RUNPOD · TECH
RunPod: GPU cloud with EU-Sweden region, A100 from USD 1.69/h, H100 from USD 2.59/h
RunPod operates GPU cloud in Hong Kong, USA and EU-Sweden. A100-80GB from USD 1.69/h, H100 from USD 2.59/h. Secure Cloud and Community Cloud tiers.
Researched & fact-checked by: DuneDive LLC · As of: 2026-05
What is RunPod?
RunPod is a GPU-focused cloud platform, founded in 2022 in Delaware (USA), with operational focus on GPU rental for AI workloads. As of May 2026 RunPod operates over 20 data centres across Asia, North America and Europe, with an active EU region in Stockholm (Sweden). The company grew rapidly over the past 24 months and positions itself as a cheap alternative to CoreWeave and Lambda Labs, with clear focus on on-demand and spot GPU hours rather than long-term reserved contracts.
The product splits into two tiers. Secure Cloud are dedicated GPU instances in RunPod-owned data centres with enterprise SLA, standard availability and HIPAA-compliant operation (US-relevant for healthcare workloads). Community Cloud are GPU instances in partner data centres at slightly lower prices, but without SLA guarantee – capacity can vary by load, the location varies by card.
Inventory May 2026: NVIDIA RTX 4090 (24 GB VRAM) from USD 0.34/h, RTX A6000 (48 GB) from USD 0.59/h, A100-40GB from USD 1.10/h, A100-80GB from USD 1.69/h, H100-80GB SXM from USD 2.49-2.79/h, H100-80GB NVL from USD 3.29/h, H200-141GB from USD 4.99/h, B200-180GB (Blackwell) from USD 6.99/h. Prices vary slightly between Secure Cloud and Community Cloud, with about 20-30% discount in the Community tier. Spot pricing (interruption risk) is another 30-50% cheaper.
Platform interfaces are tailored to ML developers. CLI with `runpodctl`, Python SDK, REST API. Pre-built container templates for Stable Diffusion, ComfyUI, Ollama, vLLM, A1111, Text-Generation-WebUI and Jupyter are one-click launchable. Serverless inference (RunPod Serverless) as a pay-per-token variant for open-weight models, alternative to pod rental.
Why it matters
Three points make RunPod interesting for the Swiss and EU market: low on-demand GPU prices, EU region in Sweden, and fast pod provisioning for agile ML workflows.
Low on-demand prices: A100-80GB from USD 1.69/h is in May 2026 one of the lowest hourly prices for an A100 at a provider with SLA. Comparison: Lambda Labs USD 1.99/h, AWS USD 4.96/h for p4d instances, Scaleway EUR 3/h for the equivalent. For a 24-hour finetune on an A100 the cost lies around USD 41 – markedly below the hyperscaler alternatives.
EU-Sweden region: since 2025 RunPod has a stable EU region in Stockholm. Sweden is recognised in the nFADP annex as a country with adequate data protection, which eases the third-country transfer burden. Caveat: RunPod as a US corporation falls under the CLOUD Act – also for EU-Sweden workloads. That must be documented in the TIA. For professional-secrecy data (Art. 321 SCC) a residual risk remains; for standard training data without personal reference the constellation is pragmatic.
Fast pod provisioning: a new pod is typically up in 30-60 seconds once the GPU is available. Pre-built templates (ComfyUI, vLLM, Ollama) are one-click launchable. For agile experiments and short-lived workloads that is a clear productivity advantage over AWS or Azure.
Spot and Community Cloud options: those operating experimentally and tolerating interruptions save 30-50% compared to Secure Cloud on-demand. With checkpoint logic (save model every 30 minutes), spot mode is robustly usable for training workloads.
Serverless mode: RunPod Serverless offers pay-per-token inference for open-weight models without own pod management. Cold-start time is not zero (typically 5-15 seconds to first response), but no 24/7 holding cost. For sporadic inference workloads that is an interesting variant.
