Novita AI

Novita AI positions itself as an "AI-native cloud for builders and agents," selling low-cost inference for more than 200 open and open-weight models behind an OpenAI-compatible API, plus GPU instances, bare-metal GPUs, and a serverless GPU tier that scales to zero in about 30 seconds. Its posted rates run from roughly $0.02 per million input tokens for small Llama models to a couple of dollars per million for premium reasoning models, with a standing 50% batch discount and cache-read pricing on select routes. That model catalog and price point put it in the same lane as Together AI, Fireworks, and DeepInfra rather than the training-hyperscaler lane occupied by CoreWeave or Lambda.

Infrastructure-wise, Novita advertises a distributed footprint across "20+ locations" and "4+ continents" (14 global regions), delivered through a mix of GPU instance types (on-demand and spot at roughly 50% off) rather than owned campuses. The company does not publicly disclose GPU counts, generation mix, MW capacity, or datacenter partners; that is a deliberate contrast to the disclosure posture of listed peers like CoreWeave. Silicon partnerships center on the Nvidia stack, and inference is served through open toolchains including vLLM and SGLang.

Commercially, Novita is best known as a supplier under other people's front doors: Hugging Face lists it among Inference Providers, OpenRouter routes traffic to it, and Quora's Poe surfaces its models. The trade-off of that channel-heavy strategy is that Novita's own brand, funding history, headcount, and cap table are not well documented in the public record; the company has not announced a priced venture round to major outlets as of mid-2026.