The Memory Divide

“America leads the frontier. China challenges the economics. The next contest is the cost of production.”

New evidence strengthens a critical qualification to our U.S.–China thesis: China may be narrowing the cost gap in AI models considerably faster than it is closing the gap in advanced computing infrastructure.

This is a material inflection for semiconductor supply and AI infrastructure. Two developments, considered together, are consequential.

First, global AI infrastructure demand is accelerating. South Korea reported on November 1 that October semiconductor exports surged 259.4% year over year to $56.44 billion, while computer exports increased 496.6%, driven partly by demand for AI servers. Those figures reflect calendar effects, including additional working days.

Second, China’s domestic advanced-memory industry remains substantially behind. Reuters reported on October 28 that CXMT intends to begin limited production of HBM3E, the high-bandwidth memory used in advanced AI processors, despite yields of approximately 20–30%, against around 70% at SK Hynix. CXMT’s reported performance also trails comparable South Korean products. These figures come from industry sources rather than audited results.

The AI Supercycle is accelerating demand for physical computing resources precisely when China’s access to the most advanced components remains constrained.

Our central proposition remains intact, but the landscape now requires a third dimension: control of the infrastructure necessary to produce intelligence economically. China’s developers have shown progress in model efficiency, affordable inference and enterprise distribution. Yet inexpensive inference does not imply inexpensive hardware.

High-bandwidth memory has become a critical constraint on the performance, economics and scalability of advanced accelerators. China’s manufacturers are attempting to address it, but low yields increase costs and complicate deployment, and Chinese AI-chip makers have already raised prices by approximately 20–50% amid rising memory costs.

America’s advantage lies in frontier capabilities and access to an advanced international computing ecosystem; China’s challenge is to combine efficient models with autonomous, commercially competitive infrastructure. Sustained growth in advanced reasoning, agentic systems and industrial AI will require reliable capacity at competitive costs.

The November 1 Korean figures complicate the argument that the recent correction in American AI equities signals an imminent collapse in demand. The physical trade data indicate continued extraordinary investment activity — which does not guarantee attractive returns.

Indeed, the Bank of Korea warned on October 27 that a slowdown in spending by major American technology companies could reduce South Korean semiconductor exports by $50–100 billion relative to earlier expectations in a pessimistic 2027 scenario. This is the financial tension at the center of the AI Supercycle: demand can remain exceptionally strong even as investors question whether the capital invested will earn acceptable returns.

Three implications follow. First, advanced memory deserves recognition as a strategic asset class, extending analysis beyond GPUs to HBM, advanced packaging and yields. Second, China’s progress should be judged on commercial production economics — reliable yields, performance, energy efficiency and cost per successful workload. Third, the cycle may become increasingly differentiated: scarce component suppliers can benefit from capacity constraints even while infrastructure operators face pressure on capex and cash generation.

Five Key Takeaways

For investors and boards

• The hardware gap remains strategically important; progress in affordable models has not eliminated China’s dependence on advanced components.
• South Korea is a critical beneficiary of the AI Supercycle; its export data reinforce the importance of memory and semiconductor supply chains.
• Manufacturing yields are becoming a competitive indicator, directly influencing the cost and scalability of Chinese AI infrastructure.
• Strong infrastructure demand does not guarantee strong returns; capital efficiency and free-cash-flow conversion remain essential.
• The next decisive U.S.–China inflection will concern the total cost of intelligence: model prices, hardware, energy and deployment evaluated together.

The latest evidence does not overturn our thesis; it strengthens its industrial and financial foundations. America continues to benefit from its frontier-model ecosystem and access to advanced computing infrastructure. China continues to challenge the cost and diffusion of intelligence, but remains constrained by important semiconductor dependencies. The competition will increasingly be determined by the ability to combine technological capability, manufacturing efficiency and industrial-scale deployment.

Principal sources: Reuters — Korea’s October semiconductor exports, November 1 · Reuters — China’s advanced-memory constraints, October 28 · Reuters — Bank of Korea’s AI investment risk assessment, October 27.

The United States leads in frontier intelligence. China is challenging the economics of its diffusion. But the next decisive frontier is the industrial cost of producing intelligence at scale.