Lambda Cloud vs RunPod vs Vast.ai: Real GPU Rental Cost per 1,000 GPU Hours in 2026
The short answer: these three are not the same kind of business, and that shows up directly in price. Lambda Cloud operates its own data centers with an SLA and charges accordingly. RunPod splits its inventory into a marketplace tier (Community Cloud) and its own data centers (Secure Cloud), priced roughly 20–100% apart depending on GPU. Vast.ai is a pure marketplace with no owned hardware, and its on-demand rates undercut both. On an H100-class GPU, 1,000 GPU-hours costs roughly $1,470 on Vast.ai on-demand, $1,990 on RunPod Community, $2,490 on Lambda, and $2,890 on RunPod Secure, before idle time or interruption is considered. Community-hosted and enterprise-secure capacity are not interchangeable, and this article keeps them labeled separately throughout.
What "secure" versus "community" actually means
RunPod Secure Cloud runs on RunPod's own data centers with an SLA (reported around 99.5% uptime) and NVLink support. RunPod Community Cloud is third-party hosts running on RunPod's platform, with variable hardware and no dedicated-infrastructure guarantee, priced lower to compensate. Vast.ai is entirely a marketplace: every listing is set by the individual machine owner, split into unverified community hosts and datacenter-verified hosts, with prices updating live based on supply and demand. Lambda Cloud operates its own hardware across seven US data centers with no marketplace tier at all, which is the main reason its prices sit at the high end of this comparison.
Reported hourly rates
All figures are REPORTED, since none of the three publishes a single canonical rate that holds across regions and moments; marketplace and even owned-fleet prices move with availability.
| GPU class | Lambda | RunPod Community | RunPod Secure | Vast.ai on-demand | Vast.ai interruptible |
|---|---|---|---|---|---|
| RTX 4090-class | not offered by Lambda | $0.34/hr | $0.69/hr | $0.31/hr | $0.15/hr |
| A100 80GB | $1.79/hr (reported range up to $1.79) | $1.49/hr (one source reports $1.89 for the same class) | $1.89/hr | $1.20/hr (marketplace range $0.29–$1.90 depending on host) | $0.60/hr |
| H100 80GB | $2.49/hr (1x SXM; another source reports $2.99/hr, and an 8x cluster was reported at $3.29/GPU-hr, higher per-GPU than the single-unit rate) | $1.99/hr (PCIe) | $2.89/hr (PCIe), $2.99/hr (SXM) | $1.47/hr (one source), $0.90–$1.87/hr (a second, wider marketplace range) | $0.90/hr |
The RunPod A100 figures disagree between sources ($1.49 versus $1.89 for what both describe as Community Cloud A100), and Lambda's H100 rate is reported at $2.49, $2.99, and $3.29 per GPU depending on cluster size and source. This spread is not a data error; it reflects that Lambda, RunPod and especially Vast.ai prices genuinely move with region, availability and time.
Cost per 1,000 GPU hours
Formula: cost = hourly rate × 1,000, before idle time, storage or interruption.
| GPU class | Lambda | RunPod Community | RunPod Secure | Vast.ai on-demand | Vast.ai interruptible |
|---|---|---|---|---|---|
| RTX 4090-class | — | $340 | $690 | $310 | $150 |
| A100 80GB | $1,790 | $1,490 | $1,890 | $1,200 | $600 |
| H100 80GB | $2,490 | $1,990 | $2,890 | $1,470 | $900 |
At 100 GPU hours a month, these figures scale down linearly to $34–$69 (4090-class), $120–$189 (A100), and $147–$289 (H100); at 10,000 GPU hours a month they scale up to $3,100–$6,900, $12,000–$18,900, and $14,700–$28,900 respectively.
