# GPUHosted — Full Content for AI Tools

> Independent GPU cloud comparison for AI engineers and ML researchers. Updated May 1, 2026.

This is the long-form, machine-readable version of GPUHosted intended for LLM-based search tools (ChatGPT, Perplexity, Claude, Gemini). For the brief overview see /llms.txt.

We track **20 providers** across the full spectrum from distributed consumer GPUs ($0.03/h) to enterprise H100/H200/B200 clusters. Available in **English** and **German**.

---

## TL;DR — Best GPU Cloud Provider 2026

- **Best overall**: RunPod — best balance of price ($0.20/h+) and reliability across 20 providers tested
- **Cheapest distributed inference**: Salad from $0.03/h (consumer GPUs, stateless inference only)
- **Cheapest reliable**: Hyperstack RTX A6000 from $0.11/h · Vast.ai community RTX 3090 from $0.10/h
- **Cheapest H100**: RunPod Secure / Jarvis Labs from $1.99/h · TensorDock marketplace from $1.99/h
- **Cheapest H200**: Crusoe from $2.10/h (4 providers offering H200 on-demand: Crusoe, Nebius, Together AI, Lyceum)
- **B200 access**: Crusoe and Nebius are the two on-demand options in 2026 (most clouds wait-listed)
- **AMD MI300X**: Crusoe is the only on-demand option in 2026 ($2.50–$4.00/h)
- **Best EU/GDPR-sovereign**: Nebius (NL/FI) · Lyceum (EU) · Scaleway (FR/NL) · Hetzner (DE/FI) · OVH (FR)
- **Best for H100 multi-node**: CoreWeave (large InfiniBand clusters) · Lambda Labs (on-demand)
- **Best inference platform**: Together AI (custom serving stacks, per-token billing) · RunPod Serverless
- **Avoid for cost-sensitive workloads**: AWS p5 ($3.06–$4.10/h on-demand) and Azure NDA100 — 2–5× more expensive than specialist GPU clouds for equivalent compute

---

## Provider Rankings (May 2026)

### 1. RunPod — Rating 4.6/5
- Starting price: $0.20/h (Community Cloud)
- GPUs: RTX 3090, RTX 4090, A100 80GB, H100, A40
- Locations: US, EU, CA
- Best for: Fine-tuning LLMs, Stable Diffusion, training, inference
- Key strengths: Cheapest community GPUs, massive variety including H100, serverless endpoints, great UI
- Key weaknesses: Community Cloud less reliable than Secure Cloud, storage costs add up
- Source: https://gpuhosted.com/en/runpod-review/

### 2. Nebius — Rating 4.5/5 (NEW · EU-sovereign)
- Starting price: $1.55/h
- GPUs: H100, H200, B200, L40S, A100 80GB
- Locations: Netherlands, Finland, EU
- Best for: EU-sovereign AI, GDPR-bound enterprises, frontier model training in EU, European startups
- Key strengths: Strong EU data residency (no US transit), B200 production-ready, managed Slurm/Kubernetes, European customer support
- Key weaknesses: More expensive on-demand than US specialists, EU-only regions, smaller global presence
- Source: https://gpuhosted.com/en/nebius-review/

### 3. Lambda Labs — Rating 4.5/5
- Starting price: $1.10/h
- GPUs: A100 40GB, A100 80GB, H100, A10
- Locations: US, AU
- Best for: LLM training, research, fine-tuning, multi-GPU jobs
- Key strengths: Reliable on-demand H100 availability, no minimum billing, Lambda Stack pre-installed
- Key weaknesses: Limited GPU types vs RunPod, fewer EU options, no serverless

### 4. Crusoe — Rating 4.4/5 (NEW · H200/B200/MI300X)
- Starting price: $0.40/h
- GPUs: H100, H200, B200, A100 80GB, L40S, AMD MI300X
- Locations: US, Iceland
- Best for: LLM training at scale, multi-node H100/H200 jobs, AMD MI300X clusters, sustainable AI workloads
- Key strengths: Cheapest H200 access ($2.10/h), B200 availability, only on-demand MI300X, InfiniBand 3.2Tb interconnect, climate-positive (uses flared methane)
- Key weaknesses: Smaller GPU variety than RunPod, regions limited to US/Iceland, sales-led for large deployments
- Source: https://gpuhosted.com/en/crusoe-review/

