RunPod vs Vast.ai for Running Open-Weight LLMs (2026): Prices, EU Regions, Setup and Which to Pick

Pick RunPod if you want a predictable menu, EU datacentres, an ISO 27001 certificate and a ready OpenAI-compatible endpoint with an API key. Pick Vast.ai if you want the lowest hourly price and are willing to choose a host yourself. On 20 October 2026 an RTX 4090 cost $0.89/h on RunPod Secure Cloud, $0.34/h on RunPod Community Cloud, and about $0.50/h (median single-GPU offer) on Vast.ai. For an afternoon with a 70B–120B model, either costs roughly $6–20. For a 24/7 private endpoint, the gap grows to hundreds of dollars a month.

Not sure which GPU your model needs? Run it through the analyzer first, then rent only what it says.

RunPod and Vast.ai are both referral partners of this site; details are in the referral section below.

The verdict in one table

Every price below was checked on 2026-10-20.

Vast prices are a live-market sample, not a list price. Sources are in the sections below.

How the two business models differ

RunPod is a cloud provider with two tiers. Secure Cloud "operates in T3/T4 data centers" for production workloads. Community Cloud "connects individual compute providers to users through a vetted, secure peer-to-peer system" (RunPod docs). Prices are fixed per GPU and tier.

You rent either a Pod (a GPU machine, billed per second, with SSH and ports) or Serverless workers. Flex workers scale to zero and bill per second while running, including a default 5-second idle timeout (RunPod Serverless pricing). Savings plans prepay 3 or 6 months for a discount and are non-refundable (RunPod Pod pricing).

RunPod's Pod pricing docs list only on-demand and savings plans. Its pricing page shows no spot or interruptible Pods, and its public API returned no spot or bid price for any GPU on 20 Oct.

Vast.ai is a marketplace. "A host is anyone who lists GPU hardware on Vast," from one gaming PC to large datacentres, and each sets its own price (Vast docs). Three things decide what you get:

• Reliability score. "A measure of a machine's historical uptime and health." New machines start at 60% (Vast docs).
• Datacentre (Secure Cloud) status. Vast's host docs require an active ISO/IEC 27001 certificate, a registered business, ID checks and at least 5 GPU servers. They get a blue datacentre label (Vast datacenter status). Vast's compliance page is looser: it calls such certificates "encouraged" but "not strictly required" (Vast compliance (vast.ai/compliance)).
• Rental type. On-demand is fixed-price with "data persistence even when stopped". Interruptible is bid-based and "may be paused if outbid or if on-demand requested", with data kept during the pause. Reserved prepays for up to 50% off (Vast instance types).

The practical difference: on RunPod you choose a GPU. On Vast you choose a GPU and a landlord.

Prices, checked 2026-10-20

RunPod numbers come from the RunPod pricing page (runpod.io/pricing) (marked "Updated October 8, 2026") and RunPod's public API on 20 Oct. I cross-checked them against getdeploying.com, which updated its RunPod data the same day. Vast has no list price. Its numbers are my snapshot of Vast's public offer API on 20 Oct: single-GPU, rentable offers with the card's standard memory, at most 64 per query. Treat Vast medians as indicative; they move every hour.

All prices are USD per GPU-hour, on-demand unless stated.

getdeploying showed several RunPod Community SKUs out of stock on 20 Oct, including the $0.34 RTX 4090. The $0.16 Vast RTX 4090 floor was a single deverified host. Vast had no single-GPU H100 PCIe or B200 offers, and its H100 SXM, H200 and A100 figures rest on very few hosts.

RunPod Serverless flex rates per hour: RTX 4090 $1.10, RTX 5090 $1.58, 48 GB class $1.75, A100 $2.72, H100 $4.79, H200 $5.93, B200 $8.64 (RunPod pricing (runpod.io/pricing), checked 2026-10-20).

Storage, transfer and billing rules

Sources: RunPod billing, RunPod Pod pricing, Vast pricing page (vast.ai/pricing), Vast billing docs. Vast storage and bandwidth medians are from my offer snapshot.

On both, stopping an instance stops the GPU bill, not the storage bill.

Which one for which job

Security, privacy and EU data residency

RunPod's compliance page lists ISO/IEC 27001:2022 certificate RUN-ISMS-20260827, valid 27 August 2026 to 26 August 2029, plus a SOC 2 Type II report and SOC 3 (RunPod compliance (runpod.io/legal/compliance)). Read the small print: the certificate "covers the management system, not individual products or services," and coverage "can vary by workload, region, provider, and deployment model." GDPR and HIPAA are described as programmes, not certifications.

