPU-hour for a 180 GB B200 (plus tax). It also sells 1-Click Clusters of 16 to 256+ B200s on 2-week to 1-year terms. Instances come with Lambda Stack (PyTorch, CUDA, drivers) preinstalled. Lambda is a good choice for training runs and teams that want predictable, enterprise-grade hardware; it is not the cheapest for hobby inference.

Other Providers and the Big Clouds

Spheron’s comparison (May-October 2026 rates) shows how wide the spread is for the same 80 GB H100 SXM:

The hyperscalers make sense when you already run your business on them, need compliance certifications or get credits. For individuals they cost two to five times more than specialist GPU clouds. Their spot prices (for example about $3.69-$3.83/hr for H100 on Google Cloud and AWS) are closer, but spot instances can be reclaimed.

Hidden Costs to Watch

• Storage. Persistent disks and volumes are billed whether or not a GPU is attached.
• Egress. Downloading big outputs (video, checkpoints) can cost money on some hosts and clouds.
• Minimum billing. Hourly rounding makes a 5-minute test cost a full hour.
• Idle time. The most expensive GPU is the one you forgot to stop. Set an auto-stop or a budget alert.
• Model downloads. A 60 GB model on a slow host can take 20+ minutes of paid time. Pick hosts with fast internet or keep models on a network volume.

Rent vs Buy: The Break-Even Maths

Divide the purchase price by the hourly rental price to get break-even hours (ignoring electricity, which makes buying even less attractive):

Buying still makes sense if you use a GPU for many hours every day, need privacy (your data never leaves your machine), want zero setup friction, or also game on the card. For occasional big jobs, renting wins easily at today’s prices. I explain why cards cost so much in GPU prices in 2026, and which cards are still worth buying in the best GPU for local LLMs.

Tips to Pay Less

1. Start on the smallest GPU that fits and scale up only if it is too slow.
2. Use quantized models (GGUF Q4-Q6, FP8) so a 24-32 GB card is enough. See the GGUF quantization guide.
3. Keep models on a network volume so you do not re-download them each session.
4. Use interruptible or Community instances for batch jobs that can resume.
5. Shut down pods when you are done and delete storage you do not need.
6. For ComfyUI, prepare and test your workflow locally on a small model, then rent only for the final high-resolution or video runs.

Your First Rental: ComfyUI or Ollama on RunPod (Step by Step)

1. Create an account at runpod.io and add a small amount of credit, for example $10. Prepaid credit is a natural spending cap.
2. Create a network volume (optional but recommended) in the same region you will rent in, sized for your models: 50-100 GB for ComfyUI image work, 100-200 GB for video models. This keeps your downloads between sessions.
3. Deploy a pod: choose the GPU (an RTX 4090 or 5090 is enough for most ComfyUI and 8B-32B LLM work), pick a ComfyUI or Ollama template, and attach the network volume.
4. Open the pod’s web port from the Connect button. ComfyUI usually runs on port 8188; Ollama’s API listens on 11434.
5. Download your models into the volume (for ComfyUI, the ComfyUI/models/ folders listed in my ComfyUI model folders cheat sheet). Use a Hugging Face token for gated models.
6. Do your work, download outputs to your PC, then stop or terminate the pod. A stopped pod with a container disk can still cost a little for storage; a terminated pod costs nothing except the network volume.

The same flow works on Vast.ai: pick an offer, choose a ComfyUI or Ollama image, open the port, and destroy the instance when you finish.

What a Typical Session Costs

Compare that with buying: even a heavy hobby user renting an RTX 5090 for 10 hours a week spends about $50 a month, far below the cost of the card at today’s prices.

Security and Privacy Tips

• Never leave ComfyUI or Jupyter open to the internet without a password. Use the provider’s authenticated proxy or an SSH tunnel.
• Avoid unverified marketplace hosts for private or client data. Use secure or datacenter-verified tiers instead.
• Use scoped tokens. Create a read-only Hugging Face token for the pod and revoke it later.
• Be careful with random custom nodes and workflows. They run code on the machine. My must-have ComfyUI custom nodes list sticks to well-known packs.

Common Problems and Fixes

• “No GPUs available” in my region: try another region or the Community tier, or pick a similar card (a 4090 instead of a 5090). Your network volume is tied to its region, so choose a busy region with many GPUs when you create it.
• Model downloads are very slow: check the host’s internet speed before renting, and download straight from Hugging Face on the pod rather than uploading from home.
• CUDA errors on RTX 50 or Blackwell cards: use a template with a recent PyTorch build (CUDA 12.8 or newer). Old images do not support these GPUs. See my ComfyUI on RTX 50 guide.
• My files disappeared: anything outside the network volume or persistent disk is wiped when the pod is terminated. Always save models and outputs to the volume.
• The bill kept running: stopping is not always terminating. Check the dashboard for running pods and idle volumes at the end of every session.

Frequently Asked Questions

What is the cheapest cloud GPU for AI?

On Vast.ai, RTX 3090 and 4090 instances from about $0.35 per hour are usually the cheapest useful option. On RunPod, the RTX A5000 at $0.27 and RTX 3090 at $0.50 per hour are the cheapest 24 GB GPUs.

How much does it cost to rent an H100 per hour?

About $1.80-$2.50 on Vast.ai verified hosts, $2.69-$3.99 on RunPod and $3.99 on Lambda in November 2026. AWS, Google Cloud and Azure on-demand prices are about $6.88-$12.29.

Is RunPod or Vast.ai cheaper?

Vast.ai is usually cheaper because hosts compete on price, but reliability varies. RunPod costs a bit more and is more predictable, with per-second billing.

Can I run ComfyUI on a cloud GPU?

Yes. RunPod and Vast.ai both have ready-made ComfyUI templates. Use a network volume for your models and custom nodes so you can stop and restart cheaply.

Is it cheaper to rent or buy a GPU in 2026?

For most people, renting is cheaper. An RTX 5090 at about $6,900 needs roughly 6,000-7,800 hours of rental to break even, which is years of daily use.

Final Thoughts

For most AI hobbyists in late 2026, the cheapest setup is a modest local GPU for everyday work plus a rented RTX 5090, RTX PRO 6000 or H100 for the big jobs. Start with Vast.ai if you want the lowest price, RunPod if you want simplicity, and Lambda if you need serious datacenter hardware. Whatever you choose, stop the instance when you are done.

Sources

• RunPod GPU cloud pricing (checked 5 Nov 2026)
• Lambda pricing (checked 5 Nov 2026)
• Spheron: Vast.ai pricing 2026 (updated 19 Oct 2026)
• Spheron: GPU cloud pricing comparison 2026 (updated 23 Oct 2026)
• Tech Insider: RTX 5090 street price (Nov 2026)
• OpenClaw: used RTX 3090 prices