▶Feldman consistently argues across multiple appearances that NVIDIA's GPU architecture is fundamentally flawed for AI, especially for inference, due to memory bandwidth bottlenecks that lead to massive underutilization.
▶He repeatedly identifies the primary constraint for large-scale AI deployment as physical power availability and grid infrastructure, rather than the supply of silicon chips.Feb 2026
▶Feldman maintains that NVIDIA's competitive moat is its dominant market share and default status, not a technical 'lock-in' from its CUDA software, which he claims is irrelevant for inference workloads.Mar 2026
▶He asserts that Cerebras's wafer-scale architecture provides a fundamental advantage by co-locating massive amounts of fast SRAM with compute, directly addressing the memory bottlenecks of GPUs.Feb 2026
▶Feldman directly debates the industry's perception of 'CUDA lock-in,' arguing it is a myth for inference tasks where simple APIs are sufficient, though he concedes software porting for training remains a challenge.
▶He challenges the conventional wisdom of a rapid, two-year depreciation cycle for AI hardware, pointing to the continued utility of older NVIDIA chips like the A100 and H100 for less demanding tasks.
▶Feldman refutes near-term predictions of massive AI-driven labor shortages, arguing that while productivity gains will be enormous, the organizational restructuring required means significant economic dislocation is more than five years away, possibly closer to fifteen.
▶He contests the idea that vertical integration (building both models and chips) is necessary for success in AI, citing the success of non-integrated players like OpenAI (using Azure) and Anthropic (using AWS/Google) as evidence.
Not enough data for timeline
Sign up free to see the full intelligence report
Get started free