RunPod and Vast.ai top most "cheapest GPU cloud" searches in 2026, and both genuinely undercut hyperscaler pricing by 3-6x. But "cheap" means something different on each platform. RunPod prices GPUs by platform tier: pick Community Cloud, Secure Cloud, or Serverless and you get a fixed rate. Vast.ai lets individual hosts set their own price, so the number you see is one listing among thousands. This gpu cloud pricing comparison puts RunPod's and Vast.ai's live hourly rates side by side by GPU tier, so you can see what you're actually paying for beyond the headline number.
Think of this post as the lookup table: rates broken out by GPU tier with the storage and billing fine print, nothing else. For the full platform-specific breakdowns, see our RunPod H100 pricing guide and Vast.ai pricing guide covering H100, H200, and B200 by host tier. If you want the narrative version, including a worked 48-hour, 8x H100 total-cost-of-a-finished-run calculation and a line-by-line reliability record for each platform, our RunPod vs Vast.ai reliability and TCO breakdown covers that ground. For the wider market beyond these two platforms, see the broader GPU cloud pricing comparison covering 5+ providers.
How RunPod and Vast.ai Price GPUs: Platform Tiers vs Marketplace Bidding
The structural difference between these two platforms explains most of the pricing spread you'll see below. RunPod is a managed platform with a rate card. Vast.ai is a peer-to-peer marketplace with no rate card at all, just thousands of individual listings.
RunPod's Community Cloud, Secure Cloud, and Serverless Tiers
RunPod splits inventory into three tiers, according to its pricing page. Community Cloud runs on third-party host hardware with no formal SLA and the lowest rates: RTX 4090 at $0.34/hr, A100 PCIe at $1.19/hr, H100 PCIe at $1.99/hr. Secure Cloud runs on RunPod-operated data centers with an SLA and costs more across the board: RTX 4090 at $0.69/hr (roughly double Community), H100 PCIe at $2.89/hr, H100 SXM at $2.99/hr. B200 is the one GPU where both tiers converge at $5.89/hr.
Serverless is a different billing model entirely: it scales to zero and bills per second of active worker time, which makes it cheaper for spiky traffic and more expensive per active hour than a comparable on-demand pod. It's not part of the side-by-side table below since it isn't priced as a straight $/hr rental.
Storage is billed separately on both tiers: container disk runs $0.10/GB/month, network storage is $0.05-$0.07/GB/month depending on volume, and high-performance network storage runs $0.14/GB/month.
Vast.ai's Host-Set Marketplace Pricing: Unverified vs Datacenter-Verified
Vast.ai isn't a cloud provider in the traditional sense. Every listing is set by whoever owns that specific machine, and the platform's own pricing page markets three instance types: On-Demand ("guaranteed uptime, best for production"), Interruptible ("50%+ cheaper, best for batch training"), and Reserved (up to 50% off for 1-6 month commitments). All three bill per second with no hourly rounding.
What Vast.ai's own marketing doesn't emphasize, and what GPUnex's independent 2026 review does, is the split between unverified community hosts and datacenter-verified hosts. Unverified listings are the cheapest: H100 SXM from roughly $0.90/hr, A100 80GB from $0.50/hr, RTX 4090 from $0.34/hr. Verified datacenter hosts, submitted with documentation and typically running in real facilities, cost more, $1.50-$1.87/hr for H100. The marketplace itself is enormous: 17,000+ GPUs across 1,400+ independent providers in 500+ locations worldwide, per GPUnex, which is what keeps the cheap end of every price range populated.
RunPod vs Vast.ai GPU Cloud Pricing Comparison: Hourly Rates by GPU Tier
Here's the direct answer: on sticker price, Vast.ai's unverified tier is cheapest for every GPU tracked below, RunPod's Community Cloud runs a close second, and the premium you pay for RunPod Secure Cloud or a verified Vast.ai host buys you a platform-backed or reputation-backed guarantee instead of raw hardware access. We pulled these numbers from RunPod's and Vast.ai's public rate cards plus Spheron's live marketplace API on 11 Aug 2026.
