Dylan Patel

Compute for AI is still a limited resource, driving sustained demand for semiconductor stocks. AI models are increasingly restrictive with tokens and downgrading to save compute, and the latest frontier models require massive compute. This backdrop supports continued upside for semiconductor stocks.

Alternative AI accelerators and startup chip companies may win some real orders, but their volumes are tiny compared with Nvidia and Google TPUs, and many startup orders are just nebulous LOIs. As long as demand outstrips production and incumbents keep improving, the bulk of AI accelerator revenue and cash flows will go to incumbents such as Nvidia and Broadcom.

Alternative AI accelerators and startup chip companies may win some real orders, but their volumes are tiny compared with Nvidia and Google TPUs, and many startup orders are just nebulous LOIs. As long as demand outstrips production and incumbents keep improving, the bulk of AI accelerator revenue and cash flows will go to incumbents such as Nvidia and Broadcom.

SMIC's latest DUV node, N+3, has achieved EUV-class density without EUV, demonstrating real progress. Although power and performance still lag the leading edge, this shows China is becoming a serious fab player. SMIC is being directed to license its N+2 and N+3 processes to other domestic fabs. If this learning spreads into AI accelerators and other chips, the choke point moves from a single sanctionable fab to an entire ecosystem, making China less dependent on TSMC and harder to contain. China doesn't need to beat TSMC; it only needs to be good enough to not need TSMC at all.

DRAM prices are going to double or triple from current levels because memory capacity can only grow at low double-digit percentages per year, and AI demand is so strong that incremental supply won't arrive until 2027 or 2028, forcing demand destruction via higher pricing.