Groq is an AI inference cloud. Inference means running an already trained AI model to produce answers, as opposed to training it. Groq runs models on the LPU, a processor for inference that Groq pioneered, in 13 data centers on four continents, and it is building LPX racks that connect LPU accelerators to NVIDIA's next-generation GPUs.
Developers use GroqCloud through an API that is compatible with OpenAI's: a program sends text or audio to api.groq.com and receives the model's output. The models come from other makers, such as OpenAI's open-weight GPT-OSS, Llama and Qwen, plus Whisper for speech-to-text and Orpheus for text-to-speech. The documentation covers reasoning, structured outputs, tool use, prompt caching and batch processing. Companies can also rent dedicated bare-metal capacity (GroqMetal), with an inference stack and enterprise controls on top.
GroqCloud has a free tier with lower limits and is otherwise billed by usage: per million tokens for language models, per hour of audio for speech-to-text and per million characters for text-to-speech.