Vast.ai is a GPU marketplace where prices are not fixed, they are set by individual hosts and move with supply. An H100 PCIe on an unverified host might list at $0.90/hr on a quiet Tuesday and disappear entirely when demand picks up. Datacenter-verified hosts offer more stability but push prices to $1.50-$2.27/hr, which is a different calculation entirely. If you are trying to understand whether Vast.ai's price ranges hold up in 2026 or how they compare to managed alternatives, this post breaks down the actual numbers by host tier, covers H200 and B200 availability, and puts Vast.ai's rates next to Spheron's live on-demand and spot pricing.
TL;DR: Vast.ai vs Spheron H100, H200, and B200 (Jul 2026)
| GPU | Vast.ai (Unverified) | Vast.ai (Verified DC) | Spheron On-Demand | Spheron Spot |
|---|---|---|---|---|
| H100 PCIe | ~$0.90-$1.60/hr | ~$1.50-$2.27/hr | $2.51/hr | N/A |
| H100 SXM | Rare | ~$1.80-$2.50/hr | $5.07/hr | $2.91/hr |
| H200 SXM | Rare | ~$3.50-$5.50/hr | $3.70/hr | $1.76/hr |
| B200 | Not available | Very limited, ~$6-$12/hr | N/A (spot only) | $5.34/hr |
Vast.ai unverified H100 PCIe can look attractively cheap, but the verified tier lands at $1.50-$2.27/hr. Spheron H100 PCIe at $2.51/hr on-demand sits just above Vast.ai's verified tier range and carries no host reliability risk. Spheron H100 SXM5 spot at $2.91/hr sits above Vast.ai's verified SXM range but comes with platform-managed reliability rather than individual host gambles. Spheron H200 at $3.70/hr on-demand is priced near the low end of Vast.ai's verified H200 range.
Pricing fluctuates based on GPU availability. The prices above are based on 01 Jul 2026 and may have changed. Check current GPU pricing → for live rates.
Vast.ai Pricing Model Explained
Vast.ai is not a cloud provider. It is a marketplace that connects GPU owners with GPU renters. Every listing comes from an individual host: a data center operator, a mining rig owner, a hobbyist with a server rack. The price you see reflects what that specific host decided to charge, not a platform rate.
This creates a few things you need to understand before comparing Vast.ai prices to managed providers.
Host categories. Vast.ai divides hosts into unverified community hosts and datacenter-verified hosts. Unverified hosts have no formal infrastructure vetting. They might be someone running GPUs from their home or a small colocation space. Datacenter-verified hosts have submitted documentation to Vast.ai and typically operate from proper data center facilities. The reliability spread between these two tiers is significant.
The interruptible vs on-demand distinction. Vast.ai listings are labeled either interruptible or on-demand. Interruptible means the host can evict you at any time, similar to spot instances on managed clouds. On-demand on Vast.ai means you can rent for as long as the host is online and the machine is available, but the platform does not guarantee the machine stays online. If the host takes their hardware offline, your workload stops regardless of what the listing called itself. This is a fundamentally different definition than on-demand on AWS, GCP, or managed providers like Spheron where the platform guarantees availability.
DLPerf scoring. Vast.ai shows a DLPerf score for each host, which measures verified deep learning throughput. Higher DLPerf generally correlates with better actual training performance, but the score reflects a historical test, not current hardware state.
Price volatility. Because supply is host-driven, prices change based on how many hosts are listing on a given day. A GPU model that has 40 listings today might have 12 tomorrow if hosts take their machines offline. This creates volatility that does not exist on managed clouds with fixed pricing.
Vast.ai H100 PCIe Pricing Per Hour
H100 PCIe makes up the majority of Vast.ai's H100 supply. Most community hosts running data center GPUs are running PCIe configurations, not the server-grade SXM5 hardware that requires a full HGX tray.
| Host Tier | Price Range $/hr | Reliability | Notes |
|---|---|---|---|
| Unverified community | ~$0.90-$1.60 | Variable | No infrastructure vetting |
| Datacenter-verified | ~$1.50-$2.27 | Higher | Formal DC documentation |
The spread within each tier is real. On a given day you might find a verified host at $1.52/hr and another at $2.27/hr for the same H100 PCIe. The difference comes from location, DLPerf score, RAM configuration, and the host's own pricing strategy. Filtering by verified and sorting by DLPerf gives you the most reliable subset, but that subset also commands the highest prices.
