Best GPUs to rent for LLM training & fine-tuning
For full LLM training runs, rent H100 or newer (H200, B200): HBM bandwidth and tensor performance dominate training throughput. For LoRA/QLoRA fine-tuning of models up to ~13B parameters, a single 24GB card such as an RTX 4090 or RTX 3090 is usually enough and costs a fraction of the price. A100 80GB remains the best value for multi-day jobs that do not need the latest silicon.
Prices below are live minimum / median USD per hour on the GPUniq marketplace, as of 2026-07-21 00:37 UTC. Refreshed every 5 minutes.
| # | GPU | VRAM | Min $/hr | Median $/hr | Available | Why this card |
|---|---|---|---|---|---|---|
| 1 | B200 | 192GB HBM3e | $8.19 | $10.74 | 54 | Blackwell flagship — fastest single-GPU training available for rent; 192GB HBM3e fits the largest models. |
| 2 | H200 | 141GB HBM3e | $4.23 | $6.05 | 199 | 141GB HBM3e — same compute as H100 with ~1.8× the memory; fewer GPUs needed per job. |
| 3 | H100 | 80GB HBM3 | $2.61 | $3.03 | 90 | The de-facto standard for serious training; best software ecosystem and availability. |
| 4 | A100 | 40–80GB HBM2e | $0.80 | $1.31 | 251 | Previous-gen workhorse — typically the lowest $/TFLOPS for multi-day training runs. |
| 5 | RTX 6000 Ada | 48GB GDDR6 | $0.74 | $1.03 | 59 | 48GB workstation card — full-precision fine-tuning of mid-size models without data-center pricing. |
| 6 | RTX 4090 | 24GB GDDR6X | $0.30 | $0.56 | 1742 | 24GB consumer card — the budget king for LoRA/QLoRA fine-tuning up to ~13B. |
Source: GPUniq live marketplace analytics (2026-07-21 00:37 UTC). Each linked page shows price history for 24h/7d/30d/365d.
Frequently asked questions
How much VRAM do I need to fine-tune a 7B model?
With QLoRA (4-bit), a 7B model fine-tunes comfortably on a single 24GB card such as an RTX 4090 or RTX 3090. Full-precision (bf16) full fine-tuning of 7B needs roughly 80–120GB across GPUs — one A100/H100 80GB with offloading, or 2+ cards.
H100 vs A100 for training — is the premium worth it?
H100 delivers roughly 2–3× A100 training throughput on transformer workloads (FP8/bf16). If your hourly H100 price is less than ~2.5× the A100 price, H100 usually wins on total job cost. Check both live prices on this page before launching.
Can I train on consumer GPUs like the RTX 4090?
Yes — for fine-tuning (LoRA/QLoRA) and small models it is the cheapest option per experiment. The limits are 24GB VRAM, no NVLink, and consumer interconnect; full multi-GPU pre-training belongs on H100-class hardware.