RTX 4090 Hourly Rates by Platform and Tier
| Platform | Rate | Notes |
|---|---|---|
| RunPod Community Cloud | $0.34/hr | Third-party host, no SLA |
| RunPod Secure Cloud | $0.69/hr | RunPod-operated, SLA-backed |
| Vast.ai marketplace | $0.34-$0.50/hr | Lower end is unverified; higher end is verified |
| Spheron on-demand | $0.58/hr | Per-minute billing, no idle storage fee |
A100 80GB Hourly Rates by Platform and Tier
| Platform | PCIe Rate | SXM Rate | Notes |
|---|---|---|---|
| RunPod Community Cloud | $1.19/hr | $1.39/hr | Third-party host |
| RunPod Secure Cloud | $1.39/hr | $1.49/hr | SLA-backed |
| Vast.ai marketplace | $0.50-$0.80/hr | $0.50-$0.80/hr | GPUnex doesn't split by form factor |
| Spheron on-demand | $1.43/hr | $1.82/hr | No spot currently listed on PCIe |
| Spheron spot | N/A | $1.15/hr | Reclaimable, checkpoint-friendly |
H100 PCIe and SXM Hourly Rates by Platform and Tier
| Platform | PCIe Rate | SXM Rate | Notes |
|---|---|---|---|
| RunPod Community Cloud | $1.99/hr | $2.69/hr | Third-party host |
| RunPod Secure Cloud | $2.89/hr | $2.99/hr | SLA-backed |
| Vast.ai unverified | ~$0.90/hr (SXM low end) | ~$0.90/hr | Cheapest listings, variable uptime |
| Vast.ai verified datacenter | $1.50-$1.87/hr | $1.50-$1.87/hr | Documented facilities |
| Spheron on-demand | $2.64/hr | $3.98/hr | SXM5 supply is tight right now |
| Spheron spot | $2.20/hr | $2.91/hr | Reclaimable |
Pricing fluctuates based on GPU availability. The prices above are based on 11 Aug 2026 and may have changed. Check current GPU pricing → for live rates.
Worth calling out: Spheron's H100 SXM5 on-demand rate is higher than RunPod Secure Cloud's right now, a reversal from earlier in 2026 when Spheron undercut RunPod across the board. GPU pricing on any marketplace-adjacent platform moves with supply, and SXM5 inventory is the tightest tier at the moment. The PCIe tier still favors Spheron, and spot pricing brings SXM5 back in line with RunPod's rate for workloads that can checkpoint through a reclamation.
Reliability and Interruption Risk: Marketplace vs Dedicated Capacity
The rate you see on either platform isn't the rate you pay if the job gets interrupted. That's the part the tables above can't show.
What "On-Demand" Actually Means on a Peer-to-Peer Marketplace
On a managed cloud, "on-demand" means the provider guarantees the instance stays available until you stop it. On Vast.ai, "on-demand" means the specific host currently won't actively evict you, which isn't the same promise. As GPUnex puts it: "the listed hourly rate assumes your instance runs uninterrupted from start to finish. On a decentralized marketplace with thousands of independent providers, that assumption does not always hold."
GPUnex models the real cost as listed rate plus restart overhead plus lost compute, and finds unverified hosts run 20-40% above sticker price once that's priced in, versus 3-13% for verified datacenter hosts. Their worked example: a $1.00/hr H100 landed closer to $1.29/hr in practice after accounting for a disconnection and restart.
RunPod isn't immune to the same failure mode. A 2026 independent review tracked 227+ outages across RunPod over nine months, describing pods that fail to start or crash mid-job while billing keeps running, and standard self-serve plans carry no formal SLA; that coverage requires committing to RunPod's Growth tier. Neither platform's rate card tells you this, which is exactly why the number on the pricing page is a starting point, not the answer. For the full math on what an interruption costs a multi-hour training run, including a worked 48-hour, 8x H100 total-cost-of-a-finished-run calculation, see the RunPod vs Vast.ai reliability and TCO breakdown.
