The speakers characterize the current AI infrastructure build-out as a combination of the internet's development, the space race, and the Manhattan Project. They argue it is 100x the scale of the 90s internet boom and that current market forecasts are still underestimating the long-term demand, as evidenced by 100% utilization of even 7-8 year old Google TPUs.
The primary bottleneck for AI expansion is no longer just chip supply, but fundamental physical resources like power, land, and permitting. This scarcity is forcing a strategic shift where data centers are built where power is available, rather than bringing power to desired locations, leading to more distributed architectures.
The unique demands of AI workloads are driving a move away from general-purpose CPUs towards highly specialized silicon. This includes custom accelerators like Google's TPUs, which offer 10-100x power efficiency for certain tasks, and the emergence of hardware specifically designed for inference versus training, each with distinct computational profiles.
The infrastructure build-out is a key front in global competition, with different national strategies emerging. China, for example, may leverage abundant power and engineering talent to optimize older 7nm chips, while the US focuses on cutting-edge 2nm designs. In the corporate sphere, companies like Cisco are creating their own silicon to provide an alternative to a potential Broadcom monopoly in networking.
Large tech companies are aggressively deploying AI internally to accelerate their own development and operations. Google used AI tools to speed up a massive code migration (TensorFlow to JAX), while Cisco is aiming for a 2-3x productivity increase for its 25,000 engineers by using AI for tasks like code generation, debugging, and legal contract review.
Keep pulling the thread on Amin Vahdat & Jeetu Patel.