H100 PCIe on Vast.ai has 80GB HBM2e VRAM, the same as the SXM variant in terms of memory capacity. The architectural difference is in the interconnect: PCIe uses the host's PCIe bus for GPU communication, while SXM uses NVLink with significantly higher GPU-to-GPU bandwidth. For single-GPU inference workloads, this distinction does not matter. For distributed training across multiple H100s, SXM is meaningfully faster. The form factor implications for training and inference are covered in detail in our H100 PCIe vs SXM form factor guide.
Spheron H100 PCIe on-demand at $2.51/hr sits just above Vast.ai's verified tier range. The difference: Spheron's rate is from a managed platform, not a host lottery. You get per-minute billing, full root access to a VM rather than a container, and no restart risk when a host takes their hardware offline.
Pricing fluctuates based on GPU availability. The prices above are based on 01 Jul 2026 and may have changed. Check current GPU pricing → for live rates.
Vast.ai H100 SXM Pricing Per Hour
SXM5 listings are rarer on Vast.ai. The SXM5 form factor requires a full 8-GPU HGX tray in a server chassis, which means the only hosts with SXM5 hardware are actual data centers or well-funded operators with HGX systems. This pushes SXM5 availability on Vast.ai toward the verified datacenter tier almost exclusively.
When SXM5 hosts list on Vast.ai, pricing runs approximately $1.80-$2.50/hr per GPU depending on the host. At 8 GPUs per node, you are typically renting the full node rather than a single GPU, which means the floor for an 8-GPU SXM5 job is roughly $14-$20/hr total. Some hosts offer per-GPU slots if the node is not fully subscribed, but availability of individual SXM5 GPU slots varies.
NVLink within the node is typically functional on verified SXM5 hosts, but verify this before running distributed training workloads. The node interconnect specs should appear in the host listing details.
Spheron H100 SXM5 on-demand at $5.07/hr is higher than Vast.ai's verified SXM range. Spheron H100 SXM5 spot at $2.91/hr also sits above Vast.ai's verified tier, though it is managed at the platform level rather than dependent on an individual host's uptime decisions. For fault-tolerant training runs with checkpointing, Spheron SXM5 spot is the more cost-effective path with lower reliability risk than Vast.ai verified hosts.
Vast.ai H200 Pricing Per Hour
H200 supply on Vast.ai is thin. H200 hardware only became widely available in 2025, and the bulk of it went to hyperscalers and large managed clouds first. Community hosts running H200 are still rare, and verified datacenter hosts with H200 capacity on Vast.ai represent a small subset of the H200 market.
When H200 is available on Vast.ai's verified hosts, pricing runs approximately $3.50-$5.50/hr per GPU. The range reflects the same host-by-host variability that applies to H100, amplified by lower overall supply. Unverified hosts occasionally list below $3.00/hr, but the reliability concerns that apply to unverified H100 hosts apply even more acutely to H200, where the hardware is newer and supply disruptions are more common.
Spheron H200 SXM5 on-demand at $3.70/hr is just above the floor of Vast.ai's verified H200 range. Spheron H200 SXM5 spot at $1.76/hr is well below any verified Vast.ai H200 host listing. For teams that need H200 for large-context inference or high-memory training runs, Spheron's pricing on H200 is currently more competitive than Vast.ai's verified tier. See H200 GPU rental on Spheron for current availability and configuration options.
Pricing fluctuates based on GPU availability. The prices above are based on 01 Jul 2026 and may have changed. Check current GPU pricing → for live rates.
Vast.ai B200 Pricing Per Hour
B200 availability on Vast.ai is very limited as of mid-2026. The B200 GPU requires Blackwell-generation HGX hardware in a full server chassis, and the installed base in community-accessible data centers is still small. B200 listings do not appear on Vast.ai with regularity.
