Abstract:
The explosive expansion of the global artificial intelligence industry is setting off an unprecedented battle for resources in the semiconductor supply chain. According to the latest industry survey report on the supply chain, the conventional memory inventory levels of global memory chip duo Samsung Electronics and SK Hynix have fallen off a cliff. The available days of spot inventory of the two giants have both fallen below the extreme warning line of 10 days. This indicates that the global memory chip market is falling into a historic supply shortage that has been rare in recent years.
In the traditional semiconductor production and sales cycle, memory chip manufacturers usually maintain safety inventory levels for 4 to 6 weeks or even longer to stabilize market supply and demand fluctuations, cope with sudden maintenance of production lines, and meet customers' flexible order needs. However, the latest data shows that the current high-performance DRAM and enterprise-level storage inventory cycles of the two South Korean giants have been sharply compressed to the extreme state of a little more than a week. The main reason for the rapid emptying of warehouse shelves is that the world's major cloud service providers, data center operators and AI server giants are aggressively purchasing goods regardless of cost in order to seize large-scale model computing power infrastructure.
This extremely tight situation is largely due to the squeezing effect of upstream wafer production capacity towards high-bandwidth memory (HBM). In order to meet the nearly unlimited throughput requirements of accelerators such as NVIDIA and AMD for HBM3E and next-generation HBM4 chips, Samsung and SK Hynix have fully allocated a large number of the most advanced advanced process DRAM wafer production lines to HBM's stack manufacturing in the past few quarters. Since the physical manufacturing process of HBM chips consumes several times the wafer area of conventional DRAM, and the packaging complexity is extremely high, this directly results in the production capacity of conventional DDR5 and LPDDR5X memory used in the general server, consumer PC and smartphone markets encountering severe derivative squeeze, triggering a chain supply shortage across the entire product line.
In addition to the tightening of supply caused by production capacity vacancies, the panic hoarding mentality of downstream customers has further accelerated the inventory crisis. Faced with the ever-extending delivery cycle and expectations of price increases, major leading technology companies have increased long-term agreement purchase volumes and locked orders in advance, leaving storage manufacturers almost in a zero-turnover buffer state of "shipping as soon as they are offline." Industry analysts predict that extremely scarce inventory will give original manufacturers unprecedented bargaining power. It is expected that server DRAM and high-end general-purpose memory prices will usher in a new round of jump increase in the second half of 2026, and contract quotations may maintain double-digit high increases for several consecutive quarters.
Semiconductor industry observers pointed out that the inventory fell below 10 days, indicating that the fragility of the global memory supply chain has been pushed to the critical point. Unexpected power outages, earthquake disturbances or equipment maintenance failures at any single production base are very likely to trigger a violent price tsunami and delivery default risk in the spot market. Although both Samsung and SK Hynix are planning capacity expansion plans for new wafer fabs, the move-in, commissioning and yield ramp-up of new machines require a long physical cycle. In the foreseeable next few quarters, the supply-demand imbalance dominated by the AI computing power arms race and the industry-wide core shortage pain may not be fundamentally alleviated.
Comments