RunPod's EU regions include EU-CZ-1, EU-FR-1, EU-NL-1, EU-RO-1, EU-SE-1, EUR-NO-1/2 and EUR-IS-1 to 5 (getdeploying; EU-RO-1 also in RunPod docs).

Vast.ai claims SOC 2 Type 2 and SOC 3, with the report available under NDA (Vast compliance (vast.ai/compliance)). It does not claim ISO 27001 for itself. HIPAA support is limited to its Secure Cloud tier. The page says nothing about GDPR or EU regions.

The bigger question on Vast is the host. Instances "are almost always Docker containers" on someone else's machine (Vast docs). A platform audit does not audit each host. A third-party review puts it bluntly: SOC 2 "covers Vast.ai's platform operations, not individual hosts" (fast.io, 25 June 2026, a commercial review; single source).

Location varies too. In my 20 Oct snapshot, 24 of 63 single RTX 4090 offers were in EU countries, 15 in the US and 3 in Russia. Filter by country and datacentre label if residency matters. My rule on both platforms: no personal data on Community Cloud or non-datacentre hosts, and assume every open port is public until you add authentication.

Setup walkthrough: RunPod (vLLM, connect from your laptop)

The cleanest RunPod path is the vLLM worker on Serverless, because it gives you an API key out of the box. You can create a RunPod account here.

This page contains affiliate links. Rent GPUs on RunPod →

• Open the vLLM worker in the RunPod Hub and click Deploy (RunPod vLLM docs).
• Enter the Hugging Face model ID. Add your HF token if the model is gated.
• Under Advanced, set Max Model Length to the context you need, for example 8192.
• Pick a GPU with enough memory. For example, Llama 3.1 8B at F16 needs 17.9 GB at 8K context.
• Create the endpoint and copy the Endpoint ID.
• Create a RunPod API key in the console.

Then call it from your laptop with any OpenAI client. The base URL format comes from RunPod's docs:

from openai import OpenAI

client = OpenAI(
    api_key="RUNPOD_API_KEY",
    base_url="https://api.runpod.ai/v2/ENDPOINT_ID/openai/v1",
)
r = client.chat.completions.create(
    model="YOUR_MODEL_ID",
    messages=[{"role": "user", "content": "Hello"}],
)
print(r.choices[0].message.content)

The same base URL works as an OpenAI connection in Open WebUI (my guide).

Expect cold starts. In RunPod's own test, a 32B FP8 model on 2x H200 took about 5 min 24 s to cold-start on defaults, and about 1 min 31 s after tuning (vendor-reported, RunPod blog (runpod.io/blog/cut-vllm-cold-starts-runpod-serverless), updated 13 Sep 2026). Users report 2–3 minute loads for 8B models (community-reported, GitHub worker-vllm #111, undated, possibly stale).

On a Pod instead. If you run vLLM or Ollama on a Pod, know that the HTTP proxy URL (https://POD_ID-PORT.proxy.runpod.net) is public and unauthenticated: "Your service becomes publicly accessible" (RunPod docs). Cloudflare also cuts requests at 100 seconds. Either start vLLM with --api-key, or skip the proxy and use an SSH tunnel:

# on the pod
vllm serve YOUR_MODEL_ID --port 8000 --api-key "$(openssl rand -hex 24)"
# on your laptop (IP and port from Connect → Direct TCP Ports)
ssh -N -L 8000:localhost:8000 root@POD_IP -p SSH_PORT -i ~/.ssh/id_ed25519

Your laptop now talks to http://localhost:8000/v1. For Ollama, forward port 11434 the same way.

Setup walkthrough: Vast.ai (vLLM or Ollama, connect from your laptop)

Vast's templates put apps behind its Instance Portal, which adds a token to every link (Vast docs). You can create a Vast.ai account here.

Rent GPUs on Vast.ai →

• Go to Templates and pick vLLM, or Open Webui (Ollama) if you want a chat UI.
• For vLLM, edit the template: set VLLM_MODEL to a Hugging Face repo ID and VLLM_ARGS to your vllm serve flags (Vast vLLM docs).
• Filter offers: enough VRAM (Llama 3.3 70B at Q4_K_M needs 45.7 GB at 8K context), reliability above 95%, and your country or the datacentre label.
• Set the disk slider above the model size. Vast's Ollama guide uses 50 GB for a 70B model on 2x RTX 4090 (Vast Ollama guide).
• Click Rent and watch the logs. The first start downloads the model.
• Click Open to reach the portal. Copy the OPEN_BUTTON_TOKEN from any portal link.