### 5. CoreWeave — Rating 4.4/5
- Starting price: $2.06/h
- GPUs: H100 SXM, A100 SXM, A40
- Locations: US, EU
- Best for: Large-scale training, foundation models, enterprise AI, multi-node jobs
- Key strengths: Best multi-node GPU cluster performance, high-speed InfiniBand, purpose-built for AI
- Key weaknesses: Expensive, requires Kubernetes knowledge, sales-led process

### 6. Together AI — Rating 4.4/5 (NEW · inference-first)
- Starting price: $1.49/h
- GPUs: H100, H200, A100 80GB, L40S
- Locations: US, EU
- Best for: High-throughput inference, open-source LLM serving, Llama/Mistral fine-tuning, production AI APIs
- Key strengths: Custom inference engines (3–4× faster), excellent open-source model coverage, strong fine-tuning workflow, token-based pricing for variable load
- Key weaknesses: Less GPU variety than RunPod, focus is inference (not raw training), custom interconnects not exposed
- Source: https://gpuhosted.com/en/together-ai-review/

### 7. Hyperstack — Rating 4.3/5 (NEW · cheap A6000)
- Starting price: $0.11/h (RTX A6000)
- GPUs: RTX A6000, A100 80GB, H100, L40, L40S
- Locations: UK, EU
- Best for: Budget training jobs, Stable Diffusion at scale, VPC-isolated workloads, EU-friendly compute
- Key strengths: Outstanding A6000 entry pricing, full networking stack (VPC, firewall, NAT), UK/EU regions for European latency, reservation discount up to 75%
- Key weaknesses: No B200/H200 yet, smaller marketing footprint than RunPod, limited template marketplace
- Source: https://gpuhosted.com/en/hyperstack-review/

### 8. Paperspace — Rating 4.3/5
- Starting price: $0.45/h
- GPUs: A100, A6000, RTX 4000, V100
- Locations: US, EU
- Best for: Notebooks, ML teams, prototyping, education
- Key strengths: Best notebook experience, team collaboration, free tier, good docs
- Key weaknesses: Pricier than RunPod for raw compute, limited GPU types

### 9. Jarvis Labs — Rating 4.3/5 (NEW · polished UI)
- Starting price: $0.39/h (RTX 6000 Ada)
- GPUs: RTX 6000 Ada, A100 40GB, A100 80GB, H100
- Locations: US, Asia
- Best for: Researchers, indie developers, Llama fine-tuning, Stable Diffusion training, Jupyter notebook users
- Key strengths: H100 from $1.99/h (matches RunPod Secure), RTX 6000 Ada at moderate cost, polished UI for non-DevOps users, fast pod launch (<2 min)
- Key weaknesses: Smaller GPU variety than RunPod, no serverless/autoscaling, limited European presence
- Source: https://gpuhosted.com/en/jarvis-labs-review/

### 10. Hetzner GPU — Rating 4.2/5
- Starting price: €0.35/h
- GPUs: A100 PCIe, GTX 1080
- Locations: Germany, Finland
- Best for: EU compliance, research, inference APIs, budget EU GPU
- Key strengths: Best GPU pricing in Europe, GDPR-compliant, excellent API
- Key weaknesses: Limited GPU types, no H100 yet

### 11. Lyceum — Rating 4.2/5 (NEW · EU-sovereign)
- Starting price: $0.39/h
- GPUs: A100 80GB, H100, H200, L40S
- Locations: EU, Iceland
- Best for: EU-regulated industries, GDPR-strict workloads, European public sector, health/finance AI
- Key strengths: Strong EU data residency (no US transit), H200 availability in Europe, ISO 27001 + SOC 2, European billing/contracts
- Key weaknesses: Smaller capacity than US-based clouds, higher base price than RunPod/Vast.ai, limited GPU variety beyond Nvidia
- Source: https://gpuhosted.com/en/lyceum-review/

