The demand for AI compute is growing at an explosive and unpredictable rate, with customers themselves unsure of their future needs by an order of magnitude. This forces infrastructure providers like Cerebras to make massive, long-term capital commitments for manufacturing and data centers in a highly uncertain environment.
NVIDIA is positioned as the dominant force, employing aggressive tactics like 'predatory pre-announcements' to maintain its lead. However, the speaker identifies weaknesses, such as high field failure rates and performance bottlenecks like memory bandwidth, which create opportunities for differentiated hardware solutions.
The primary limitations to AI's growth are not just chip supply, but a cascade of systemic issues. These include a critical shortage of AI expertise, insufficient university output, restrictive immigration policies, and the multi-year timelines required to build new semiconductor fabs and data centers.
While AI training gets significant attention, the market for AI inference is described as vastly larger and more accessible for competitors. For inference workloads, customers prioritize simple APIs over NVIDIA's complex CUDA ecosystem, lowering the barrier to entry for alternative hardware providers.
The market's heavy reliance on a few large tech companies (Magnificent Seven) for growth creates significant, underappreciated sector risk. Investors who view the S&P 500 as a diversified index are unknowingly making a concentrated bet on the AI sector, which could derail the broader market if AI development hits a snag.
Keep pulling the thread on Andrew Feldman.