Then call the API. Find the external port mapped to 8000 (vLLM) or 11434 (Ollama) under the instance's IP button:

curl -k https://INSTANCE_IP:EXTERNAL_PORT/v1/chat/completions \
  -H "Authorization: Bearer OPEN_BUTTON_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"model":"YOUR_MODEL_ID","messages":[{"role":"user","content":"Hello"}]}'

The -k flag is there because Vast uses a self-signed certificate, per its own docs. That means your client does not verify who it is talking to. I prefer an SSH tunnel, using the SSH command shown on the instance card:

ssh -N -L 8000:localhost:8000 root@INSTANCE_IP -p SSH_PORT

Point your laptop's tools at http://localhost:8000/v1.

Fine-tuning on each

RunPod's fine-tuning page is "powered by Axolotl" (RunPod docs). You pick a Hugging Face model and dataset, deploy a fine-tuning Pod, edit /workspace/fine-tuning/config.yaml and run axolotl train config.yaml. You can test the adapter with vllm serve ... --enable-lora.

Vast offers an Unsloth Studio template with a web UI (Vast docs). Its guide lists QLoRA 4-bit at about 5 GB for a 7B model, LoRA 16-bit at 19 GB, and 70B QLoRA at 41 GB (vendor-reported). It exports a merged model, a LoRA adapter of about 100 MB, or GGUF for Ollama and LM Studio. On Vast interruptible instances, checkpoint to external storage: a paused instance keeps its data but does not train.

Worked cost examples (my arithmetic)

My sums from the prices above, checked 2026-10-20. GPU sizing is my arithmetic; the analyzer is the authority. Short jobs include one hour for download and setup.

1. Try a 70B–120B model for an afternoon (4 h + 1 h setup = 5 h). On an 80 GB card such as an A100 80 GB or H100, a 120B MoE such as gpt-oss-120b at MXFP4 needs 63.2 GB, fits 80 GB at 8K context. A 70B dense model such as Llama 3.3 70B at Q4_K_M needs 45.7 GB, fits 80 GB.

A single 24 GB home GPU cannot hold either model, so renting wins this job outright.

2. A 24/7 private endpoint for a month (730 h, one RTX 4090, 100 GB disk). 3. QLoRA fine-tune of an 8B model (assume 3 h training + 1 h setup = 4 h on one RTX 4090). The 3-hour run time is my assumption; it depends on dataset size and settings. For memory, see the vendor-reported QLoRA figures in the fine-tuning section above. Against buying a 24 GB GPU. The EU average household electricity price was €0.2896/kWh in the second half of 2025, from €0.1082 in Hungary to €0.4042 in Ireland (Eurostat, data from April 2026). I assume a 4090-class desktop draws 0.5 kW under load and 0.15 kW on average as a mostly idle endpoint. Those draw figures are my assumptions, not measurements.

• Afternoon test: 0.5 kW × 5 h = 2.5 kWh ≈ €0.72. But the model does not fit.
• QLoRA run: 0.5 kW × 4 h = 2 kWh ≈ €0.58.
• 24/7 endpoint: 0.15 kW × 730 h ≈ 110 kWh ≈ €32 a month; at full load, 365 kWh ≈ €106.

Then add the card itself (an RTX 4090; check its current price below). Break-even in months = card price ÷ (monthly cloud bill − monthly electricity). Against a ~$385 Vast month, a home card pays back fast if you truly run it 24/7. Against a $66 serverless month, it may never pay back. Put your own numbers into the cost calculator, and read local vs cloud cost for the long version.

Referral programmes: what you get

• RunPod. A referred European customer gets a fixed $5 after loading $10. Outside Europe, the bonus is a random $5–$500, and about 96% get $10 or less. You must sign up with Google SSO and be a first-time user. I earn 3% of your Pod spend and 5% of your Serverless spend for six months, as credit (RunPod referrals).
• Vast.ai. The referral docs mention no bonus for you. I earn 3% of your spend for the lifetime of your account (Vast referral programme).

Neither programme changes your price.

My take

For most readers here, I would start on RunPod. You do not judge strangers' machines, there are no transfer fees, and the EU regions have names. The vLLM Serverless worker is the fastest route to an authenticated endpoint, if you accept minute-long cold starts.

I would move to Vast.ai once price matters more than convenience. Its median RTX 4090 and A100 offers were well below RunPod Secure, and interruptible instances suit experiments that checkpoint. The cost is diligence: check reliability, country and datacentre label every time. And if you will run one model 24/7 for a year, do the break-even sum first. A home GPU can win that job.

Get started

• Cloud GPU rental guide: how GPU rental works across providers.
• Local vs cloud cost: when renting beats buying.
• Cost calculator: plug in your own hours and prices.
• Analyzer: find the smallest GPU that fits your model before you rent.
• Ready to rent: RunPod or Vast.ai (affiliate links below).

Rent GPUs on RunPod → Rent GPUs on Vast.ai →

Frequently asked questions