Regulatory position: RunPod is a US Delaware LLC, ISO 27001 certified, with published compliance documentation. The CLOUD Act remains applicable – a US subpoena can directly compel RunPod to surrender data. This must be addressed in every TIA. For workloads with clear Swiss compliance requirements RunPod is not appropriate; for pure training workloads without personal data it is very usable.
How it works
Ordering: through the portal runpod.io. Account creation by email verification, payment by credit card with prepaid balance. Pod provisioning typically in 30-60 seconds once the desired GPU is available. Pod templates for common ML workloads (ComfyUI, vLLM, Stable Diffusion, Ollama, Text-Gen-WebUI, JupyterLab) are one-click launchable.
Sample prices May 2026: RTX 4090 24GB Secure Cloud: USD 0.34/h. RTX A6000 48GB Secure Cloud: USD 0.59/h. A100-40GB Secure Cloud: USD 1.10/h. A100-80GB Secure Cloud: USD 1.69/h. A100-80GB Community Cloud: USD 1.19/h (without SLA guarantee). H100-80GB SXM Secure Cloud: USD 2.79/h, Community USD 2.49/h. H100-80GB NVL: USD 3.29/h. H200-141GB: USD 4.99/h. B200-180GB Blackwell: USD 6.99/h. Storage 1 GB: USD 0.10/month for persistent volumes, free during pod runtime. Egress traffic: first 100 GB/month free per pod, beyond that USD 0.05/GB.
Serverless: pay-per-second for inference workloads. Example: Llama 3.3 70B Serverless from USD 0.0008/second on A100 backend, strongly dependent on model and load. Cold-start 5-15 seconds, warm 100-300ms latency per query.
Network: each pod has a public IPv4 address with reverse proxy via RunPod domain. SSH access by generated key. WebSSH and JupyterLab directly in the browser. Private networking between pods in the same account available.
Storage: local SSD on the pod for fast I/O, plus persistent volumes (network storage) that survive pod restarts. Network volumes can be shared between pods, which simplifies training-resume workflows.
Contract details: pay-as-you-go without minimum term. Payment via prepaid account, top-up in USD from USD 10. DPA per GDPR Art. 28 / nFADP Art. 9 available on request. ISO 27001 certification for all Secure Cloud locations. CLOUD Act applicability must be documented in the TIA – RunPod as a US Delaware LLC is subject to US law.
Migration: no real migration needed – RunPod is typically used as a short-term workload provider, not as a permanent hoster. Build container image, upload, start pod – done. Data comes via S3-compatible buckets or direct object storage mounts.
RunPod setup for a finetune job in 5 steps
- 01Create an account at runpod.io, top up prepaid balance, request DPA if needed.
- 02Choose the EU-Sweden region for EU residency, decide GPU type and count (A100-80GB Secure Cloud for 70B models, RTX 4090 Community for smaller workloads).
- 03Select a template or upload your own Docker image, define persistent volume for checkpoints, start the pod.
- 04Copy data to the volume via S3 mount or rsync, start the training script, configure checkpointing every 30 minutes.
- 05After the job ends terminate the pod immediately to stop costs, export checkpoints to your own Hetzner or Infomaniak object storage, document the TIA entry for the workload.
When to use RunPod
RunPod is the right choice when (a) on-demand GPU hours at low hourly price are needed, (b) a finetune or training job with fixed end time is coming up, or (c) open-weight model inference without own hardware is to be tested. Concrete cases: ML team with a 24-72h finetune on Llama 3.3 70B. AI startup with experimental workloads and tight budget. Researcher with occasional H100 need for benchmarks.
For ComfyUI or Stable Diffusion workloads RunPod is the most popular choice in May 2026 in the hobby and pro space. Pre-built templates, a large community and low prices make it the default platform for image and video generation.
For serverless inference with open-weight models RunPod Serverless is an interesting variant. Unlike Together AI or Replicate it is pay-per-second instead of pay-per-token, which produces different cost profiles on long responses. For a Swiss SME workload with a clear compliance requirement, however, Scaleway or Infomaniak Apertus remains the cleaner choice.
