DeepInfra

DeepInfra is a pay-as-you-go inference cloud with APIs for 100+ open models across text, speech, image, video and embeddings, plus on-demand GPU rental. It stresses low cost, zero data retention, and SOC 2 and ISO 27001 compliance.

DeepInfra is an inference cloud that serves machine-learning models through developer-facing APIs. Instead of provisioning GPUs and tuning serving stacks, a team picks a hosted model, calls it over an API and pays for the tokens or compute it uses. The catalog covers more than 100 models across text generation, embeddings, rerankers, speech recognition, text to speech, text to image, text to video, text to music, world models and zero-shot image classification. The text-generation shelf is the most visible part of the homepage. It lists large open-weight families from DeepSeek, Moonshot AI (Kimi), Qwen, Z.ai (GLM), Xiaomi (MiMo), Tencent and NVIDIA, each with per-million-token input and output rates. Some listings add a cached-input rate, a context window of up to roughly one million tokens, a precision tag such as fp8 or fp4, and a zero data retention marker. Temporary discounts are flagged on certain models. Beyond hosted models, the site offers on-demand GPU rental (a DGX B300 instance is listed by the instance-hour), a DeepCluster offering, a DeepStart program and a browser chat page for trying models. The audience is developers, startups and enterprises that want to run open models in production without owning hardware. The pitch centers on low cost, speed, simplicity and reliability, with scale from small experiments to very high token volumes and hands-on technical support. The company says it runs on its own inference-optimised hardware in US-based data centers, holds SOC 2 and ISO 27001 certifications, and applies a zero retention policy to inputs, outputs and user data. It also announced a $107M Series B in May 2026. In practice, DeepInfra sits among open-model inference providers and GPU clouds. It suits teams comparing per-token prices across many model families, or those who want to swap models without changing vendors. Buyers who need a single proprietary frontier model, a visual no-code builder or a fixed monthly plan will find it less direct, since the product is built around APIs and usage-based billing.