Runpod H100 Pricing 2026: Per-Hour and Serverless Cost vs Spheron

Runpod lists H100 GPUs across three distinct pricing tiers: Community Cloud (community-hosted, variable), Secure Cloud (Runpod-operated, SLA-backed), and Serverless (scale-to-zero, per-second billing). Each tier targets a different tradeoff between cost, reliability, and latency. This post covers exact per-hour rates for each tier, the per-second serverless math, hidden storage fees, and a direct comparison against Spheron, Lambda Labs, and CoreWeave. For where Runpod's rates sit in the full market, including hardware purchase costs, see the NVIDIA H100 price overview.

Runpod H100 Pricing: Three Tiers Explained

Runpod's tiered model puts different GPU inventory under different guarantees. Understanding which tier fits your workload matters before comparing raw $/hr figures.

Community Cloud connects you to third-party hosts who rent their hardware through Runpod's marketplace. These are the cheapest listings, roughly $1.99/hr for H100 80GB PCIe and $2.69/hr for SXM5, but availability is host-dependent and hardware condition varies. If a host goes offline, your pod can be interrupted with limited recourse.

Secure Cloud is Runpod's own data center capacity. Hardware is Runpod-owned, availability is more predictable, and uptime guarantees apply. Prices are higher: H100 80GB (PCIe) runs around $2.89/hr and SXM5 variants are approximately $2.99/hr. Most production inference deployments land here.

Serverless is not an always-on instance. You deploy a container template and Runpod spins up H100 pods on demand, billing per second of active execution. When traffic drops to zero, you pay zero. This is ideal for APIs with unpredictable or bursty traffic patterns, though the per-active-hour rate is higher than on-demand. Runpod's tiered model isn't the only one that hides real numbers behind qualifiers: how Denvr Dataworks prices GB200 access is worth checking too, since its H100 and A100 rates are published but H200, GB200, and B200 all route to a sales quote.

Runpod H100 SXM5 vs PCIe Per-Hour Rates

Most H100 listings on Runpod show "H100 80GB" without specifying interconnect. The majority are PCIe models. SXM5 variants with NVLink are listed separately and typically carry a small premium due to higher memory bandwidth (3.35 TB/s vs 2 TB/s) and better multi-GPU scaling.

For single-GPU inference where bandwidth headroom is not critical, the PCIe variant is fine. For distributed training or large model inference requiring multi-GPU NVLink bandwidth, the SXM5 variant is the right call.

Runpod Serverless GPU Pricing: Per-Second Billing and Cold-Start Math

Runpod Serverless charges for H100 access per second of active execution, which works out to roughly $4.55/hr (about $0.00126/sec) while a pod is running. That is higher than the on-demand rate, $2.89/hr for PCIe Secure Cloud, so on a per-active-hour basis Serverless is the more expensive option. What Serverless buys you is scale-to-zero: you pay nothing during idle periods. Whether it saves money comes down entirely to how much of the time your pod would otherwise sit idle.

Every time a pod initializes (a cold start), Runpod bills the full per-second rate for the initialization period. H100 pods typically take 20-60 seconds to become ready, depending on image size and container startup time. That adds $0.025-$0.076 per cold start, which is negligible individually but accumulates at scale.

The break-even against an always-on on-demand pod ($2.99/hr) lands around 66% utilization. If an on-demand pod would be busy less than roughly two-thirds of the time, Serverless is cheaper because you skip the idle hours. Above that, keep an on-demand pod running.

Serverless becomes less attractive when:

• Your API needs sub-100ms first-token latency (cold-start adds 20-60s of delay).
• You are running training jobs (continuous, no idle time, cold-starts irrelevant).
• Your workload is already above ~66% utilization around the clock, where an on-demand pod is cheaper.

For the inference throughput math behind token costs, see Ollama vs vLLM and KV cache optimization.

Hidden Costs on Runpod H100

The per-hour compute rate is not the total cost. Runpod charges separately for storage and network volumes:

• Network volumes: $0.07/GB/month for persistent NFS-backed storage under 1TB, dropping to $0.05/GB/month above 1TB.
• Container volumes: billed on the same $0.07/GB/month storage tier for pod-local disk.
• Idle pod billing: If you leave a pod running but idle (paused state), Runpod may still bill at a reduced storage-only rate depending on pod status.
• Cold-start billing: Serverless pods bill the full per-second rate (~$0.00126/sec) during initialization, before any work is processed.
• Community Cloud egress: Some hosts apply network transfer fees. The Runpod platform does not explicitly itemize this; check with individual hosts.

A realistic monthly bill for a team running 2x H100 SXM5 instances:

vs purely looking at the $2.99/hr headline: $4,305.60. Storage adds only ~0.5% in this example, but for storage-intensive pipelines (checkpointing, large datasets), it adds up faster.