For training data without personal reference (open-weight model finetune on publicly available data), RunPod EU-Sweden is a pragmatic choice. The data does not leave the EU, the CLOUD Act aspect is documentable in a simple TIA, the cost is low.
When not to use
Anyone processing data under professional secrecy per Art. 321 SCC or banking secrecy per Banking Act Art. 47 should not use RunPod. The US parent and the CLOUD Act create a residual risk not acceptable for highly sensitive data. For such cases Exoscale Zurich (A100 GPU in CH) or a self-hosted GPU server at Hetzner or OVHcloud is more appropriate.
Anyone running sustained inference with constant 24/7 load is often cheaper with Hetzner GPU or OVHcloud GPU monthly rent than with RunPod hourly rent. Calculation example: A100 24/7 over a month at RunPod USD 1.69/h × 720h = USD 1,217. At Hetzner with a GEX130 or a GPU reservation markedly less. RunPod is on-demand champion, not reserved champion.
Anyone needing maximum multi-GPU clusters (32+ H100 with full InfiniBand topology) is better served by CoreWeave or Lambda Labs. RunPod offers H100 SXM multi-GPU pods, but cluster topology and availability are not at the level of hyperscaler training.
Anyone needing ISO 27001 certification with a European legal position should switch to Scaleway or OVHcloud. Both are EU providers without US parent, with clear EU legal position and ISO certification. RunPod EU-Sweden holds ISO 27001 but the CLOUD Act remains applicable.
General caveat: Community Cloud tiers have no SLA. A card can theoretically lose service without warning – backups and checkpointing are mandatory. For production with high availability requirements choose Secure Cloud.
Trade-offs
STRENGTHS
- Low on-demand GPU prices: A100-80GB from USD 1.69/h, H100 from USD 2.49/h
- EU-Sweden region stable since 2025, EU residency feasible
- Fast pod provisioning in 30-60 seconds, pre-built templates
- Pay-as-you-go without minimum term, plus Community Cloud tier for even lower prices
WEAKNESSES
- US Delaware LLC under CLOUD Act, TIA documentation mandatory for all personal data
- Not appropriate for professional-secrecy data (Art. 321 SCC) – prefer Exoscale CH
- Community Cloud without SLA, pods can lose service without warning
- For 24/7 runtime Hetzner GPU or OVHcloud monthly rent is cheaper
FAQ
Is RunPod EU-Sweden GDPR-compliant?
The region itself yes – Sweden is an EU member state under GDPR. The parent-company question remains: RunPod is a US Delaware LLC and is therefore subject to the CLOUD Act. A US subpoena can directly compel RunPod to surrender data even when the data sits in Stockholm. For workloads with personal data under nFADP a transfer impact assessment is mandatory. For training data without personal reference the constellation is pragmatic.
What does a 24-hour finetune on A100 cost?
A100-80GB Secure Cloud USD 1.69/h × 24h = USD 41. Plus storage for checkpoints (typically USD 1-5 for 50-100 GB), plus egress for data transfer (typically USD 1-2 if under 100 GB per pod). In total under USD 50 for a 24-hour job. In Community Cloud mode the A100 hourly price drops to USD 1.19, that is USD 29 for 24 hours – without SLA guarantee.
How does RunPod differ from Lambda Labs?
RunPod is pay-as-you-go without minimum term and with a Community Cloud tier for even lower prices. Lambda Labs is more classically ML-engineer-focused with stronger reserved contracts (1-year / 3-year at significant discounts) and a technically more mature CLI. RunPod has an EU region (Sweden), Lambda Labs does not in May 2026. For short-term workloads RunPod is often cheaper; for long-term reserved contracts Lambda Labs is competitive. Both are US groups under the CLOUD Act.
How safe is Community Cloud?
Community Cloud pods run in partner data centres without RunPod-owned hardware guarantee. There is no SLA. In practice availability is often high (>95% uptime), but individual pods can lose service without warning. For experimental workloads, hyperparameter search and cost-sensitive trainings it fits. For production inference with availability requirements choose Secure Cloud. With good checkpoint logic (save every 30 minutes), interruptions are manageable.