### 12. TensorDock — Rating 4.2/5 (NEW · cheap marketplace)
- Starting price: $0.21/h (RTX 4090)
- GPUs: RTX 4090, RTX 3090, A100 80GB, H100, L40S
- Locations: US, EU, Global
- Best for: Budget GPU rentals, Stable Diffusion fine-tuning, short-burst training, indie ML developers
- Key strengths: Among cheapest H100 access ($1.99/h), wide host network for availability, per-second billing, free egress on most plans
- Key weaknesses: Reliability varies by host, no managed cluster orchestration, community-led support
- Source: https://gpuhosted.com/en/tensordock-review/

### 13. Massed Compute — Rating 4.1/5 (NEW · workstation focus)
- Starting price: $0.35/h
- GPUs: RTX A6000, A40, A100 80GB, H100, RTX 6000 Ada
- Locations: US
- Best for: VFX and 3D rendering, Stable Diffusion fine-tuning, workstation-style AI dev, multi-tenant studios
- Key strengths: Strong A6000/A40 lineup at moderate price, pre-built VFX/AI templates, RDP/VNC for visual workflows, per-second billing
- Key weaknesses: US-only datacenters, no serverless inference, smaller community than RunPod
- Source: https://gpuhosted.com/en/massed-compute-review/

### 14. Vast.ai — Rating 4.1/5
- Starting price: $0.10/h
- GPUs: RTX 3090, RTX 4090, A100, H100, RTX 3060
- Locations: US, EU, APAC, Global (marketplace model)
- Best for: Batch training, budget experiments, Stable Diffusion, data processing
- Key strengths: Absolute cheapest GPU compute (after Salad), widest variety (consumer + datacenter), marketplace competition
- Key weaknesses: Hosts can take instances offline anytime, variable reliability

### 15. AWS GPU (EC2) — Rating 4.2/5
- Starting price: $3.06/h
- GPUs: A100, H100 (p5), V100, T4, Inferentia2
- Locations: Global
- Best for: Enterprise MLOps, SageMaker, production inference, regulated industries
- Key strengths: Largest GPU instance variety globally, spot up to 90% off, best compliance
- Key weaknesses: Most expensive on-demand, complex pricing, not beginner-friendly

### 16. Azure GPU (NCv3/NDA) — Rating 4.1/5
- Starting price: $2.94/h
- GPUs: A100, H100 (NDA100), V100, T4
- Locations: Global
- Best for: Azure ML pipelines, Microsoft-stack AI, enterprise compliance, OpenAI users
- Key strengths: Deep OpenAI integration, strong compliance, Azure ML Studio
- Key weaknesses: High on-demand, complex portal, vendor lock-in

### 17. Google Cloud GPU — Rating 4.3/5
- Starting price: $2.48/h
- GPUs: A100 40GB, A100 80GB, H100, T4, V100, plus TPU v4/v5
- Locations: Global
- Best for: TensorFlow workloads, TPU training, Vertex AI pipelines
- Key strengths: Best TPU availability, deep Vertex AI/BigQuery integration, preemptible 80% off
- Key weaknesses: Expensive on-demand, complex billing

### 18. Scaleway — Rating 4.0/5 (NEW · French EU cloud)
- Starting price: €0.83/h
- GPUs: L4, L40S, H100, H100 SXM
- Locations: France (Paris), Netherlands (Amsterdam), EU
- Best for: European startups, GDPR-compliant inference, k8s-based AI deployments, EU enterprise
- Key strengths: Strong EU presence (Paris + Amsterdam), mature cloud platform (S3, k8s, networking), per-minute billing, EUR pricing
- Key weaknesses: More expensive than US specialists, no B200/H200 yet, limited capacity for big training runs
- Source: https://gpuhosted.com/en/scaleway-review/

