At the turn of the year, a new bottleneck in artificial intelligence development emerged. The shortage of processors dedicated to training models eased, but with the rapid growth in AI usage, the demand for various types of memory chips sharply increased. These are necessary not only for the operation of models but also for storing rapidly growing results.
Read more 10-year-old rollerbladed onto the streets in Virbalis and ended up in a hospital in Kaunas
Rise and fall
Until now, the memory chip business was considered quite boring, cyclical, and highly competitive. Companies’ margins and valuations were relatively low because manufacturers had no clear competitive advantage – for the user, it basically did not matter whose memory chip was in their computer or phone.
However, with the rapid construction of data centers, it became clear that existing memory chip production capacities were insufficient. Memory prices rose 6–10 times in a short period, production was sold out a year in advance, and profit margins soared towards 70%.
Such conditions turned memory chip manufacturers’ stocks into some of the main drivers of market growth. Shares of Micron Technology, SK Hynix, SanDisk, and Samsung Electronics increased 3–10 times in the first half of the year. Companies reported record profits, and by traditional valuation metrics, they even appeared cheap because their results grew not by percentages but by multiples per year.
The majority of technology giants’ investments in AI infrastructure were directly felt by chip manufacturers. The South Korean market, dominated by these electronics sector names, became one of the fastest-growing in the world – it doubled in six months, and at its peak, Micron became the eighth-largest company in the US. In fact, due to massive investments, most of the technology giants’ profits were transferred to chip manufacturers.
But the miracle did not last long. Investors began to critically assess the huge expenses on data centers, memory price growth slowed, and AI model developers sought ways to use memory more efficiently. Moreover, seeing a highly profitable market niche, manufacturers began aggressively increasing capacities, and China announced faster memory chip production expansion.
Fear arose in the market that the current deficit could quickly turn into a surplus. The chip manufacturers’ index fell about 30% from its peak, with market leaders like SK Hynix or Sandisk dropping 40–60%. The South Korean market lost about 40% of its value in roughly 40 days – a faster decline than during the 2008 crisis.
The market reaction, seeing still record profits, may seem exaggerated. However, in cyclical industries, the most important factor is not the profit level itself but the moment when profit growth begins to slow. In this case, the market simply expected the very favorable period to last longer.
AI development paradoxes
Although the current sell-off is very painful, it is certainly not the first and is full of paradoxes. Its intensity was determined not only by fundamentals but also by technical reasons. South Korean retail investors particularly like to invest with borrowed money and speculate dangerously in other ways – this gave additional momentum during the rise but equally intensified the market decline. In such situations, painful fluctuations are possible even without changes in companies’ prospects.
On the other hand, the sector itself is changing. Several open-source Chinese AI models have appeared that are not far behind the leading OpenAI or Anthropic models. On one hand, this could accelerate the AI arms race between China and the US and require even more investments – such a situation would continue to favor memory and other infrastructure manufacturers. On the other hand, fierce competition among model developers would mean lower AI prices for users and even worse investment returns for advanced model developers.
Read more After news about deadly dangerous bacteria in the Baltic Sea – expert commentary
According to current trends, it seems we will have more and cheaper artificial intelligence, so the greatest value should be captured by AI users. AI appears to be commoditizing – becoming a raw material. This would mean both greater demand for memory chips and rapidly growing supply. Margins in this sector will inevitably shrink, and exponential profit growth will no longer occur, but structurally higher memory demand still means better sector prospects than a few years ago.
Where will the AI boom turn?
AI sector trends are changing very quickly: questions constantly arise about whether this is a bubble, new market leaders emerge, and new development constraints appear.
There are many signs that this investment cycle is already mature and will not be able to grow as fast. Additional investments are financed by debt; technology giants’ revenues are no longer sufficient. The bond market has started to view this debt suspiciously – bond yields are rising while overall borrowing conditions remain favorable, and credit default swaps have begun to rise sharply.
Monetary policy can play a crucial role here. Investment cycles of this scale usually end when interest rates suddenly increase. Currently, the market expects a mild interest rate rise in the US, but if the new FED chair does not control inflation and interest rates rise 1–2% in a short time, it would stop or at least significantly slow investments in data centers.
Investors are closely watching the investment plans of major technology companies. So far, they continue to grow, but the market is increasingly critical of rapidly rising expenses. The “Magnificent Seven” stocks have fallen nearly 5% since the beginning of the year and are certainly no longer the market darlings they once were.
If investments continue to grow rapidly, it will mean further profit growth for memory chip and other data center infrastructure manufacturers. However, if companies decide these investments are not paying off, semiconductor manufacturers will face a painful sell-off.
Still, everything is not yet clear. Microsoft’s results show that cloud revenues are growing very rapidly, and model developers are looking for ways to monetize AI usage. So far, at least some hyperscalers manage to achieve positive returns rather than just burning money.
Avoiding major monetary policy shocks, the investment cycle, though slowing, may continue. But the longer it lasts, the more painful the correction may be in the future.
Read more The baby death that shocked Lithuania drew attention to a rare disease: what is Patau syndrome?
