The world of computing has long been grappling with a persistent crisis: the memory and storage shortage, often dubbed the RAMpocalypse. For hardware enthusiasts and server administrators, rising prices and constrained supply have defined the landscape. But amidst this ongoing crunch, some industry giants are suggesting that the economics of the market have undergone a fundamental shift, driven by the explosive growth of artificial intelligence.
At the International Supercomputing Conference this past week, one major player offered a bracing perspective on the future: Lenovo reportedly said that memory prices will never return to the lows seen in previous years. This statement, framed within their presentation, hints at a complex reality where demand—specifically from AI infrastructure—is set to redefine supply dynamics for years to come.
The company’s reasoning is rooted in market economics. They argue that even as new manufacturing capacity comes online starting around 2028, the enormous demand generated by AI will absorb much of that additional output. This suggests that the memory industry’s growth is now tightly coupled with AI infrastructure spending.
This view is supported by major memory producers. For instance, SK hynix has announced plans to triple its memory production capacity by 2034, illustrating the massive investment pouring into expanding supply chains. This effort aligns with the expectation that manufacturers will prioritize maximizing output for AI workloads rather than seeking the razor-thin margins that characterized earlier market cycles.
The long-term forecasts from these companies paint a clear picture: while supply constraints are expected to persist through at least 2027, gradual improvements will only begin in 2028 or later. These predictions are backed by multi-year supply agreements worth billions, demonstrating how seriously hyperscalers are securing their memory needs for the long haul.
This shifting landscape has also changed how system architects approach server design. With new dual-socket servers incorporating up to 16 memory channels per processor in the near future, maximizing bandwidth is critical. This means that even modest configurations require significantly more installed memory to fully exploit available capabilities, making capacity planning a much more important factor than ever before.
Furthermore, the focus on AI has inadvertently steered manufacturers toward a new hierarchy of memory types. To meet intense AI accelerator demands, DRAM makers have redirected production capacity toward higher-margin High Bandwidth Memory (HBM). This dynamic is eating into the supply of conventional system memory like DDR5 and LPDDR5.
Ironically, this shift means that while demand for all types of memory remains high, the relative value proposition of HBM has increased. The economic pressure has caused a re-evaluation: maximizing GPU utilization using HBM may now make more financial sense than simply packing more conventional system DRAM into the host machine.
Ultimately, the narrative is evolving from a pure supply crunch to an optimization challenge. As hyperscalers continue pouring billions into AI infrastructure and vendors prioritize high-margin enterprise products, the future of memory won’t just be about scarcity; it will be about intelligent design that maximizes efficiency, ensuring that the next generation of computing runs smoothly and profitably.
Credit: Tom’s Hardware