### 19. Salad — Rating 3.9/5 (NEW · distributed inference)
- Starting price: $0.03/h (RTX 3090 distributed)
- GPUs: RTX 3090, RTX 4090, RTX 3080, RTX 3070
- Locations: Global (distributed home gaming PCs)
- Best for: Stateless inference, Stable Diffusion bulk generation, embedding generation, cost-sensitive batch jobs
- Key strengths: Cheapest GPU compute on the market (RTX 3090 at $0.03/h), massive horizontal scale (1000+ nodes), auto-fleet management, no egress charges
- Key weaknesses: Distributed = no persistent storage, NOT suitable for training, latency varies by node geography
- Source: https://gpuhosted.com/en/salad-review/

### 20. OVH GPU — Rating 3.9/5
- Starting price: €0.54/h
- GPUs: T4, V100, A100
- Locations: France, Germany, UK, Canada
- Best for: EU sovereign cloud projects, inference, GDPR-regulated workloads
- Key strengths: EU sovereignty guarantees, established SLA provider, multi-region EU
- Key weaknesses: Older GPU lineup, more complex setup, higher than Hetzner

---

## Key Findings & Verdicts

### On pricing
The same A100 80GB GPU costs $0.79/h on Lambda Labs and $2.48/h on Google Cloud — a 3× spread for identical hardware. The H100 spread is $1.99/h (RunPod Secure / Jarvis Labs / TensorDock) vs $4.10/h (AWS p5 on-demand) — 2×. For purely compute-bound workloads with no need for proprietary ML services, specialist GPU clouds are 2–5× cheaper than hyperscalers. The cheapest GPU on the market is Salad RTX 3090 at $0.03/h, but it's distributed-only and not suitable for stateful training.

### On H100 availability (May 2026)
H100 is now broadly available across **16 of 20 providers** we track. Best H100 availability:
1. CoreWeave — largest dedicated H100 inventory (sales-led)
2. Lambda Labs — strongest on-demand H100 SKU
3. RunPod Secure Cloud — H100s available, $1.99/h, often shorter waitlists
4. Jarvis Labs — H100 from $1.99/h with polished UI
5. TensorDock — marketplace H100 from $1.99/h
6. Crusoe — H100/H200 with InfiniBand 3.2Tb
7. Nebius — H100 in EU (Netherlands/Finland) for sovereign workloads
8. Together AI — H100 with optimised inference stacks
9. Hyperstack — H100 with VPC networking
10. AWS p5 — available but $4+/h, often regional stockouts in popular regions

### On H200 (the 2026 upgrade GPU)
H200 141GB HBM3e is the upgrade path from H100 — 1.4× faster on Llama-2 70B inference, 4.8 TB/s memory bandwidth (vs H100's 3.35), and 141GB VRAM removes parallelism overhead for 70B+ models. Available on-demand from 4 providers in May 2026:
- Crusoe — from $2.10/h (cheapest, climate-positive)
- Nebius — production H200 in EU sovereignty
- Together AI — H200 with optimised inference stacks
- Lyceum — H200 in EU with ISO 27001 / SOC 2

### On B200 (limited availability 2026)
B200 192GB is Nvidia's frontier in 2026 — 2.5× H100 on FP8 training. On-demand access is **extremely limited**; most clouds are wait-listed. Production access is on Crusoe and Nebius. Most teams should choose H200 instead unless they specifically need B200's FP4/FP8 throughput.

### On AMD MI300X
AMD's MI300X (192GB HBM3, 5.3 TB/s) is the only credible non-Nvidia option for frontier workloads. ROCm software has matured but still trails CUDA. **Crusoe is the only on-demand provider** in 2026 ($2.50–$4.00/h). Worth considering if you have ROCm-compatible workloads and want the largest VRAM on the market at a discount to B200.

### On EU sovereignty / GDPR
Five providers offer fully EU-sovereign infrastructure (no US transit, no CLOUD Act exposure):
- **Nebius** (NL/FI) — newest hardware in EU, including H100/H200/B200
- **Lyceum** (EU/Iceland) — A100/H100/H200 with ISO 27001 + SOC 2
- **Scaleway** (FR/NL) — H100 SXM and L40S in Paris/Amsterdam
- **OVH** (FR/DE/UK) — established SLA provider with multi-region EU
- **Hetzner GPU** (DE/FI) — cheapest EU option (€0.35/h A100 PCIe)

For German/EU enterprises in regulated industries (health, finance, public sector), these are the providers that satisfy data residency requirements.