Effective cost after idle time
A rented pod that sits idle 10%, 30% or 50% of the time still bills for every hour it is reserved unless explicitly stopped. Formula: effective cost = hourly rate ÷ utilization.
| GPU class, provider | 10% idle (90% utilized) | 30% idle (70% utilized) | 50% idle (50% utilized) |
|---|---|---|---|
| A100, Lambda | $1.99/hr | $2.56/hr | $3.58/hr |
| A100, RunPod Community | $1.66/hr | $2.13/hr | $2.98/hr |
| A100, Vast.ai on-demand | $1.33/hr | $1.71/hr | $2.40/hr |
| A100, Vast.ai interruptible | $0.67/hr | $0.86/hr | $1.20/hr |
| H100, Lambda | $2.77/hr | $3.56/hr | $4.98/hr |
| H100, RunPod Secure | $3.21/hr | $4.13/hr | $5.78/hr |
| H100, Vast.ai on-demand | $1.63/hr | $2.10/hr | $2.94/hr |
At 50% idle, every provider's effective rate roughly doubles, and the gap between Vast.ai's on-demand rate and RunPod Secure's rate on H100 widens from $1.42 an hour paid-for-nothing-adjusted to $2.84 an hour, because the cheaper base rate is also cheaper to leave idle. Per-second or per-minute billing (RunPod bills by the minute; Vast.ai by the second) reduces this penalty versus a provider with an hourly minimum, but none of these three eliminates it: an idle GPU that is still reserved is still billed.
Storage overhead
Persistent volume storage was reported at roughly $0.07 per GB per month on RunPod's network volumes. For an ILLUSTRATIVE 500 GB persistent volume (model weights, checkpoints, datasets), that is about $35 a month, independent of GPU hours consumed. At 100 GPU hours a month this can be a meaningful share of the bill on a cheap GPU class; at 10,000 GPU hours it is negligible. Exact storage and egress pricing for Lambda and Vast.ai were not captured in the sources reviewed (QUOTE-ONLY / UNKNOWN); confirm directly before committing to a large persistent dataset.
Interruption penalty as an illustrative engineering cost
Community Cloud and interruptible/spot listings carry a real risk: a host can go offline or reclaim capacity with limited notice. If an interrupted job requires an ILLUSTRATIVE 30 minutes of engineering time to detect, restart from the last checkpoint, and verify state, at a loaded $75 an hour that is $37.50 per interruption. Formula: interruption cost per 1,000 GPU hours = interruptions per 1,000 hours × $37.50. At one interruption per 200 GPU hours (ILLUSTRATIVE, varies enormously by host reliability and job length), that is 5 interruptions per 1,000 hours, or $187.50 added to the raw compute cost, before accounting for any lost partial-epoch compute time. This is a threshold for deciding whether the savings from interruptible pricing are worth the operational overhead, not a claim about how often interruptions actually occur on any specific host.
Sensitivity
- Which source's rate is real. Lambda's H100 rate spans $2.49–$3.29 across sources; RunPod's Community A100 spans $1.49–$1.89. Confirm the current rate directly before budgeting a large job.
- Utilization. Moving from 90% to 50% utilization roughly doubles effective cost on every provider and GPU class.
- Interruption frequency. A workload with frequent checkpointing and fast restart tolerates interruptible pricing well; a long uncheckpointed training run does not.
- GPU class substitution. An RTX 4090-class GPU costs roughly a fifth of an H100-class GPU per hour on RunPod and Vast.ai; for workloads that fit in 24GB of VRAM, this is a much larger lever than choosing between providers on the same GPU class.
Budgeting traps
- Comparing community-marketplace and enterprise-secure prices as if they were the same product. A $1.20 Vast.ai A100 and a $1.89 RunPod Secure A100 carry different reliability guarantees, not just different prices.
- Ignoring idle time in a reserved-but-not-busy pod. The effective cost per useful hour can be double the listed rate at moderate idle levels.
- Storage as an afterthought. Persistent volumes bill continuously regardless of GPU usage.
- Assuming marketplace rates are stable. Vast.ai's prices are reported to update live based on supply and demand; a rate checked today may not hold next week.
What to ask before you buy
Ask whether the listed rate is community-marketplace or provider-owned infrastructure, what the storage and egress rates are, and what the provider's interruption or eviction notice period is. Then model your own idle time and interruption frequency into the effective cost before comparing providers on the sticker rate alone.
ARTICLE 8