Feldman defines the AI chip landscape, contrasting general-purpose GPUs with specialized ASICs like Cerebras' wafer-scale chip, which is 58 times larger than a GPU. He explains that AI inference is memory-bound, and Cerebras uses on-chip SRAM to move model weights 2,500 times faster than HBM-based GPUs, solving the '100 HD movies' problem. This speed is crucial for reasoning models, agents, and reinforcement learning, where latency is amplified. Feldman argues that Nvidia's CUDA moat is shrinking for inference, as switching is easy. He discusses Cerebras' partnership with AWS, where Trainium handles prefill and Cerebras handles decode, and its 750MW deal with OpenAI. He concludes that fast AI will reshape SaaS by enabling instant tool creation and breaking down organizational silos, and that today's models will be the worst we ever use.