Vast.ai is the largest GPU rental marketplace. Anyone with spare NVIDIA capacity can list it, set a price, and start earning. That model produces the lowest floor prices in the market, and also the widest variance in what you actually get: hardware condition, host uptime, network quality and job continuity depend on the individual host, not a platform guarantee.
packet.ai owns and operates its own GPU infrastructure. It has one rate card. Jobs do not get pre-empted. This comparison is about whether the Vast.ai price floor is worth the reliability trade-off for your specific workload.
Platform overview: packet.ai vs Vast.ai
What Vast.ai offers
Vast.ai's catalogue is whatever hosts have listed. Typical on-demand starting rates in September 2026: RTX 4090 from about $0.35/hr, RTX 5090 from $0.41/hr, L40S from $0.47/hr, RTX 6000 Pro from about $1.00/hr, A100 80GB from $1.28/hr, B200 from around $4.34/hr with a median across verified hosts closer to $7.78/hr. Interruptible bids run lower still.
- Per-second billing; storage and bandwidth billed separately by each host at their own rates
- Host reliability score and datacentre/verified flags help you filter, but there is no platform SLA
- Interruptible instances can be outbid and stopped mid-job
- Docker-based; you bring an image and expose ports, similar to RunPod Pods
What packet.ai delivers
Vast.ai's floor prices are real, but they are bids on third-party hosts. The RTX 4090 floor of $0.35/hr is an interruptible instance on a random host that can outbid you and stop your job mid-run. The hosts with verified datacentre status and reliability scores good enough to trust for production work cost meaningfully more than the floor. packet.ai has no floor: every rate is a published, fixed price on first-party infrastructure that does not pre-empt your workload.
The relevant comparison is not Vast.ai's floor against packet.ai's rates. It is what a reliable Vast.ai instance actually costs versus packet.ai:
- Dynamic: scheduler-enforced multi-tenant on packet.ai's own hardware. Never interruptible, never outbid, never stopped mid-job. RTX 4090 equivalent at $0.39/hr versus Vast.ai's $0.35/hr floor on interruptible hosts. L40S at $0.92/hr versus Vast.ai's $0.47/hr floor. RTX 6000 Pro at $0.66/hr versus Vast.ai's $1.00/hr floor (packet.ai is cheaper here). B200 at $3.75/hr versus Vast.ai's $4.34/hr floor.
- Dedicated: a whole card, exclusively yours, 99% SLA. RTX 4090 at $0.39/hr, L40S at $0.92/hr, A100 80GB at $1.43/hr, B200 at $6.99/hr. Rates are published and do not change based on competing bids.
What Vast.ai has that packet.ai does not: a much wider GPU selection (including cards packet.ai does not list), interruptible pricing for jobs that can tolerate interruption, and per-second billing. If your job is stateless and resumable and you are optimising purely for cost floor, Vast.ai interruptible bids can be cheaper. If you need the job to finish without being pre-empted, packet.ai's published rate on owned hardware is the more honest comparison.
packet.ai vs Vast.ai at a glance
packet.ai rates are fixed published prices. Vast.ai rates are on-demand floor bids; interruptible bids and verified-host non-interruptible rates will differ.
| GPU | packet.aiYOU | Vast.ai (floor, on-demand) |
|---|---|---|
| RTX 4090 (24GB) | $0.39/hr Dedicated | ~$0.35/hr interruptible floor |
| L40S (48GB) | $0.92/hr Dedicated | ~$0.47/hr floor |
| RTX 6000 Pro (96GB) | $0.66/hr Dynamiccheaper | ~$1.00/hr floor |
| A100 80GB | $1.43/hr Dedicatedcheaper | ~$1.28/hr floor |
| B200 | $3.75/hr Dynamiccheaper | ~$4.34/hr floor (~$7.78/hr median verified) |
| Pre-emption risk | None (Dynamic is never outbid) | Yes on interruptible bids |
| Platform SLA | 99% on Dedicated | None (host-level only) |
| Infrastructure | First-party, owned by packet.ai | Third-party hosts, peer-to-peer marketplace |
| Rate stability | Fixed published rate | Market-driven, varies by bid and host |
| Billing | Hourly; monthly commits up to 20% off | Per-second; host storage/bandwidth billed separately |
Detailed comparison
The floor versus the real price
Vast.ai's advertised floor rates are interruptible bids. A job running on an interruptible instance can be stopped at any time by a higher bidder. For training runs with checkpointing, this is manageable overhead. For long inference jobs, API serving or anything stateful, it is a real operational risk.
Non-interruptible on-demand instances from verified datacentre hosts run higher than the floor. The B200 floor of $4.34/hr is an interruptible bid; the verified-host median of $7.78/hr is closer to what a production-grade Vast.ai B200 actually costs. Against that number, packet.ai's $3.75/hr Dynamic is substantially cheaper, and Dynamic is never pre-empted.
Where Vast.ai has the edge
GPU breadth. Vast.ai lists hardware packet.ai does not: RTX 5090, H100 in various configurations, MI300X, and older cards like A6000 and A5000. If you need a specific card not on packet.ai's list, Vast.ai likely has a host running it somewhere.
Interruptible pricing for the right workload. Distributed training with frequent checkpointing, batch processing jobs and any job that can restart cleanly can exploit Vast.ai's bid market effectively. Per-second billing also favours very short jobs.
What packet.ai trades
Smaller GPU selection and hourly billing granularity. packet.ai does not match Vast.ai's card range and its minimum billing unit is an hour. For GPU types and short burst jobs, Vast.ai is the more flexible option.
Who should choose which
Choose packet.ai if
- Your job cannot be interrupted - serving, stateful training, long inference
- You want a fixed published rate with no bid market uncertainty
- You need RTX 6000 Pro or B200 cheaper than Vast.ai's verified-host rates
- A 99% SLA on the card matters for your use case
Choose Vast.ai if
- Your job checkpoints frequently and can resume after interruption
- You need a GPU type not on packet.ai's current list
- You want per-second billing for very short jobs
- You are optimising for the lowest possible cost floor and can accept host variance
Conclusion
Vast.ai and packet.ai answer different questions. Vast.ai answers: what is the cheapest GPU-hour available anywhere right now? packet.ai answers: what is a reliable, published GPU-hour on first-party hardware? On RTX 6000 Pro, B200 and A100, packet.ai is cheaper than Vast.ai's floor. On everything else, Vast.ai's floor is lower, but that floor comes with pre-emption risk and host variance that packet.ai does not carry.
Same silicon. Smarter economics.
Every rate on this page is a published starting price you can deploy against today. Dynamic launches in under five minutes; no credit card to start.
Additional resources
- packet.ai GPU pricing: the full Dynamic, Dedicated and Clusters rate card.
- packet.ai documentation: API, CLI and platform guides.
- How intelligent scheduling works: why Dynamic keeps peak performance at a lower price.
- Clusters and wholesale quotes: multi-node InfiniBand capacity at ~30% below retail.
- Rent GPUs on packet.ai: specs, pricing and workload fit for every card.