GPU Cloud Pricing Comparison 2026

It depends on the card, the provider tier and whether you take spot or on-demand capacity. In AltStreet's July 2026 verification pass: H100 ran from about $2.46/hour (specialized-provider spot) to $6.16/hour (on-demand 8-GPU nodes), with hyperscale on-demand at roughly $10.35 to $12.29/hour per GPU; A100 80GB from about $1.21/hour to $2.79/hour at specialized providers and $0.68 (spot) to $4.10/hour at hyperscalers; L40S from about $0.39/hour on-demand; and RTX 4090 from about $0.14/hour spot on decentralized marketplaces. The live dashboard shows the current figures, refreshed from provider pages, and the calculator turns an hourly rate into a cost per million tokens for your workload.

Yes, two. The calculator on this page turns a provider's hourly rate into a cost per million tokens, per training epoch or per month for your workload, including egress, storage and compute time. For a fuller total cost of ownership, AltStreet's AI Infrastructure Cost Calculator at /tools/ai-infrastructure-calculator models training, inference, GPU utilization, data labeling and operational overhead together. Both take the rates you enter or the current rates on the live dashboard; neither is a quote from a provider.

No single provider is cheapest for every card. In the July 2026 pass, decentralized marketplaces such as Vast.ai had the lowest RTX 4090 spot rate (about $0.14/hour), RunPod had the lowest L40S on-demand rate (about $0.39/hour) and a $0.34/hour 4090 spot rate, CoreWeave had the lowest specialized H100 spot rate (about $2.46/hour) and A100 80GB spot rate (about $1.21/hour), Lambda Labs priced H100 at about $2.49 to $3.99/hour and A100 80GB on-demand at about $2.79/hour, and the hyperscalers (AWS, GCP, Azure) were the most expensive per hour for H100 on-demand. Lower price usually buys lower availability guarantees and fewer enterprise features, which is why the dashboard pairs every rate with an availability score.

In the July 2026 verification pass, A100 80GB pricing at specialized providers ranged from about $1.21/hour (CoreWeave spot) to $2.79/hour (Lambda Labs on-demand). At hyperscalers it ranged from about $0.68/hour on Azure spot to $4.10/hour on Azure's on-demand ND A100 v4 instances. Decentralized marketplace rates for the A100 80GB, including Vast.ai, move too often to print here and are shown on the live dashboard with the time they were observed.

This page and its live dashboard track nine providers, seven of them verified daily by scraping their published rates: Lambda Labs, RunPod, Vast.ai, CoreWeave, AWS, GCP, Azure and others. Each offer is listed with its hourly price, pricing type (spot, on-demand or committed), region, GPU count per instance and an availability score, and the yield-score ranking orders cards by performance per dollar. The ranges on this page are a dated snapshot; the dashboard is the current view.

As of our July 2026 verification pass, RTX 4090 spot instances start around $0.14/hour on decentralized marketplaces (batch-priority distributed capacity around $0.20/hour), and L40S starts near $0.39/hour on-demand at specialized providers. For larger models requiring 80GB VRAM, A100 80GB spot pricing was verified as low as $0.68/hour (hyperscale spot) with on-demand rates of roughly $1.21-4.10/hour depending on provider type. Actual rates vary by provider, region, availability, and interruption tolerance.

As of our July 2026 verification pass, H100 hourly rates at specialized providers ranged from about $2.46/hour (spot) to $6.16/hour (on-demand 8-GPU HGX nodes), with hyperscale on-demand rates of roughly $10.35-12.29/hour per GPU. Decentralized marketplace rates change continuously; see the live dashboard for current listings. Price differences reflect availability guarantees, support levels, and SLA commitments.

In AltStreet's July 2026 verification pass, RTX 4090 spot capacity ranged from about $0.14/hour on Vast.ai to about $0.34/hour on RunPod spot, with SaladCloud batch-priority capacity around $0.20/hour. On-demand RTX 4090 rates run higher and move daily; the live dashboard on this page shows the current quotes.

In the July 2026 pass, specialized-provider H100 pricing ran from about $2.46/hour (CoreWeave spot) to $6.16/hour (CoreWeave on-demand 8-GPU nodes), with Lambda Labs on-demand H100 at roughly $2.49 to $3.99/hour inside that range. RunPod H100 and decentralized-marketplace H100 quotes were not verified in that pass and are reported on the live dashboard rather than as a static range.

RunPod on-demand L40S was the lowest verified L40S rate in the July 2026 pass at about $0.39/hour (spot from about $0.44/hour at other providers), with CoreWeave on-demand at the top of the specialized range at $2.25/hour. Hyperscaler L40S availability was limited.

GPU yield score is a performance-per-dollar metric calculated as Performance Index / Hourly Price. Higher scores indicate more computational output per dollar. This metric helps identify cost-efficient options but should be considered alongside workload requirements, availability needs, and performance constraints.

Decentralized GPU marketplaces often offer meaningfully lower hourly rates (sometimes 50%+ savings) than hyperscale clouds by aggregating supply from diverse sources. Trade-offs commonly include lower availability guarantees (often cited as 70-85% vs 99.9%), no enterprise SLAs, and variable network performance. Decentralized providers work well for development, batch workloads, and cost-sensitive applications, while production services requiring high reliability generally favor hyperscale or specialized providers.

Cost per 1M tokens depends on GPU throughput, hourly price, model architecture, and optimization. As an example, RTX 4090 instances running smaller models may achieve costs in the $0.50-1.00 per 1M tokens range, while H100 instances can be more efficient for larger models despite higher hourly rates due to superior throughput. Use our calculator with your specific model parameters for accurate estimates.

For workloads with consistent 24/7 utilization at 80%+ capacity, purchasing hardware may break even in approximately 8-12 months compared to cloud rental. However, ownership requires significant upfront capital, ongoing power and cooling costs, maintenance, and lacks flexibility for scaling or GPU type changes. Cloud rental generally proves more cost-effective for variable workloads, short-term projects, or when testing different GPU configurations.