OpenAI plans to integrate 750 megawatts of Cerebras’ chip technology within its operations until 2028, enhancing its AI services. This aligns with OpenAI’s strategy to outperform Nvidia by using Cerebras’ wafer-scale technology, valued over $10 billion. These chips boast a massive SRAM capacity, facilitating rapid inferences with an impressive memory bandwidth of 21 Petabytes per second, far surpassing Nvidia’s capabilities. This significant upgrade allows models like GPT-OSS 120B to deliver faster results, enabling longer reasoning without losing interactivity. However, Cerebras’ SRAM-centric design poses certain limitations, requiring multiple chips for larger models due to substantial power requirements. Anticipated future Cerebras designs may address these constraints, allowing for enhanced memory efficiency and the incorporation of advanced data types. Incorporating a model router in GPT-5 aids in using memory efficiently, by allocating smaller, cheaper models for simpler tasks, and reserving Cerebras’ technology for advanced queries. This partnership could signify a shift in data center architecture, contrasting with traditional setups focused on Nvidia and AMD as Cerebras technology rises in relevance.