What’s the cheapest way to get 8 GPUs for a week?

The cheapest way depends on whether you mean cloud GPUs or buying/renting physical hardware, but for a one-week need, the cheapest overall is usually:

1) Use spot/preemptible cloud instances

These are often much cheaper than on-demand and are the best first place to look.

Good options

• AWS Spot Instances
• GCP Preemptible / Spot VMs
• Azure Spot VMs
• GPU-focused marketplaces like:
  • RunPod
  • Vast.ai
  • Lambda Cloud (sometimes competitive, less variable)
  • CoreWeave (often more enterprise-oriented)

Why this is cheapest

• You pay only for the week
• Spot pricing can be 30–90% cheaper than normal pricing
• You can often choose older/less in-demand GPUs (A10, T4, V100, etc.) to save more

Downside

• Spot instances can be interrupted
• You may need checkpointing / restart logic

---

2) Choose the right GPU model

If your workload doesn’t need top-end GPUs, cheaper cards save a lot.

Usually cheaper:

• NVIDIA T4
• NVIDIA A10 / A10G
• NVIDIA L4
• Older V100 / P100 on some marketplaces

Usually expensive:

• A100
• H100
• 4090s in small-volume rentals can vary a lot

If “8 GPUs” just means 8 cards, not necessarily 8 high-end cards, then 8× T4/A10 on a marketplace is often the lowest-cost route.

---

3) Use GPU rental marketplaces

These are often cheaper than hyperscalers for short-term rentals.

Best for price

• Vast.ai is often the cheapest if you’re flexible
• RunPod is usually simple and reasonably priced
• TensorDock can also be competitive

You can sometimes find:

• 8 GPUs on one machine
• Or multiple smaller nodes totaling 8 GPUs

---

4) If you need guaranteed uptime, compare to reserved/higher-end rentals

If interruption is unacceptable, spot pricing may stop being “cheapest” once you account for lost work.

In that case, the cheapest reliable options are often:

• RunPod reserved
• Lambda Cloud
• Vast.ai dedicated hosts from reputable providers

---

Rough rule of thumb

For 8 GPUs for 7 days, the lowest-cost path is usually:

1. Vast.ai / RunPod spot
2. Cloud spot instances
3. Dedicated GPU rental
4. On-demand cloud
5. Buying hardware (not economical for just one week)

---

How to minimize cost

• Pick older or midrange GPUs
• Use spot/preemptible
• Put all 8 GPUs on one host if your software benefits from that
• Avoid large storage/egress fees
• Shut down instances during idle time
• Use checkpointing to survive interruptions

---

If you want the absolute cheapest

Tell me:

1. Do you need all 8 GPUs on one machine, or just 8 total?
2. What GPU model do you need, if any?
3. What workload is it? (training, inference, rendering, etc.)
4. What region matters?

Then I can suggest the cheapest realistic option and estimate the cost.