TL;DR If you like Lambda’s managed feel but want lower on-demand prices or simpler dev UX, start with Thunder Compute (A100 80 GB at $0.78 per hour, H100 at $1.47 per hour) and compare against Runpod, Crusoe, Voltage Park, Modal, Paperspace, and marketplace options like Vast.ai.
Quick picks
<ul> <li><strong>Cheapest on-demand A100/H100</strong>: <strong>Thunder Compute</strong> (<a href="https://www.thundercompute.com/pricing?utm_source=chatgpt.com">pricing page</a>) – per-second billing, persistent storage, one-click VS Code integration.</li> <li><strong>Large marketplace with consumer cards</strong>: <strong>Vast.ai</strong> (<a href="https://vast.ai/article/high-performance-deep-learning-with-cloud-gpus?utm_source=chatgpt.com">overview of GPU marketplace</a>) – crowdsourced supply, many RTX 4090 options.</li> <li><strong>Enterprise-y alternative with public rates</strong>: <strong>Crusoe</strong> (<a href="https://support.crusoecloud.com/hc/en-us/articles/37421109850907-FAQ-Determining-On-Demand-Pricing-for-Crusoe-Offerings?utm_source=chatgpt.com">on-demand pricing FAQ</a>).</li> <li><strong>Low H100 headline price at scale</strong>: <strong>Voltage Park</strong> (<a href="https://www.voltagepark.com/pricing?utm_source=chatgpt.com">pricing</a>) – H100 from $1.99 per hour.</li> <li><strong>Serverless and workflows</strong>: <strong>Modal</strong> (<a href="https://modal.com/pricing?utm_source=chatgpt.com">pricing</a>) – per-second rates translate to ~$2.50/hr for A100, ~$3.95/hr for H100.</li> <li><strong>Broad ecosystem and notebooks</strong>: <strong>Paperspace</strong> via DigitalOcean (<a href="https://docs.digitalocean.com/products/paperspace/machines/details/pricing/?utm_source=chatgpt.com">pricing details</a>) – H100 on-demand ~$5.95/hr; A100 also available.</li> </ul>
Pricing snapshot (A100 and H100)
Rates are on-demand list prices where published. Some providers sell multi-GPU nodes; figures shown are per-GPU where the provider publishes per-GPU pricing.
Marketplace vs managed clouds (important if you need consumer GPUs)
Marketplaces can deliver the lowest cost, but host consistency varies.
<ul> <li><strong>Vast.ai</strong> is a decentralized, peer-to-peer marketplace aggregating GPUs from both individuals and datacenters—including consumer-grade GPUs like RTX 4090—resulting in high supply variability and often lower prices.</li> <li><strong>Runpod Community Cloud</strong> also lists consumer GPUs with transparent starting prices and community-provided capacity.</li> </ul>
If consistent performance, multi-GPU NVLink, or enterprise networking matters, managed clouds (Thunder Compute, Lambda Labs, Crusoe, Voltage Park) are more predictable.
Why teams pick Thunder Compute
<ul> <li><strong>Lowest on-demand A100/H100 rates in this comparison</strong>—A100 80 GB for $0.78/hr; H100 for $1.47/hr.</li> <li><strong>Developer velocity</strong>—one-click VS Code, per-second billing, persistent disks, snapshots, dynamic vCPU/RAM adjustments.</li> <li><strong>Simple pricing model</strong>—storage at $0.15/GB/month.See the <a href="https://www.thundercompute.com/pricing?utm_source=chatgpt.com">Thunder Compute pricing page</a> for up-to-date details.</li> </ul>
How to choose
<ul> <li><strong>For multi-GPU training with fast interconnect</strong>: opt for managed providers that explicitly publish SXM node specs and interconnect performance.</li> <li><strong>For fast prototyping or fine-tuning</strong>: prioritize per-second billing, quick restart speeds, and persistent storage.</li> <li><strong>To minimize cash burn</strong>: compare hourly A100 vs H100 costs—A100 often offers more cost-effective compute per token for prototyping models.</li> <li><strong>If you need consumer GPUs</strong>: good for specific workloads like image generation or lightweight training, but verify VRAM, driver compatibility, and host stability.</li> </ul>