Best GPU clouds for fine-tuning LLMs (2026)
Fine-tuning sits between training and inference: short bursts of GPU work interleaved with eval, dataset shuffling, and human review. Billing granularity (per-second vs per-hour), persistent storage, and dataset egress hit harder than the headline $/hr. Three picks across the axes.
Live pricing + 30-day reliability update on every page load. Curation refreshed manually.
Hyperstack · 1× A100-80GB
LoRA / QLoRA on a 70B model fits a single A100 80GB. Region-aware availability across 3 DCs (CANADA-1, NORWAY-1, US-1), 100% data coverage in our 30-day window. Free egress.
Lambda · 4× H100
Per-minute billing, persistent filesystems on 1-Click Clusters, snapshot workflow that doesn't require a manual EBS dance. You can stop a 4×H100 box overnight without losing your dataset cache.
Runpod · 1× H100
Per-second billing + persistent volumes at $0.07/GB-mo. Iterative fine-tuners burn 30-50% of wall-clock idle (debugging, eval, waiting for human review). Per-second vs per-hour on a 14-min run = pay for 14 min vs 60 min — a ~4× difference invisible in the listing price.
Avoid: Vast (RTX 4090)
Consumer cards (24 GB VRAM) for anything past 7B QLoRA force aggressive memory offload that 5×'s wall-clock time vs an A100 80GB at 2× the hourly rate. Net loss on TCO every time.
The caveat we wish more pages mentioned
Dataset egress is the silent killer. If your dataset lives in S3 and you fine-tune on Lambda, AWS charges egress every time you re-pull. A 2 TB dataset = $180 one-time exit fee at $0.09/GB. Co-locate compute with storage, OR park your dataset in Cloudflare R2 ($0 egress) as a hub.
Data we don't yet show — and how it might change the call
- Snapshot/checkpoint create + restore time for a 70B image — nobody publishes this
- Persistent storage $/GB-mo per provider
Honesty about gaps beats false confidence. We add data as it becomes structurally available.