Samsung Expects Memory Shortage to End by 2028—Signaling a Shift in AI Growth Expectations

Samsung, a major player in the memory market, expects the current shortage to end by 2028—and in the author's logic, this is a far more significant signal than just a semiconductor forecast. If the company that profits from AI infrastructure building arguably more than anyone is preparing in advance for cooling demand, it means the industry is seriously considering a scenario where the current data center boom and accelerator purchases won't grow infinitely. The key context here is explosive demand for memory, without which modern AI infrastructure simply doesn't work.

This refers to DRAM, VRAM, and especially HBM, which is necessary for accelerators and server systems for training and running inference on models. In recent months, prices for such components have risen sharply, and some consumer-grade RAM has gotten significantly more expensive, according to data cited in the article. Against this backdrop, the market has even earned its own nickname—RAMpocalypse.

Three manufacturers benefit most from this: Samsung, SK Hynix, and Micron. They form the foundation of the entire supply chain, selling memory to companies that then assemble AI hardware and scale cloud infrastructure. But high margins don't mean manufacturers are ready to blindly increase capacity.

The article cites a SEMI forecast: output of 300mm wafers for such memory will grow at roughly 7% annually from late 2024 through 2028. Meanwhile, AI data centers are expected to consume up to 70% of the produced memory chips in the corresponding categories by 2026. This creates a strange situation: demand seems enormous, but any miscalculation in assessing future demand will be too costly.

If companies sharply invest in new fabs and production lines, then the market cools, they'll be left with expensive capacity, oversupply, and margin pressure. This is exactly why Samsung's forecast about ending the shortage by 2028 is interpreted in the article as a sign of cautious reversal. Especially since as recently as January, the company was supposedly considering significant expansion of memory production by year-end.

Now the emphasis is shifting: don't chase short-term frenzy at any cost, but hedge against weaker demand in two or three years. For the market, this matters because Samsung isn't a disinterested observer but one of the main beneficiaries of the AI boom. If even such a major player is factoring in a downturn scenario, the risk no longer looks theoretical.

The article's author connects this caution to a broader problem across the entire AI economy. His thesis is that part of the growth is being fueled by circular financing: model developers, cloud platforms, and chip suppliers simultaneously invest in each other and thereby sustain demand within a closed system. Against this backdrop, companies that spend enormous sums on training and operating models but haven't yet shown comparable returns look particularly vulnerable.

In this reading, the memory market becomes not just a consequence of the AI boom, but one of the best indicators of its sustainability: as long as data centers are aggressively built out, shortages persist; if the investment cycle breaks, the memory shortage quickly begins to ease. This leads to the main conclusion. Samsung's forecast by itself doesn't prove that the AI bubble will necessarily burst or that it will happen in any specific year.

Large corporations always build several scenarios and hedge against overheating in advance. But the very fact that one of the largest memory manufacturers is already thinking about the end of shortages and doesn't want to expand recklessly changes the optics. The story about endless AI infrastructure growth is starting to look not like the only possible future, but like a bet with substantial downside risk.

And for the entire market, this is perhaps more important than any loud proclamation about the next wave of AI.