Vast.ai

Vast.ai is a GPU rental marketplace. The console on the public page listed example cards such as RTX 5090 around $0.27–$0.57 / hr and H200 NVL around $1.48 / hr. Those are live offers, not a contract. The old "17,000+ GPUs / 1,400 hosts / 50-80% cheaper than AWS" line is not on the pricing page we used, so it is out.

Compare Modal if you want a first-party serverless GPU price list, or E2B if you wanted an agent sandbox rather than a rented card.

Key Features

Marketplace, not a reserved cloud: Hosts post cards. You pick interruptible or on-demand.
Per-second style billing: Confirm the current minimum credit and disk meters in the console.
Docker / SSH / Jupyter: Typical launch path is still a container plus SSH.
Reliability tiers: The UI still distinguishes cheaper community hosts from verified / datacenter offers. Cheapest is not always cheapest completed.

Use Cases

Short GPU jobs that can die and restart.
Fine-tunes that would be painful at hyperscaler list prices, if you accept host variance.
People who do not need a sandbox product, just a card.

Limitation: listed $/hr moves every refresh. Do not paste a 2025 A100 band into a contract. We did not re-verify SOC 2 / ISO claims on a public compliance page.

Pricing

Examples from the 2026-08-17 console, not a rate card:

Offer seen that day	About
RTX 5090 interruptible	~$0.27–$0.36 / hr
RTX 5090 on-demand	~$0.57 / hr
H200 NVL interruptible	~$1.48 / hr

Confirm vast.ai/pricing or the search UI before you quote any GPU.

Getting Started

Create an account at vast.ai.
Filter for the GPU and interruptible vs on-demand.
Launch a template, then SSH.
Destroy the instance when the job ends. Marketplace idle time is still your bill.

Frequently Asked Questions

Is it cheaper than AWS?

Often, on the offers we saw. It is not a guaranteed discount.

Can I use it as an agent sandbox?

You can exec in a container. It is not E2B.

Are the 17k GPU stats still right?

Not as a first-party sentence we kept.

Alternatives

Modal: First-party GPU seconds and a sandbox API.
E2B: Firecracker sandboxes for agents.
Daytona: Persistent dev environments if you wanted a workspace, not a marketplace GPU.

Tips

Sort by verified hosts if downtime costs more than the sticker.
Confirm disk and bandwidth extras. They are not always in the GPU hour.
Do not keep GTX 1080 rows as if they were the 2026 catalog.

Conclusion

Vast.ai is still a host marketplace for GPUs, not a fixed price list. Start at vast.ai/pricing, pick a live offer, and compare Modal if you wanted a vendor SLA instead of a bid.