Salad: Utilizes distributed consumer hardware to drastically lower inference costs, but cannot match enterprise GPU availability or SLA guarantees. RunPod 每月有 2,881,043 次造訪，月增率 +5.9%，定價模式為付費。Salad 每月有 713,519 次造訪，成長率 +10.4%，定價模式為免費試用、付費、需聯絡報價。
面向 AI 開發者的 GPU Pod、Serverless 推論與叢集雲平台
按需、可中斷或預留方式租用 GPU 執行個體，涵蓋 30 多種 GPU 型號，從 RTX 4090 到 B200、B300 等級加速卡，部署於全球 31 個區域。
把任意容器部署為 API 端點，按需從 0 擴展到數千個 worker，僅為實際使用的算力付費。
啟動多節點 GPU 叢集用於分散式訓練與大規模推論，可掛載共享儲存，且無需長期承諾。
從精選中樞部署熱門開源模型，或使用公開端點呼叫圖像、音訊與文字模型，無需自建基礎設施。
編排全球閒置消費級 GPU 的分散式 GPU 雲
在數千個分散式 GPU 節點上調度、監控並重新分配 Docker/OCI 工作負載，節點消失時自動容錯移轉。
隨需存取覆蓋 190 多個國家的閒置消費級 GPU 分散式算力池，優先定價約從每小時 0.015 美元起。
面向開源權重模型的託管 token 端點，按 token 計費，你的資料不會被用於訓練。
在組織內部運行同一套編排堆疊，把閒置工作站 GPU 匯聚成內部相容 OpenAI 的服務。
專項更強Provides a unified API gateway across hundreds of models, but entirely abstracts away underlying GPU infrastructure management.
專項更強Provides a unified API gateway across hundreds of models, but entirely abstracts away underlying GPU infrastructure management.
專項更強Specializes in lightning-fast serverless execution for generative media models, but lacks general-purpose GPU pod rental options.
專項更強為 AI 管線提供輕鬆的 Python 無伺服器執行，但專注於 Function-as-a-Service 而非裸機虛擬機。
Offers a polished API and cloud platform for running open models, but provides less direct infrastructure-level control than RunPod.
Offers a polished API and cloud platform for running open models, but provides less direct infrastructure-level control than RunPod.
專項更強Offers extremely cheap serverless model hosting via API, but does not provide raw containerized GPU instances.
Provides managed development studios and guaranteed on-demand GPUs, but with a more structured framework-centric workflow than RunPod.
Provides managed development studios and guaranteed on-demand GPUs, but with a more structured framework-centric workflow than RunPod.
Combines serverless model APIs with on-demand GPU instances, offering a similar developer experience at competitive global price points.
Focuses on large-scale enterprise GPU clusters for training, rather than lightweight consumer container workloads.
Focuses on large-scale enterprise GPU clusters for training, rather than lightweight consumer container workloads.
Features second-billed persistent SSH GPU instances, but lacks a distributed consumer hardware backbone.
Features second-billed persistent SSH GPU instances, but lacks a distributed consumer hardware backbone.