Runpod H100 vs Spheron H100: Per-Hour and Per-Million-Token Cost

Spheron's live pricing for H100 as of 18 Jul 2026, fetched from the Spheron GPU rental marketplace:

For inference workloads, per-million-token cost matters more than raw $/hr. Using vLLM serving Llama 3.1 70B at approximately 1,200 tokens/sec on an H100 SXM5:

Cost per million tokens = (1,000,000 / 1,200 tokens/sec) / 3,600 * price

Spheron's on-demand H100 SXM5 at $3.90/hr is higher than Runpod Secure Cloud's $2.99/hr for the same variant. The comparison flips on PCIe: Spheron PCIe on-demand at $2.09/hr beats Runpod PCIe at $2.89/hr. On spot, Spheron SXM5 at $1.66/hr is the cheapest option in this comparison, dropping the per-million-token cost to $0.384 vs $0.692 at Runpod Secure Cloud. Check H100 on Spheron for current SXM5 and PCIe availability.

Pricing fluctuates based on GPU availability. The prices above are based on 18 Jul 2026 and may have changed. Check current GPU pricing → for live rates.

Runpod vs Lambda vs CoreWeave H100 Pricing Side-by-Side

Lambda Labs typically has strong availability but bills hourly and often requires on-demand reservations to guarantee capacity, a tradeoff our RunPod vs Lambda Labs pricing and reliability comparison breaks down in more detail alongside RunPod's own outage record. CoreWeave's on-demand rate is the highest in this table, and it's sold only as an 8-GPU HGX node with no single-GPU tier; see our full CoreWeave H100 and H200 pricing breakdown for the per-GPU math and how its two pricing pages quote different numbers. Spheron aggregates supply from 5+ providers, which generally keeps spot and on-demand availability higher than single-datacenter platforms. Nebius runs $3.85/hr for HGX H100 on-demand with preemptible instances at $2.15/hr; for the full breakdown of Nebius per-hour rates including H200 pricing and committed-use math, that post covers the details.

For a wider market survey covering AWS, Azure, Lambda, CoreWeave, and other providers, see GPU cloud pricing 2026.

Pricing fluctuates based on GPU availability. The prices above are based on 18 Jul 2026 and may have changed. Check current GPU pricing → for live rates.

When Runpod Serverless Beats On-Demand (and When It Doesn't)

Serverless wins for workloads where idle time is significant enough that scale-to-zero offsets the higher active rate:

• Bursty inference APIs: If your API sees traffic spikes with more than a third of the time idle between bursts, the scale-to-zero billing beats paying for an always-on pod that mostly sits idle.
• Prototyping and low-volume endpoints: Keeping an H100 inference endpoint warm at $2.89/hr for occasional queries is wasteful. Serverless lets you pay only for the seconds you actually use.
• Event-driven pipelines: Batch jobs that run overnight or on a schedule benefit from not paying for idle daytime hours.

Serverless loses for:

• Sustained training runs: A 72-hour fine-tuning job runs continuously with no idle time. At 100% utilization, Serverless ($4.55/hr) costs far more than Secure Cloud ($2.99/hr), and training frameworks often need persistent state that conflicts with stateless pod restarts. On-demand or spot is the right call here.
• Low-latency production inference: If first-token latency must be under 500ms, a 20-60s cold start is a non-starter. On-demand with a warm pod is the only option.
• High-throughput continuous inference: Above ~66% sustained utilization, an on-demand pod at $2.99/hr is cheaper than Serverless at $4.55/hr active, and more predictable. Reserve Serverless for the genuinely bursty case.

Spheron H100 Spot Availability and Marketplace Pricing

Spheron's marketplace model aggregates H100 capacity from data center partners globally, which creates more spot inventory than a single-provider platform can typically offer. Spot pricing for H100 SXM5 sits at $1.66/hr, roughly 57% below the $3.90/hr on-demand rate for the same SKU. That gap can narrow when supply tightens.

Spot instances suit workloads that checkpoint regularly. Training jobs using tools like Determined AI or custom PyTorch checkpoint loops can tolerate interruption every few hours without losing significant progress. For these workloads, running spot and re-queuing on interruption is cheaper than paying on-demand rates continuously.

Spheron's marketplace also differs structurally from Runpod's Community Cloud: Spheron vets providers for Tier 2/3/4 data center compliance before listing their hardware. This is a different reliability tradeoff than Runpod Community Cloud, where hosts range from enterprise to individual operators.

Check Spheron spot GPU instances for current availability and pricing. For a detailed comparison including architecture differences, see Spheron vs Runpod.

Runpod H100 SXM5 Secure Cloud runs $2.99/hr. Spheron H100 PCIe starts at $2.09/hr on-demand, with SXM5 spot from $1.66/hr for fault-tolerant workloads. No storage markups, no cold-start billing, per-minute granularity.

H100 pricing on Spheron → | View all GPU pricing →