Software should work, and talking about it needn't be boring (www.theregister.com)

Meta has unveiled its MTIA 400 chip, designed for dual purposes: training AI models and serving advertisements. This chip boasts enhanced performance, outpacing Nvidia's Blackwell architecture in certain tasks, although it hasn't yet established itself as a direct competitor to AMD or Nvidia. The MTIA 400’s ability to efficiently handle both training and inference workloads represents a significant step forward, particularly for developers engaged in complex AI applications that demand quick processing and adaptability. In a parallel development, OpenAI is preparing to launch its Jalapeño chip, which promises to deliver unparalleled inference capabilities with a staggering 128 chips, achieving up to 1.7 exaFLOPS and 27 TB of high-bandwidth memory (HBM). This positioning puts OpenAI in a strong competitive stance against industry leaders, indicating a potential shift in the hardware landscape that could empower more sophisticated AI solutions. Together, these advancements reflect a growing trend where companies are increasingly optimizing hardware for specific AI tasks, which could ultimately accelerate innovation in machine learning capabilities across various sectors.