As "AI bubble" becomes a topic of widespread discussion, Montage Technology released a Q3 earnings preview showing net profit growth exceeding 150% year-over-year. Setting aside stock price fluctuations and valuation levels, the true value of this report lies in its provision of technical testimony regarding the authenticity of AI demand from a critical niche within the supply chain.
The market's fear of an AI bubble essentially stems from skepticism about whether "computing power accumulation can translate into actual productivity." Montage Technology's performance growth reveals a silent shift occurring in the AI industry—moving from the training era to the inference and agent era—a transition that imposes fundamentally different requirements on hardware systems.
The causal chain outlined in the announcement is worth examining closely: the rapid development of AI inference and Agent applications has driven sustained growth in server CPU demand, which in turn increased the usage of memory modules and memory interconnect chips. This is not a simple linear narrative of "AI is hot, so chips sell well." In the traditional AI training phase, GPUs were the absolute protagonists, with a CPU-to-GPU ratio of approximately 1:8, where CPUs played more of a supporting role. However, Agentic AI is entirely different: agents require frequent task scheduling, tool invocation, logic orchestration, and state management—tasks that are precisely the domain of CPUs. AMD has publicly stated that the CPU-to-GPU ratio in AI infrastructure is converging from 1:4 or 1:8 toward 1:1. As CPUs return to the center stage from their supporting roles, the associated memory modules and interface chips have also seen increased demand and value.
Montage Technology is positioned exactly at this intersection. As the global leader in DDR5 memory interface chips, its products serve as the "throat" component of server memory modules—without these chips, efficient communication between CPUs and memory cannot be achieved. In the third quarter, the shipment proportion of the company's third, fourth, and fifth sub-generation RCD chips for DDR5 further increased, and the gross margin for interconnect chips reached 68.0%~68.5%, rising both year-over-year and quarter-over-quarter. In hardware manufacturing, a gross margin approaching 70% signifies extremely high technical barriers and pricing power; this figure cannot be achieved through low-price competition but is a direct reflection of product generational leadership.
This leads to a key judgment: if an AI bubble truly exists, it should manifest as inflated demand, homogenized products, and squeezed profits. What Montage Technology presents is precisely the opposite—demand driven by genuine architectural evolution, products winning through technological gaps, and profits expanding amid structural upgrades. The company even forecasted that demand for related products would further increase in the fourth quarter, indicating that management's confidence in the business cycle is well-founded.
Of course, a single quarterly report cannot endorse the entire AI industry. There are indeed segments in the market that rely solely on narratives without performance support, and valuations for some concept stocks do appear inflated. However, the case of Montage Technology draws a clear line: bubbles exist in segments far removed from real bottlenecks, while at hardcore strategic positions within the supply chain, demand remains robust and profits remain substantial. Memory interface chips, though seemingly a niche track, have become a touchstone for testing the quality of demand due to their irreplaceable technical positioning in the AI inference era.
When the noise subsides, technologies that genuinely resolve computing bottlenecks will speak for themselves through their performance. Montage Technology's Q3 report is not a wholesale rebuttal of the bubble theory, but rather a technical footnote from deep within the supply chain—it reminds us that the industrial foundation of AI is far more solid than what stock price curves suggest.