When B200 does appear on Vast.ai, expect pricing in the $6-$12/hr range for verified datacenter hosts. Supply is intermittent enough that treating Vast.ai as a reliable B200 source would be a planning mistake. If a team needs B200 for production inference or training and cannot tolerate availability uncertainty, a managed provider is the practical choice.
Spheron B200 SXM6 is currently available as spot at $5.34/hr per GPU. There is no on-demand offer right now, so treat B200 capacity as interruptible. The spot rate undercuts the floor of Vast.ai's occasional B200 listings when they do appear, though spot can be reclaimed without notice. For context on B200 pricing across providers, see our Nvidia B200 cloud pricing 2026 guide and the B200 GPU rental page.
Vast.ai Hidden Costs and Reliability Trade-offs
The per-hour rate on a Vast.ai listing is not your actual cost. Four factors consistently push the real cost above the headline number.
Storage billing. Vast.ai charges for allocated disk storage whether or not your instance is running. If you allocate 200GB for a training dataset and pause your instance for a weekend, storage fees accumulate. At approximately $0.10-$0.15/GB/month, 200GB paused for 10 days costs roughly $6.67-$10 in storage alone (200GB x $0.10/GB/month = $20/mo, prorated to 10 days). Scale this across multiple instances paused between jobs and the billing adds up in ways the hourly GPU rate does not reflect.
Host reliability. Unverified hosts can go offline without warning, taking your running workload with them. Even verified hosts have maintenance windows and hardware issues. A host with 95% reliability sounds solid, but across a 72-hour training run the math is less comfortable: modeling reliability as 0.95 per hour, the probability of at least one interruption over 72 hours is approximately 1 - 0.95^72 using a simplified hourly model, which works out to approximately 97%. For runs where checkpointing is not enabled or checkpoints are infrequent, an interruption means lost compute and real money. Verified datacenter hosts improve the picture significantly but do not eliminate it.
Bandwidth costs. Vast.ai's bandwidth policy is not standardized across hosts. Some hosts charge for network egress; others include it. Always check the host listing's bandwidth policy before starting a large data transfer workload. Missing a per-GB egress charge on a host that transfers 500GB of training data changes the economics of that instance entirely.
Per-hour billing rounding. Vast.ai bills in hourly increments on most listings. A 43-minute inference evaluation job costs a full hour. Run 10 such jobs in a day on 4 GPUs and you pay for 40 GPU-hours while consuming 28.7. At $1.50/hr per GPU that billing rounding adds $17/day in waste. Spheron's per-minute billing eliminates this entirely.
Full details on Spheron's billing model are at docs.spheron.ai.
Vast.ai Interruptible vs On-Demand: What the Labels Actually Mean
The terminology mismatch between Vast.ai and managed cloud providers is a source of confusion that affects cost planning.
On a managed provider like AWS, GCP, or Spheron, "on-demand" means the platform guarantees the instance is available when you request it and stays running until you stop it. The platform's infrastructure backs that guarantee, not an individual machine's uptime.
On Vast.ai, "on-demand" means something different. An on-demand listing means the host is currently offering the machine at the listed price and will not actively evict you, but the machine stays available only as long as the host keeps it online. If the host performs maintenance, moves the hardware, or simply takes the machine offline for any reason, your "on-demand" instance stops. There is no platform-level guarantee that the machine remains available.
Vast.ai's "interruptible" listings can be evicted with short notice for any reason, similar to spot instances on hyperscalers, though without the same degree of advance warning infrastructure.
Spheron's on-demand tier operates on platform-managed bare-metal from vetted data center partners, meaning the platform is responsible for uptime rather than an individual host. Spheron's spot tier offers lower pricing at the cost of explicit interruptibility, managed at the platform level. The distinction matters for production workloads: a Vast.ai "on-demand" instance has host-level reliability; a Spheron on-demand instance has platform-level reliability.
