Cloud Rentals — vast.ai, RunPod, Lambda

The "rent before you buy" tier of self-hosted AI. vast.ai is the cheapest hourly GPU marketplace; RunPod is the slicker option; Lambda Labs is the premium US choice; TensorDock, Modal, Hyperstack are newer entrants. For occasional fine-tuning, validating a build, or spike workloads, renting beats buying. For steady use, hardware wins after <1 year — see cost math.

Cross-links: hardware tier guide; overview; fine-tuning for the most common reason to rent.

vast.ai

★ vast.ai (vast.ai) — peer-to-peer GPU marketplace.

• Hosts list their idle GPUs; you bid (interruptible) or pay-on-demand (uninterruptible).
• Cheapest tier of any provider; community / hobbyist hardware.
• May 2026 representative pricing:
  • RTX 3090 24GB: $0.20–0.30/hr
  • RTX 4090 24GB: $0.40–0.55/hr
  • RTX 5090 32GB: $0.55–0.75/hr
  • A100 40GB: $0.80–1.20/hr
  • A100 80GB: $1.10–1.60/hr
  • H100 80GB: $1.80–2.50/hr
  • H200: $2.20–3.00/hr
• Pros: cheap, broad hardware selection, simple SSH-in workflow.
• Cons: quality varies host-to-host; spot instances can get reclaimed; trust the hosts with your data.
• ★ Best for: short jobs, fine-tuning, "I want to try a 70B before buying," batch generation.

RunPod

★ RunPod (runpod.io) — cleaner UX, hosted in two tiers.

• Community Cloud — like vast.ai; cheap, less guaranteed.
• Secure Cloud — RunPod-operated, pricier, more reliable.
• Pods (long-running VMs) and Serverless (per-second autoscale).
• Persistent volumes; templates for common AI stacks (Ollama, ComfyUI, vLLM, Stable Diffusion).
• May 2026 representative pricing on Secure tier:
  • RTX 4090: $0.45–0.60/hr
  • A100 80GB: $1.40–2.00/hr
  • H100 80GB: $2.50–3.50/hr
• Pros: smoother UX than vast.ai, persistent volumes, serverless option.
• Cons: noticeably pricier than vast.ai for similar hardware.
• ★ Best for: production-flavoured workloads, persistent dev environments, occasional fine-tuning where reliability matters.

Lambda Labs

★ Lambda (lambdalabs.com) — premium US datacenter GPU rental.

• H100 / H200 focus; A100 still available.
• Higher quality / lower variance than vast.ai or RunPod community.
• Pricier; closer to AWS/Azure than vast.ai.
• ★ Best for: serious training runs where you need consistent performance and US-based ML expertise / support.

TensorDock, Hyperstack, Modal, Replicate

• TensorDock (tensordock.com) — competitive rates; smaller than RunPod.
• Hyperstack — UK / Europe focus; H100 forward.
• Modal (modal.com) — Python-first serverless GPU; pay-per-second; great for batch jobs and inference serving as code.
• Replicate (replicate.com) — model-as-API hosting; not raw GPU rental but related; great for "host my fine-tuned model and call it via API."

AWS / Azure / GCP

• Available; pricier than the above for same hardware.
• Worth it only if you're already in that ecosystem or need specific compliance (SOC2, HIPAA).

When to rent vs. buy

• ★ Rent for spike workloads — fine-tuning a 70B, generating 1000 video clips, validating a build.
• ★ Rent for hardware tiers you can't afford to buy — H100, H200, MI300X.
• ★ Rent before buying — spend $30–50 on a week of 3090 time on vast.ai before dropping $900 on a used 3090; verify your workflow.
• Buy for steady daily use — break-even on a $1,200 build is <1 year if you'd rent equivalent hours every day.

Common patterns

• Fine-tune on rented H100, deploy locally. Train on $2/hr H100 for 6 hours = $12; serve with the resulting LoRA on your local 3090.
• Burst inference on serverless. Modal / Replicate / RunPod Serverless for occasional heavy jobs; near-zero cost when idle.
• Validate a hardware purchase. Rent a 3090 for a week ($30) before dropping $900 on used hardware.
• Privacy-zone work locally + heavy work in cloud. Sensitive data on local Tier 2 hardware; bulk batch jobs in cloud.

Honest caveats

• vast.ai security model is "trust the host." Don't put truly sensitive data on community-cloud spot instances. Use RunPod Secure or Lambda for that.
• Spot interruptions. vast.ai bidding can get evicted; checkpoint frequently.
• Cold-start time. Pulling a 70B model fresh on a new instance is 5–10 minutes; persistent volumes help.
• Egress / bandwidth. Most providers don't charge for inbound; outbound can surprise you on big batch outputs.
• Currency volatility on global hosts — vast.ai listings sometimes shift with exchange rates.

Pick this if…

• Cheapest hourly GPU: ★ vast.ai.
• Slicker UX, persistent volumes: RunPod.
• Premium US training: Lambda.
• Serverless / per-second / pay-as-you-go: Modal or Replicate.
• Validating a build: vast.ai for $30, prove your workflow, then buy hardware.
• Heavy fine-tuning: Rent H100 on vast.ai or Lambda for the duration; cheaper than buying.
• Steady daily use: stop renting, build hardware.