Cerebras Deploys Trillion-Parameter Models on Custom Chips

Cerebras, a specialist in custom AI hardware, has deployed trillion-parameter frontier models across its WSE-3T wafers, a significant step in scaling large AI systems. The deployment achieves inference speeds of 1,000–2,000+ tokens per second, addressing a critical bottleneck in making massive models practical for real-world applications.

The milestone underscores how specialized hardware can unlock performance gains for cutting-edge AI models. Trillion-parameter systems represent the frontier of model scale, and demonstrating fast, stable inference at that scale is a key technical and commercial challenge.