Vast.ai vs Spheron: Direct Comparison
| Metric | Vast.ai (Verified DC) | Spheron |
|---|---|---|
| H100 PCIe on-demand | ~$1.50-$2.27/hr | $2.51/hr |
| H100 SXM on-demand | ~$1.80-$2.50/hr | $5.07/hr |
| H100 SXM spot | Rare / not reliably available | $2.91/hr |
| H200 on-demand | ~$3.50-$5.50/hr | $3.70/hr |
| B200 (spot only) | Limited, ~$6-$12/hr | $5.34/hr spot |
| Billing granularity | Per hour | Per minute |
| Host control | Individual hosts | Managed platform |
| Root access | Container (restricted) | Full VM/bare-metal |
| Idle storage billing | Yes | No |
| Egress fees | Host-dependent | Included |
| Uptime SLA | None | Platform-managed |
The structural difference is that Vast.ai is a marketplace where you rent from individual hosts and Spheron is a managed cloud that aggregates vetted bare-metal capacity from data center partners worldwide. Spheron controls reliability at the platform level. Vast.ai's verified tier reduces but does not eliminate host-level risk. For the full breakdown of what this means for VM access, container restrictions, and production workload suitability, see Spheron vs Vast.ai.
Spheron Live Rates (Jul 2026):
| GPU | On-Demand $/hr | Spot $/hr | Notes |
|---|---|---|---|
| H100 SXM5 | $5.07 | $2.91 | Per-minute billing |
| H100 PCIe | $2.51 | N/A | Per-minute billing |
| H200 SXM5 | $3.70 | $1.76 | Per-minute billing |
| B200 SXM6 | N/A | $5.34 | Spot only; reclaimable |
No commitment required for any tier. SSH setup takes under 2 minutes. Full deployment documentation is at docs.spheron.ai.
Pricing fluctuates based on GPU availability. The prices above are based on 01 Jul 2026 and may have changed. Check current GPU pricing → for live rates.
When to Use Vast.ai vs a Managed Bare-Metal Cloud
Vast.ai does some things well, and it is worth being direct about when it makes sense.
Vast.ai is a good fit for:
Short experiments and prototyping where the cheapest listed price matters more than reliability. If a job fails, you restart it. Consumer GPU workloads (RTX 4090) at very low rates for development and testing. Hobbyist projects where occasional interruptions are acceptable overhead. Situations where you can actively monitor the host reliability score and pick verified hosts with 99%+ uptime history.
A managed provider like Spheron is the better call when:
Training runs exceed 4 hours where an interruption means losing meaningful compute cost. Production inference APIs that need uptime and consistent latency. Workloads requiring kernel-level access, custom CUDA drivers, or system-level configuration. Vast.ai containers block this; Spheron VMs give you full root access. Teams that need predictable billing for budget planning. Teams where per-hour billing rounding on short jobs adds real cost. Workloads that need H200 at competitive rates, where Spheron's on-demand H200 SXM5 access currently beats Vast.ai verified host pricing.
For teams that specifically need on-demand H100 access without host reliability uncertainty, on-demand H100 access on Spheron with per-minute billing is the direct alternative.
Vast.ai vs Other GPU Clouds: Provider Comparison
| Provider | H100 On-Demand $/hr | Spot/Preemptible $/hr | Billing | Notes |
|---|---|---|---|---|
| Vast.ai (Verified DC) | ~$1.50-$2.27 (PCIe) | Interruptible ~$0.90-$1.60 | Per hour | Marketplace, host-dependent |
| Spheron (SXM5) | $5.07 | $2.91 | Per minute | Managed platform |
| Spheron (PCIe) | $2.51 | N/A | Per minute | Managed platform |
| RunPod (Secure Cloud) | ~$3.29 | Available | Per minute | |
| Lambda Labs | $3.29-$3.99 | None | Per hour | No spot option |
| Nebius | $3.85 | $2.15 (preemptible) | Per hour | EU-focused |
For the full provider breakdown including AWS, GCP, Azure, CoreWeave, and others, see:
- RunPod H100 pricing 2026
- Lambda Cloud H100 pricing 2026
- Nebius H100 and H200 pricing 2026
- Full GPU cloud pricing comparison 2026
- Vast.ai alternatives
Vast.ai verified datacenter H100 PCIe at $1.50-$2.27/hr is among the cheapest verified-tier H100 options on this list. The trade-off is host-level reliability rather than platform-level reliability, per-hour billing, container-only access, and idle storage fees. Spheron H100 PCIe at $2.51/hr sits just above Vast.ai's verified range with platform-managed reliability and per-minute billing. Whether that trade-off is worth it depends entirely on your workload's tolerance for host interruptions and job failures.
