CICC: The path of cost reduction and diversification for AI hardware has opened up, which is expected to stimulate more demand.

According to a research report from China International Capital Corporation, in the current context of persistent inflation in memory costs and the negative cash flow beginning to appear in overseas CSPs, the market's focus has shifted back to the core issue of "AI hardware investment returns." In July, stocks in the AI hardware sector experienced significant adjustments. Unlike the previous round of declines, key indicators such as capital expenditures from major manufacturers and the annualized revenue of AI Labs are already at high levels, and market trading congestion is also higher than before. Through research conclusions on core issues like single-token costs and AI Lab API profits, China International Capital Corporation believes that the demand for AI hardware continues to grow strongly. On the hardware cost side, China International Capital Corporation believes that the inflation in memory costs cannot resist the efficiency gains from GPU technology iterations, leading to a rapid decrease in single-token costs. On the model side, closed-source models rely on premiums while open-source models focus on cost reductions, ensuring a solid profit foundation for all vendors, with upstream demand for computing hardware continuing to be stimulated. Furthermore, the refinement of the inference load structure is expected to facilitate the implementation of customized architecture chips and innovations in system interconnectivity. As inference applications shift from traditional Chat-only formats to agents and other forms, AI hardware may accelerate single-token cost reductions through customized architectures and optimized system interconnections, or overcome latency and throughput bottlenecks, providing performance premiums for real-time interactions and agent scenarios. The diversification of hardware cost reduction is expected to stimulate further demand.