Cerebras swaps AI architecture for wafer-scale chips

Cerebras replaces conventional AI processing architecture with wafer-scale chips. The technology is not yet available in Brazil.

Cerebras has developed a method for running trillion-parameter AI models on wafer-scale chips. Each of the company's wafer-scale chips has 44 GB of SRAM, which allows large volumes of data to be processed efficiently. The Kimi K3 model, with 3 trillion parameters, is used as an example of the technology. The model is split into 93 layers, with 16 GB per layer, which makes it easier to distribute the workload. Communication between wafers takes place via RDMA, with a latency of 2 microseconds per hop, ensuring speed and accuracy. Cerebras offers a free API and support for model deployment, at no cost.

Cerebras leads innovation in trillion-parameter AI

Cerebras is leading innovation in wafer-scale chips for trillion-parameter AI, while C12 and Thales are developing a quantum technology to improve drone tracking. Synology is also integrating AI into cameras to improve physical security. C12 and Thales's technology, called QuantumTrack, won the Quantum Effects Award 2026.

Synology, for its part, is expanding the use of AI in security systems, allowing cameras to analyze images and respond to queries as if they were human. The company has already demonstrated how the method works with the Kimi K3 model, split into 93 layers, each with 16 GB of memory. This approach allows trillion-parameter models to run efficiently, something competitors have not yet been able to replicate at the same scale. Cerebras replaces the traditional AI processing architecture with wafer-scale chips, allowing trillion-parameter models to run with reduced latency and integrated memory. Synology created an AI technology that lets users "talk" to cameras to search for videos by description.