Memory Mayhem: Elevated Prices and AI’s Demand
Right now, it’s a prime time to be involved in the memory sector, akin to discovering cash in an old coat. With the AI datacenter industry flourishing, revenues for SK Hynix and Micron have surged threefold over the past year, and Samsung’s figures have nearly doubled. However, despite their windfall from the AI transformation, the scenario can quickly shift. This trend is typical in the memory industry historically.
Currently, the soaring demand for high-bandwidth memory (HBM), DDR5, and NAND flash memory required for GPU servers has consumed any remaining production capacity. This situation has led to shortages that have inflated prices for everything from consumer electronics to AI infrastructure. Nowadays, even a budget smartphone is hard to come by.
Major Investments, Major Setbacks
The leading three memory suppliers are now in the midst of pouring hundreds of billions into expanding new fabrication capacity. In June, South Korean President Lee Jae Myung unveiled a $576 billion investment spearheaded by SK Hynix and Samsung aimed at enhancing chip output and reinforcing AI supply chains. On Thursday, Micron announced it would invest as much as $3 billion to fortify the US semiconductor supply chain, and recent reports indicate that the Idaho-based chipmaker is also enhancing production across its facilities in Singapore, Taiwan, and Japan.
The Long Journey Ahead
Sadly, this is a lengthy endeavor. Semiconductor production ranks among the most intricate and resource-demanding fields globally, and establishing a new DRAM or NAND flash wafer fabrication facility is a significant challenge. Before any chips leave the assembly line, financing has to be arranged, a site chosen, permits obtained, and tens of millions of dollars’ worth of support facilities—from power conditioning to ultra-pure water filtration—must be put in place. Even after clean rooms are established, hundreds of millions of dollars need to be allocated for specialized lithography, wafer transport, and testing equipment to be installed and validated. Once everything is prepared for activation, it can take months to fine-tune all aspects and achieve acceptable yield levels. This entire process can span years, even in the absence of delays.
The Delay Dilemma
Thus, while several new memory fabrication facilities are already underway, anything inaugurated today by SK, Samsung, or Micron will take at least three years to become operational, with an even longer timeline for production ramp-up. As a result, memory prices are set to remain elevated for the foreseeable future. A recent IDC report warns that relief from the RAMpocalypse might not arrive until at least 2028. This bodes well for memory producers, whose revenues will continue to be robust. However, it poses significant challenges for AI startups and model developers, who will bear the brunt of increased infrastructure costs until the situation changes.
OpenAI and several other companies have invested the last four years and hundreds of billions in venture capital into creating increasingly capable models, agents, and tools. The question is no longer whether the technology is effective, but if the benefits warrant ongoing investment at present or elevated levels. Eventually, these startups must turn a profit, and the soaring memory prices do not assist in achieving any semblance of margin in the cost per token.
Can Memory Meet Demand?
The pressing question now is whether the memory manufacturers can bring new capacity online before the major AI enterprises deplete their VC-funded runway and the music stops. Traditionally, memory is a commodity, experiencing drastic price fluctuations typical of boom and bust cycles. Thus, memory vendors rely on boom periods to finance new fabs, acutely aware that once operational, this additional capacity might collapse their pricing.
The Shift in Dynamics
As noted late last year, the AI surge has significantly altered this landscape. Rather than expecting memory prices to drop throughout 2025 and 2026, we have observed the opposite as AI infrastructure consumes every ounce of DRAM and NAND available. However, if the anticipated demand for AI doesn’t materialize, the industry will falter, and memory vendors could find themselves facing an unprecedented market slump. On a positive note, the soaring price of memory will no longer be the reason preventing you from buying a new laptop or smartphone.
Conclusion: Memory, the Tech Yo-Yo
So there you have it. Memory behaves like a yo-yo, rising and falling unpredictably. Beneficial for sellers at the moment, not so convenient for those of us seeking a new gadget without parting with a kidney. Stay alert, as this ride is far from over.