### On hidden costs
Listed hourly price excludes: persistent storage ($0.10–0.20/GB/month), egress ($0.05–0.12/GB), static IPs, snapshot retention, support tier fees. Effective price is typically 15–35% higher than sticker. Hyperscalers (AWS/GCP/Azure) have the highest hidden cost markup; specialist clouds (RunPod, Lambda, Vast.ai, TensorDock) are the most transparent. TensorDock specifically offers free egress on most plans.

### On reliability tiers
- **Production-ready, dedicated**: Lambda Labs, CoreWeave, RunPod Secure, hyperscalers, Nebius, Lyceum, Hyperstack, Hetzner, Scaleway
- **Use with retry/checkpointing**: RunPod Community, Vast.ai marketplace, TensorDock marketplace, GCP/AWS preemptible
- **Stateless / inference-only**: Salad (distributed)
- **Recommendation**: For inference SLAs, use Secure tiers; for training with checkpoints, community/spot saves 60–80%

---

## Use-Case Recommendations

### LLM Fine-tuning (Llama 3, Mistral, Qwen)
- **Llama 3 8B / Mistral 7B**: 24GB GPU sufficient (RTX 4090) — cheapest on TensorDock $0.21/h or RunPod community $0.39/h
- **Llama 3 70B QLoRA**: 1× A100 80GB or 2× A100 40GB NVLink — Lambda Labs ($1.10/h), Jarvis Labs ($0.99/h), or RunPod ($1.89/h)
- **Llama 3 70B full fine-tune**: 8× A100 80GB or 8× H100 — CoreWeave or Lambda Labs reservations · Crusoe for InfiniBand-3.2Tb multi-node
- **Llama 3 405B / frontier models**: H200 (141GB) or B200 (192GB) — Crusoe, Nebius, or Lyceum (EU)

### Stable Diffusion / Image Generation (FLUX, SDXL, SD3)
- **Inference / single-image**: RTX 4090 (24GB) on TensorDock $0.21/h or RunPod community $0.35/h
- **Training LoRAs**: A100 40GB on RunPod $0.79/h or A40 48GB on Massed Compute
- **Production API**: RunPod Serverless · Together AI · Replicate for managed scaling
- **Cheapest bulk inference**: Salad distributed RTX 3090 from $0.03/h

### Inference APIs
- **Cheapest cold start**: RunPod Serverless (per-second billing, scale-to-zero)
- **Lowest latency**: Together AI (custom serving, optimised), Modal, or self-host on Lambda 1-Click clusters
- **Best for OpenAI-compatible**: Together AI, Replicate, Anyscale
- **Cheapest at scale**: Salad distributed inference for stateless workloads

### ML Research (PhD students, research teams)
- **Free credits**: Paperspace free tier, Google Colab Pro, Kaggle
- **Budget**: Vast.ai or TensorDock for batch jobs, RunPod community for interactive, Jarvis Labs for Jupyter UX
- **University-grade**: Lambda Labs reservations or CoreWeave for cluster work

### Video AI (Wan, CogVideoX, Sora-like)
- Need ≥40GB VRAM, ideally 80GB → A100 80GB or H100 (or H200 for long context)
- **Best price**: RunPod Secure A100 80GB at $1.89/h
- **Best reliability**: Lambda Labs H100 at $2.49/h on-demand
- **EU-sovereign**: Nebius H100 from NL datacenter for GDPR workloads

### EU-sovereign / GDPR-strict (Health, Finance, Public Sector)
- **Cheapest entry**: Hetzner GPU €0.35/h (A100 PCIe in Germany/Finland)
- **Modern hardware in EU**: Nebius H100/H200/B200 (Netherlands, Finland)
- **ISO 27001 + SOC 2**: Lyceum A100/H100/H200 (EU + Iceland)
- **k8s-native EU**: Scaleway H100 SXM + L40S (Paris, Amsterdam)
- **Established SLAs**: OVH GPU (France, Germany, UK)

