A recent market rumor suggests that Google is exploring purchasing DRAM memory from China's ChangXin Memory Technologies (CXMT) for its next-generation AI accelerators.
The core reason behind this potential move is simple: cost and supply. The global demand for memory chips, especially the high-performance HBM needed for AI, has skyrocketed. This has led to a severe supply shortage and soaring prices. Even Apple's CEO, Tim Cook, recently stated that memory cost inflation is now so high that price increases for their products are "unavoidable." For a company like Google, which plans to deploy millions of AI chips (TPUs), this cost pressure is immense, creating a powerful incentive to find alternative, more affordable suppliers.
However, it's important to understand what this rumor likely entails. Google's TPUs rely on specialized, on-package HBM for maximum performance. CXMT is currently strongest in producing standard DDR5 memory, the type used for general server functions. Therefore, this exploration is probably not about replacing the critical HBM in TPUs but rather about using CXMT's DDR5 in other parts of their data centers. By doing so, Google could free up supply from its primary vendors (Samsung, SK hynix, Micron) for HBM and reduce its overall memory bill.
This is where the story gets complicated, though. The biggest hurdle is the geopolitical tension between the U.S. and China. While the U.S. government recently decided not to add CXMT to its trade blacklist (the Entity List), this could be temporary. More importantly, a U.S. law, Section 5949 of the NDAA, is set to prohibit federal agencies from using services that include semiconductors from CXMT by late 2027. This poses a major compliance risk for Google, as any cloud infrastructure serving U.S. government clients would be affected.
In essence, the situation is a balancing act. On one hand, the economic pressure from the AI-driven memory shortage makes diversifying suppliers a logical step. On the other hand, the looming policy and compliance risks create a significant barrier. The decision to hold off on blacklisting CXMT has opened a narrow window for companies like Google to explore their options, but the long-term path remains uncertain.
- DRAM (Dynamic Random-Access Memory): The standard memory used in most computers and servers to store data for active applications.
- HBM (High Bandwidth Memory): A high-performance type of memory stacked vertically and placed very close to a processor, allowing for much faster data transfer speeds, which is essential for demanding AI tasks.
- TPU (Tensor Processing Unit): Google's custom-designed chip specifically for accelerating AI and machine learning workloads.
- Entity List: A U.S. government trade blacklist that restricts targeted companies from receiving U.S. technology and goods without a special license.