Most chip designers, including NVIDIA, lay out circuits on a large wafer, then dice it into smaller dies. Those dies are interconnected to build AI accelerators, improving yield by avoiding scrapping an entire wafer due to defects.

One company refuses to follow that path. It treats the whole wafer as a single chip—a 21.5 cm x 21.5 cm square roughly matching the largest rectangle you can cut from a 12-inch wafer—packing 4 trillion transistors and 900k cores; the company is Cerebras. The bottleneck for GPU inference is the 'memory wall'...