$CBRS It surged last night, reaching as high as $390. Traditional GPU clusters, like those used by NVIDIA, require hundreds or thousands of GPUs connected and communicating with each other to train or run a very large AI model. Data needs to be moved back and forth between these tiny chips, which is slow, power-intensive, and complex, especially during inference. Frequent data waiting can make AI noticeably slower. Cerebras devised a clever solution: they created a single giant chip from an entire wafer, approximately 50 times larger than an NVIDIA H100 GPU. All cores on the wafer are directly connected using ultra-short nanowires, achieving a bandwidth of over 200 Pb/s and extremely low latency. GPU clusters are like workers in many small rooms using delivery trucks to transport materials. Cerebras, on the other hand, crams all workers, materials, and tools into a vast, open factory hall, where everyone works face-to-face, handing things over, with materials barely moving. The result is significantly faster speeds; a single chip can house a giant model, saving power and costs while maintaining extremely low latency.