Where Spheron Fits for Teams That Need Predictable Uptime
Spheron aggregates bare-metal capacity from 5+ providers under one platform, so you're not betting on a single host's uptime history the way Vast.ai's marketplace model requires, and you're not choosing between a no-SLA self-serve tier and a $50,000 enterprise commitment the way RunPod's structure forces. H100 GPU rental on Spheron starts at $2.64/hr on-demand with $2.20/hr spot on the PCIe variant, and A100 GPU rental runs $1.43-$1.82/hr on-demand depending on form factor, all billed per minute with no idle storage fees and no host lottery.
Unlike Vast.ai's containerized instances, Spheron rentals give you full VM or bare-metal root access, which matters if your workload needs custom CUDA kernels or driver-level tuning. For the deeper architecture comparisons, see Spheron vs RunPod and Spheron vs Vast.ai. Full docs on Spheron's billing model and deployment flow are at docs.spheron.ai.
The GPU Cloud Pricing Comparison Verdict: Dev/Test vs Production Inference
The right platform depends on what happens if a run gets interrupted, not just what the hourly number says.
- Quick experiments and dev/test: Vast.ai's unverified tier or RunPod Community Cloud wins on raw cost. An interruption costs you minutes, not a lost multi-hour run, so the reliability tax barely matters.
- Batch training with checkpointing: Vast.ai's Interruptible instances or Spheron's spot pricing both work here. Budget for the effective markup GPUnex documents rather than the sticker rate.
- Production inference APIs: this is where sticker price stops being the deciding factor. RunPod Secure Cloud, Vast.ai's verified datacenter hosts, or Spheron's on-demand tier are the three realistic options, and the choice comes down to whether you need a formal SLA (RunPod, at a real cost premium for the Growth tier), a reputation score you vet yourself (Vast.ai verified), or platform-managed bare-metal capacity with per-minute billing and no idle fees (Spheron).
- Multi-day training runs: interruption risk compounds with runtime. Stick to verified or SLA-backed capacity regardless of platform; our spot GPU training resilience and checkpointing guide covers how to shrink the loss window if you do run on interruptible capacity.
If neither RunPod's tier structure nor Vast.ai's marketplace model fits, the RunPod alternatives roundup and Vast.ai alternatives roundup cover the wider field, and the top 10 cloud GPU providers comparison ranks the market on more than just price.
If you're done tracking down individual hosts to avoid a mid-run interruption, Spheron's on-demand bare-metal GPUs give you fixed pricing and per-minute billing without the marketplace lottery.
Frequently Asked Questions
By sticker price, Vast.ai's unverified tier is the floor: RTX 4090 from $0.34/hr, A100 80GB from $0.50/hr, and H100 SXM as low as $0.90/hr, per Vast.ai's own listings and GPUnex's 2026 marketplace review. RunPod's Community Cloud sits close behind at the same $0.34/hr for RTX 4090. Neither number is your real cost: GPUnex found unverified Vast.ai hosts run 20-40% above listed price once restarts and downtime are priced in.
Community Cloud is cheaper on every GPU RunPod publishes: RTX 4090 is $0.34/hr on Community versus $0.69/hr on Secure Cloud, and H100 PCIe is $1.99/hr versus $2.89/hr, per RunPod's pricing page. The gap is the SLA and the data center guarantee. Community Cloud runs on third-party host hardware with no formal uptime commitment, the same structural tradeoff as Vast.ai's unverified tier.
Not exactly. Vast.ai's own pricing page markets on-demand instances as 'guaranteed uptime,' but that guarantee is host-specific, not platform-wide. GPUnex's 2026 review models the real cost as listed rate plus restart overhead plus lost compute, and finds unverified hosts run 20-40% over sticker price versus 3-13% for datacenter-verified hosts. A $1.00/hr H100 in their worked example landed closer to $1.29/hr in practice.
Spheron's live on-demand H100 PCIe rate is $2.64/hr, below RunPod Secure Cloud's $2.89/hr and above Vast.ai's verified-host ceiling of roughly $1.87/hr. On A100 80GB, Spheron on-demand runs $1.43/hr (PCIe) to $1.82/hr (SXM4), with spot from $1.15/hr. Spheron doesn't try to beat Vast.ai's unverified floor: it competes on bare-metal access and per-minute billing with no idle storage fees, positioned between marketplace bidding and fixed-tier platform pricing.
