Novita AI

What we found

Novita AI is a cloud platform that sells access to open AI models and the GPUs to run them, now pitched as an AI-native cloud for builders and agents. Its serverless API offers more than 200 models, by its own count, for text, images, video, audio and embeddings through an OpenAI-compatible interface, with batch inference at an introductory half price. Dedicated endpoints run a chosen model on reserved GPUs with per-second billing that scales to zero. Developers can also rent GPU instances and bare-metal servers, and an Agent Sandbox offers isolated runtimes for AI agents. Novita has partnered with the vLLM and SGLang inference projects, displays a SOC 2 badge and claims 99.5 per cent uptime; its homepage lists Hugging Face, Quora, OpenRouter, Fish Audio and Kilo Code among users.

Little is published about the company itself. Its terms name only Novita AI under Delaware law, its press releases are datelined San Francisco, and Junyu Huang is a co-founder and chief operating officer; we could not confirm its founding year or any funding.

Pricing is usage-based with no subscription. Language models range from $0.02 per million input tokens for Llama 3.1 8B through $0.14 for DeepSeek V4 Flash to $3 in and $15 out for Kimi K3, a few models are priced at zero, and Qwen-Image costs $0.02 an image. Dedicated endpoints cost $0.61 an hour per RTX 4090 GPU, $1.99 per H100 and $2.99 per H200. GPU instances start at $0.17 an hour for an RTX 4090 on spot pricing or $0.33 on demand, and an H100 SXM costs $3.39 an hour on demand or $1.70 on spot. Container disks include 60 GB free, then $0.005 per gigabyte a day.

We could not read independent review sites. Novita suits developers who want cheap access to open models and flexible GPUs; teams needing a contractual SLA should ask sales first.

Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.