The discussion centers on Cerebras and its wafer-scale chip, which is 58 times larger than a typical chip and offers up to 1,000 times faster AI inference than GPUs by using a different memory architecture. CEO Andrew Feldman explains that the chip was built from scratch for inference, avoiding key industry bottlenecks like HBM memory and TSMC's 3nm process. He argues that open-source AI models are significantly cheaper per unit of intelligence than closed-source ones, though slightly less capable. Feldman also notes that Nvidia's CUDA software moat is shrinking, as leading models no longer rely on it for inference. The conversation covers the company's relationship with major customer G42, the strategic importance of semiconductors, and the challenges of building US fabs. Finally, Feldman reflects on the company's IPO, creating over 800 millionaires, and the balance between long-term innovation and quarterly pressures.