Cerebras

Cerebras builds wafer-scale AI chips — the WSE-3 packs 4 trillion transistors and 900,000 AI cores with 44GB of on-chip SRAM on a single chip, sidestepping the memory-bandwidth wall that limits GPU inference. Cerebras Inference exposes this through an OpenAI-compatible cloud API (api.cerebras.ai) serving open models like Llama, Qwen, and GPT OSS at speeds independent benchmarks have measured above 1,800 tokens/second on Llama 3.1 8B — roughly an order of magnitude faster than hyperscale GPU clouds. Pricing starts around $0.10 per million tokens for 8B models and $0.60 for 70B, with a free tier (no credit card) for evaluation. Customers include Meta, Mayo Clinic, GSK, and Notion, and OpenAI has reportedly been testing Cerebras hardware as a non-Nvidia inference option.

The honest picture: raw speed is real, but the model menu is narrower than GPU clouds — niche fine-tunes or custom LoRA adapters need compilation and verification for the wafer-scale architecture, so availability lags on new releases. Throughput claims are workload-dependent (large batch sizes favor GPUs). Still, for latency-sensitive agentic workloads that hammer an LLM repeatedly, it is one of the most price-performant options available.

Create an account at cloud.cerebras.ai for a free trial with credits. Generate an API key and point any OpenAI-compatible client at https://api.cerebras.ai/v1 — e.g. OpenAI(base_url=..., api_key=...) in Python.