Assume the RAM Disaster Is Unhealthy Now? It Simply Retains Getting Worse

Unless the AI bubble bursts and sends all these deliberate information middle initiatives packing, you received’t see an finish to reminiscence pricing woes anytime quickly. One of the massive three semiconductor corporations able to making reminiscence doesn’t think about will probably be in a position to meet all client demand for not less than one other 4 years.

The DRAM and NAND reminiscence market is dominated by three main gamers, Micron, Samsung, and SK Hynix. The latter falls beneath the worldwide conglomerate SK Group. The South Korean-based RAM maker’s chairman, Chey Tae-won, instructed Bloomberg that the corporate is increasing memory-making capability. Unfortunately, it received’t be capable to meet demand till “around” 2030.

Chey spoke to reporters exterior of GTC 2026 earlier this week. He mentioned that capability for the fundamental wafers the corporate makes use of for its chips is lagging 20% behind demand. All this inevitably trickles right down to shoppers. We’ve seen costs enhance for virtually each gadget, from laptops to smartphones just like the latest Samsung Galaxy S26, all the way in which right down to enthusiast-level miniature computing chips just like the Raspberry Pi.

The blame for the reminiscence scarcity falls on the toes of the AI increase. The largest AI datacenter initiatives, like OpenAI’s multi-state Stargate challenge, have such an enormous demand for high-bandwidth reminiscence (HBM). Semiconductor corporations are making such a revenue from these high-end reminiscence chips that they’ve lowered capability for consumer-level DRAM and SSD storage. Even main corporations like Valve are struggling to supply inexpensive RAM for its Steam Machine. One of the corporate’s employees reportedly joked to trade insiders at GDC 2026, “If you have a line on a bunch of RAM, we are in the market and would like to buy it.”

RAM makers aren’t eager on rising provide

The motive why the key semiconductor corporations are transferring slowly to extend provide is—satirically—a concern of getting an excessive amount of RAM. Korean outlet Chosun Biz reported, primarily based on nameless trade sources, that Samsung hopes the worldwide semiconductor market could reverse course in 2028. It reportedly doesn’t wish to scale capability a lot due to “uncertainties in demand forecasting.”

Essentially, Samsung could also be involved about scaling up too quick after which the RAM provide being an excessive amount of as soon as AI information middle demand for high-bandwidth reminiscence falters. Currently, 50% of SK Hynix and Samsung reminiscence output is within the type of HBM slated for AI information facilities.

SK Hynix is increasing capability, in response to Chosun Biz. It’s constructing extra facilities situated round Korea resembling Icheon, Cheongju, and Yongin. It’s spending near $13 billion on an enormous new meeting plant made solely to produce HBM. The work is ready to begin in April and never be completed till the top of 2027, in response to Reuters.

A single Nvidia Vera Rubin chip—the corporate’s newest, strongest AI coaching SoC (system on a chip)—requires as a lot as 288GB of HBM. That’s 9 instances as a lot as your typical gaming-ready PC. The older Grace Blackwell-based B300 chips require simply barely much less reminiscence. And this AI increase is how Nvidia plans to drive $1 trillion in income.

Data facilities plan to stack tons of of those chips collectively, only for the sake of AI cloud compute. Until demand for AI information facilities bottoms out, we’ll be in for years of inflated gadget costs.