According to the CICC research report, at present, against the backdrop of continued inflation in memory costs and overseas CSP free cash flows beginning to turn negative, the market focus is once again shifting back to the core issue of “AI hardware return on investment,” and AI hardware circuit stocks were drastically adjusted in July. Unlike the previous round of decline, core indicators such as capital expenses of major manufacturers and AI Lab's annualized revenue are already at a high level, and market transaction congestion is higher than in the previous period. Through research findings on core issues such as single token costs and AI Lab API profits, CICC believes that demand for AI hardware is still growing strongly. On the hardware cost side, CICC believes that storage inflation will not be able to withstand the iterative efficiency gains of GPU technology, and the cost of a single token continues to fall rapidly. On the model side, the closed source model relies on premium and open source models to reduce costs, the profit base of various vendors is stable, and the demand for upstream computing power hardware continues to be driven. Furthermore, refinement of inference load structures is expected to drive the implementation of customized architecture chips and system interconnection innovation. As inference applications shift from traditional chat-only to agents, etc., AI hardware may accelerate single-token cost reduction or break through latency and throughput bottlenecks through customized architectures, providing a performance premium for real-time interaction and agent scenarios. Once the path of diversification of hardware cost reduction begins, it is expected to stimulate more demand.

According to the CICC research report, at present, against the backdrop of continued inflation in memory costs and overseas CSP free cash flows beginning to turn negative, the market focus is once again shifting back to the core issue of “AI hardware return on investment,” and AI hardware circuit stocks were drastically adjusted in July. Unlike the previous round of decline, core indicators such as capital expenses of major manufacturers and AI Lab's annualized revenue are already at a high level, and market transaction congestion is higher than in the previous period. Through research findings on core issues such as single token costs and AI Lab API profits, CICC believes that demand for AI hardware is still growing strongly. On the hardware cost side, CICC believes that storage inflation will not be able to withstand the iterative efficiency gains of GPU technology, and the cost of a single token continues to fall rapidly. On the model side, the closed source model relies on premium and open source models to reduce costs, the profit base of various vendors is stable, and the demand for upstream computing power hardware continues to be driven. Furthermore, refinement of inference load structures is expected to drive the implementation of customized architecture chips and system interconnection innovation. As inference applications shift from traditional chat-only to agents, etc., AI hardware may accelerate single-token cost reduction or break through latency and throughput bottlenecks through customized architectures, providing a performance premium for real-time interaction and agent scenarios. Once the path of diversification of hardware cost reduction begins, it is expected to stimulate more demand.