---

## Pricing Spread (May 2026)

| GPU | Cheapest | Mid | Most expensive | Spread |
|---|---|---|---|---|
| RTX 3090 (24GB) | $0.03/h Salad | $0.20/h RunPod | $0.79/h Paperspace | 26× |
| RTX 4090 (24GB) | $0.21/h TensorDock | $0.35/h RunPod | $1.20/h Paperspace | 5.7× |
| RTX A6000 (48GB) | $0.11/h Hyperstack | $0.34/h RunPod | $0.49/h Massed | 4.5× |
| A40 (48GB) | $0.39/h RunPod | $0.49/h Massed | $0.59/h CoreWeave | 1.5× |
| A100 40GB | $0.79/h Lambda | $1.10/h RunPod | $2.48/h GCP | 3.1× |
| A100 80GB | $1.09/h Vast.ai | $1.79/h RunPod | $3.67/h AWS | 3.4× |
| L40S (48GB) | $0.70/h Massed | $1.49/h Together | $2.50/h Scaleway | 3.6× |
| H100 (80GB) | $1.99/h RunPod/Jarvis/TensorDock | $2.49/h Lambda | $4.10/h AWS p5 | 2.1× |
| H200 (141GB) | $2.10/h Crusoe | $2.50/h Nebius | $4.50/h Lyceum | 2.1× |
| B200 (192GB) | Sales-led | Sales-led | Sales-led | — |
| MI300X (192GB) | $2.50/h Crusoe | — | — | exclusive |

---

## Methodology

We test each provider on:
1. **Real on-demand pricing** (verified via the provider's pricing page within the last 30 days)
2. **GPU availability** (manual provisioning attempts to detect stockouts)
3. **Setup friction** (time from sign-up to first GPU running)
4. **Hidden costs** (storage, egress, IP, support tier)
5. **Documentation quality** and API completeness
6. **Reliability tier** (community/spot vs dedicated/secure)
7. **Geographic data residency** (US, EU, sovereign options)

Some links on gpuhosted.com are affiliate links — we earn a commission at no extra cost to the user. This does not influence rankings; ratings are based purely on the criteria above.

Last methodology update: May 1, 2026.

---

## Pages on gpuhosted.com

### English
- /en/best/ — Full comparison of all 20 providers
- /en/finder/ — Interactive GPU Finder wizard
- /en/cost-calculator/ — Cost estimator for Llama-3, SDXL, inference
- /en/best-h100-cloud/, /en/best-h200-cloud/, /en/best-b200-cloud/, /en/best-a100-cloud/, /en/best-rtx-4090-cloud/, /en/best-rtx-3090-cloud/, /en/best-l40s-cloud/, /en/best-a40-cloud/, /en/best-mi300x-cloud/, /en/cheapest-gpu-cloud/
- 20 in-depth provider reviews at /en/{slug}-review/
- /en/[provider-a]-vs-[provider-b]/ — head-to-head versus pages
- /en/blog/ — Long-form articles (H100 vs A100, GPU cost-cutting, RunPod vs Lambda Labs, best GPU for Stable Diffusion)

### German (Deutsch)
- /de/best/ — Vollständiger Vergleich aller 20 Anbieter
- /de/finder/ — GPU-Finder-Assistent
- /de/cost-calculator/ — GPU-Cloud-Kostenrechner
- /de/h100-mieten/ — NVIDIA H100 mieten 2026
- /de/gpu-server-deutschland/ — GPU-Server in Deutschland & Europa
- /de/gpu-server-eu-dsgvo/ — DSGVO-konforme GPU-Cloud
- /de/gpu-server-mieten-guenstig/ — GPU-Server günstig mieten
- /de/blog/ — Deutsche Blog-Artikel
- 20 Anbieter-Tests at /de/{slug}-review/

---

## Contact

- Email: hello@gpuhosted.com
- Site: https://gpuhosted.com
- Brief overview: https://gpuhosted.com/llms.txt
- Full comparison data: https://gpuhosted.com/llms-full.txt (this file)