Vast.ai's verified H100 SXM range at $1.80-$2.50/hr actually undercuts Spheron SXM5 on-demand at $5.07/hr significantly. Spheron SXM5 spot at $2.91/hr is more competitive for fault-tolerant training workloads. Teams running SXM workloads that cannot tolerate interruptions should compare total cost including restart overhead before assuming Vast.ai's lower listed rate is the cheaper option.
Vast.ai's H200 verified listings at $3.50-$5.50/hr are generally more expensive than Spheron's H200 SXM5 at $3.70/hr on-demand or $1.76/hr spot. For H200 workloads, Spheron is the more cost-effective managed option.
The picture for Vast.ai is straightforward: cheapest verified H100 PCIe access in the market at the cost of host-level reliability and container restrictions. For teams where those costs are acceptable, Vast.ai verified hosts are a legitimate option. For teams where they are not, the managed tier is the better path.
Vast.ai's marketplace gives you the cheapest listed prices, but verified-host rates land close to Spheron's on-demand tier with none of Spheron's reliability guarantees or per-minute billing. If you need predictable cost and uptime, Spheron's H100 and H200 instances are the straightforward alternative.
Quick Setup Guide
Go to vast.ai and use the search filter. Set GPU type to H100, filter by 'verified' to see datacenter-backed hosts. Note the spread between the cheapest and most expensive listing - the spread reflects supply volatility. Look at DLPerf (deep learning performance score) and reliability scores, not just price.
Take the listed $/hr and add estimated storage cost ($0.10-0.15/GB/month, divided into hourly). Factor in the host reliability score: a host with 90% reliability means 10% of your rented hours could be interrupted with potential job loss. For training runs over 10 hours, always verify the host has at least 99% reliability score.
Visit spheron.network/pricing/ and filter by H100. Check both on-demand and spot rates. Spheron H100 PCIe on-demand is available at $2.51/hr with per-minute billing, no host lottery, and no idle storage fees. H100 SXM5 on-demand at $5.07/hr or spot at $2.91/hr for fault-tolerant workloads.
Sign in at app.spheron.ai, select H100 from the GPU catalog, choose on-demand or spot, and deploy. SSH root access in under 2 minutes. Per-minute billing starts on first use, not on provisioning request.
Frequently Asked Questions
Vast.ai H100 pricing varies by host tier: unverified hosts list H100 from roughly $0.90-$1.60/hr, while datacenter-verified hosts run $1.50-$2.27/hr. These are marketplace spot rates - the actual price fluctuates with supply. Spheron lists H100 PCIe at $2.51/hr on-demand and H100 SXM5 at $5.07/hr on-demand, billed per minute with no host variability.
Vast.ai does not offer traditional on-demand pricing with guaranteed availability. All listings are marketplace-driven and interruptible. The platform distinguishes between 'interruptible' (spot-equivalent) and 'on-demand' offers, but 'on-demand' on Vast.ai still means a specific host's machine and can be unavailable when that host is offline. This differs from managed providers where on-demand means guaranteed availability from the platform.
Vast.ai H200 listings are sparse as of 2026. When available, H200 on verified hosts ranges from approximately $3.50-$5.50/hr depending on the host, with unverified hosts occasionally listing below $3.00/hr. Spheron H200 SXM5 on-demand is $3.70/hr with per-minute billing and no host lottery.
Vast.ai's headline $/hr is not your total cost. Watch for: storage charged per GB per hour even when stopped or paused; bandwidth usage on some hosts; potential restart overhead when a host goes offline mid-job; and reliability variance between hosts. Unverified hosts can disappear with no compensation. Spheron bills compute only - no idle storage fees, no bandwidth add-ons.
Vast.ai unverified hosts can list H100 below $1.00/hr, which is cheaper than Spheron's PCIe on-demand rate. However, verified datacenter hosts on Vast.ai run $1.50-$2.27/hr - comparable to or slightly below Spheron's H100 PCIe at $2.51/hr on-demand. When factoring in restart overhead, idle storage billing, and host reliability risk, Spheron's total cost is often competitive or lower for production workloads.