We benchmarked all three on the same fine-tuning run — Llama 3.1 8B LoRA on 4x A100 for 6 hours.
RunPod Spot: $1.39/hr per GPU, spot market, interruptible. Lost one run to a preemption at hour 4. Total cost $33.
RunPod On-Demand: $1.89/hr per GPU, reserved for your session. No interruptions, $45 total. Best balance of price and reliability for medium runs.
Lambda Labs: $1.50/hr per A100, reliable infrastructure, great docs, no spot weirdness. $36 total. Slightly slower disk for dataset loading.
Vast.ai: $0.89/hr spot — cheapest, but you get what you pay for. Two pods disappeared mid-run, one had a flaky network. Usable for stateless batch jobs with checkpointing everywhere.
Verdict: RunPod on-demand for anything that matters, Vast.ai for disposable batch experiments, Lambda Labs for teams that want AWS-like reliability without AWS complexity.
Golden rule for all three: checkpoint every 30 minutes or cry later. GPU preemption doesn't care about your deadline.