Samsung expects the memory shortage to end by 2028, and that changes how the AI boom looks

Samsung, the biggest player in the memory market, expects the current shortage to end by 2028. In the author's reading, that is a far more important signal than a routine semiconductor forecast. If a company that earns as much from the AI buildout as almost anyone is preparing in advance for cooling demand, the industry is seriously considering a scenario in which the current boom in data centers and accelerator purchases does not keep growing indefinitely.

The key context is explosive demand for memory, without which modern AI infrastructure simply does not work. That means DRAM, VRAM and especially HBM, which accelerators and server systems need for training and inference of models. In recent months prices for these components have jumped, and according to the article, some consumer RAM has become several times more expensive. The market has even earned its own nickname: RAMpocalypse.

Three manufacturers benefit most: Samsung, SK Hynix and Micron. They sit at the base of the whole supply chain, selling memory to companies that then build AI hardware and scale cloud infrastructure.

But high margins do not mean the manufacturers are ready to add capacity recklessly. The article cites a SEMI forecast: output of 300-mm wafers for this kind of memory will grow roughly 7% a year from the end of 2024 through 2028. Meanwhile, in 2026 AI data centers may consume up to 70% of the memory chips produced in the relevant categories.

That creates an odd situation: demand looks enormous, yet any misjudgment of the future would be far too costly. If companies pour money into new fabs and lines and the market then cools, they are left with expensive capacity, oversupply and pressure on profits.

That is why the article treats Samsung's forecast of an end to the shortage by 2028 as a sign of a cautious turn. All the more so because, as the publication states, back in January the company was considering a notable expansion of memory production by the end of the year. Now the emphasis is shifting: not chasing short-term frenzy at any cost, but hedging against weaker demand in two or three years.

This matters for the market because Samsung is not a bystander but one of the main beneficiaries of the AI boom. If even such a player builds a downturn scenario into its plans, the risk no longer looks theoretical.

The author ties this caution to a broader problem across the AI economy. His thesis is that part of the growth is fed by circular financing: model developers, cloud platforms and chip suppliers invest in one another at the same time, and so sustain demand inside a closed system. Against this backdrop, companies that spend huge sums on training and running models but do not yet show comparable returns look especially vulnerable.

Read this way, the memory market is not just a consequence of the AI boom but one of the best indicators of its durability: while data centers are being built aggressively, the shortage persists; if the investment cycle breaks, the memory crunch starts to fade quickly.

The main conclusion follows. Samsung's forecast does not by itself prove that the AI bubble will necessarily burst, or that it will happen in a specific year. Large corporations always build several scenarios and hedge against overheating in advance. But the fact that one of the largest memory makers is already thinking about the end of the shortage and does not want to expand recklessly changes the picture. The story of endless AI infrastructure growth starts to look less like the only possible future and more like a bet with a noticeable risk of reversal. For the market as a whole, that is arguably more important than any loud slogan about the next wave of AI.