Verified July 2026 (spheron.network):
- B200 verified datacenter: $6-$12/GPU-hr (median ~$9)
- B200 supply: Intermittent; listings appear and disappear with deployment cycles
Treating Vast.ai as a reliable B200 source would be a planning mistake. The bulk of B200 capacity sits with CoreWeave, Lambda, and RunPod — providers that own or have priority access to Blackwell hardware. For production Blackwell workloads, Vast.ai is at best a supplementary source for opportunistic capacity.
A100, L40S, and older generation pricing
For older-generation GPUs, Vast.ai's marketplace rates are the cheapest published rates in the category.
Verified July 2026 (gpucloudcost.com):
- A100 80GB marketplace floor: $0.79/GPU-hr interruptible
- L40S marketplace median: $0.89/GPU-hr
- RTX 4090 marketplace median: $0.45/GPU-hr (consumer GPU)
- RTX 3090 marketplace floor: $0.21/GPU-hr (consumer GPU)
Compared to RunPod Community Cloud A100 PCIe 40GB at $1.19/GPU-hr and Lambda A100 PCIe 40GB at $1.99/GPU-hr, Vast.ai A100 80GB at $0.79/GPU-hr is 30-50% cheaper. The A100 marketplace on Vast.ai is mature with stable supply — these rates are reliable enough for production training workloads that can tolerate host-level reliability.
Interruptible vs on-demand vs reserved
Vast.ai's three pricing tiers match the major cloud models with marketplace-specific reliability implications.
Verified July 2026 (vast.ai/docs/guides/instances/pricing):
- Interruptible: Lowest cost (50%+ cheaper than on-demand); host can evict at any time; ideal for batch training and fault-tolerant workloads
- On-demand: Fixed pricing; high priority; host cannot actively evict but machine stays online only while host keeps it online; no platform-level uptime guarantee
- Reserved: Discounted rate with pre-payment commitment; up to 50% off on-demand; high priority; preempts interruptible on the same host
The interruptible tier is Vast.ai's signature offering — 50% or more below on-demand pricing. For training workloads with checkpoint discipline (save every 30 minutes or less), interruptible at $1.49/GPU-hr for H100 PCIe is the cheapest credible option in the GPU cloud market.
Side-by-side Vast.ai vs RunPod vs Lambda pricing
| GPU / SKU | Vast.ai Verified | Vast.ai Interruptible | RunPod Community | RunPod Secure | Lambda |
|---|
| Cheapest published | A100 80GB $0.79 | RTX 3090 $0.21 | RTX A6000 $0.33 | L40S $0.99 | Quadro RTX 6000 $0.69 |
| A100 80GB | ~$1.00 | $0.79 | $1.19 (PCIe 40GB) | $1.39 | $1.99 (PCIe 40GB) |
| H100 PCIe | $1.50-$2.27 | $0.90-$1.60 | $1.99 | $2.89 | $3.29 (1x) |
| H100 SXM | $1.80-$2.50 | Rare | $2.69 | $3.29 | $3.99 (8x) |
| H200 | $3.50-$5.50 | Rare | $3.59 | $4.59 | Not listed |
| B200 | $6-$12 | Not available | $5.98 | $6.79 | $6.69 (8x) |
| Egress | Host-dependent | Host-dependent | Free outbound | Free outbound | Free |
| Billing | Per-second | Per-second | Per-second | Per-second | Per-minute |
How to use Vast.ai safely in 2026
Three rules for using Vast.ai without losing your training job.
- Use datacenter-verified hosts only. Unverified community hosts can pull hardware offline at any moment. Sort by DLPerf score within the verified host filter for reliability and performance.
- Checkpoint every 30 minutes or less. Vast.ai interruptible instances can be reclaimed with short notice. Save model state to S3-compatible storage every 30 minutes; checkpointing discipline is the cost of the 50%+ interruptible discount.
- Verify host reliability before committing compute. Check the host's uptime history, DLPerf score, and prior tenant reviews. A verified host with 95%+ uptime is meaningfully more reliable than one with 70% uptime at the same price.
What enterprise buyers should do next
Three actions for organizations evaluating Vast.ai in 2026.
- Run a 2-week pilot on interruptible H100 PCIe verified datacenter at $1.50-$2.27/GPU-hr. Same workload as Lambda or RunPod. Compare cost (likely 30-50% lower) against reliability variance (likely higher). Quantify both before committing to a multi-month training workflow.
- Model storage and egress costs separately. Vast.ai charges for storage per GB-month and host-dependent network transfer. For data-intensive workloads, these line items can add 10-15% to the headline GPU rate.
- Reserve capacity for time-sensitive workloads. Reserved pricing (up to 50% off on-demand with pre-payment commitment) is the right choice for training jobs with hard deadlines. Interruptible is fine for exploratory work and research; reserved is mandatory for production.
What to watch next
Three near-term datapoints. First, Vast.ai's B200 and Blackwell Ultra listings — supply is currently intermittent; expect more listings as Blackwell hardware proliferates to community data centers by Q1 2027. Second, Vast.ai Serverless pricing — the platform now offers auto-scaling serverless workers with no separate pricing tier, which makes Vast.ai competitive with Modal and Replicate for spiky inference workloads at lower GPU rates. Third, Vast.ai reserved pricing adoption — if reserved capacity scales, Vast.ai's verified datacenter H100 PCIe at $1.50-$2.27/GPU-hr reserved (50% off = $0.75-$1.13/GPU-hr) would undercut every specialist-cloud reserved contract in the market (vast.ai/pricing, July 2026).