Keep pulling the thread on Amin Vahdat & Jeetu Patel.
The current AI infrastructure build-out is 100 times larger in scale than the build-out of the internet during the late 1990s and early 2000s.
Current market projections are grossly underestimating the future demand for AI infrastructure.
Google's seven and eight-year-old TPU models are still operating at 100% utilization due to high demand.
Cisco aims to achieve a 2x to 3x productivity increase for its 25,000 engineers within the next year by leveraging AI tools.
The supply of AI infrastructure is expected to lag demand for the next three to five years due to constraints in power, land, permitting, and the supply chain.
Due to power scarcity, new data centers are being built in locations where power is already available, rather than bringing power to desired data center locations.
The entire computing stack, from hardware to software, is predicted to be unrecognizable within the next five years.
For certain computations, Google's TPUs are 10 to 100 times more power-efficient per watt than a CPU.
China's semiconductor industry is currently capable of producing 7-nanometer chips, but not 2-nanometer chips.
China's strategy to compete in AI hardware involves leveraging its abundant power and engineering resources to optimize older 7-nanometer chip technology, compensating for its lack of access to leading-edge process nodes.
The power consumption of large-scale AI clusters fluctuates so dramatically between computation and networking phases that it is noticeable to power utility companies at the scale of tens to hundreds of megawatts.
While the cost of AI inference is being reduced by factors of 10x to 100x, end-user costs are not decreasing because market demand immediately shifts to newer, higher-quality models that consume these efficiency gains.