The price increases driven by the AI boom show no signs of abating in the short to medium term and are starting to impact consumer electronics. Some see it as an additional factor likely to trigger a systemic crisis.
For months, the phenomenon has been worrying. We are now seeing the concrete repercussions. During June, Apple announced significant price increases on its products (15 to 20% for Macs, 15 to 25% for iPads), justified by an unprecedented increase in memory costs. The Apple group’s shares fell sharply on the stock market following these announcements. But one person’s misfortune is another’s happiness. Memory chip suppliers are rubbing their hands.
Full box for memory chip specialists
Barely three years after recording the worst losses in its history, memory chip specialist Micron is today one of the stars of Wall Street and has joined the very closed circle of companies valued at more than a trillion dollars. The company’s profits in the latest quarter are almost fifteen times higher than they were a year earlier in the same period, well beating Wall Street predictions. It is expected to achieve nearly $100 billion in revenue over the next twelve months, more than Meta or Berkshire Hathaway. One of Micron’s most popular memory chips in data centers sells for around $1,300 today, up from $350 a year ago, according to Circular Technology, a Massachusetts-based seller of data center hardware.
The Korean SK Hynix, another memory chip giant, is not left out, since according to analysts’ forecasts, SK Hynix’s turnover could triple this year. Its profits could be multiplied by more than five. The company, like Micron, has exceeded $1,000 billion on the stock market.
And Samsung, the third major manufacturer of memory chips, is not left behind. The Korean giant posted a net profit of 28.2 billion euros between January and March, up 755% compared to profits generated in the first quarter of 2025.
A market driven by the AI boom
Tensions on the memory market are mainly linked to the frenzied construction of data centers by AI giants. In addition to the graphics processors (GPUs) provided by Nvidia and AMD, the latter need memory chips to operate. “All the power of a GPU is linked to its ability to manage parallelized calculations directly on memory without needing to access a disk,” explains Philippe Charpentier, director of Solutions Engineering at NetApp, a specialist in data storage. Each Nvidia GPU has high bandwidth memory (HBM). Built for AI, it accelerates processing speeds while reducing energy consumption. This market segment is largely dominated by SK Hynix, which made the strategic choice to position itself in this area around ten years ago, a bet which has paid off.
In addition to the memory of the graphics processors themselves, an AI data center also requires server RAM, or dynamic random access memory (DRAM). A market shared by Samsung (leader with around 40% market share), SK Hynix and Micron, which between them hold 95% of the market. This temporarily stores the data used by the processor or the server, with very fast access, but volatile, that is to say it is erased when the device turns off. Each GPU server can carry hundreds of GB of classic DRAM for host CPUs, data management and exchanges between GPUs.
There remains storage, another determining element, which relies on NAND flash memory. Slower than DRAM, it retains data without power and is therefore used for permanent storage of training datasets, model checkpoints and logs. An LLM training cluster can require petabytes of flash storage.
The SemiAnalysis firm estimates that memory will represent this year 30% of investment spending by hyperscalers, compared to only 13.5% in 2025.
“Whether we are talking about HBM, NAND or DRAM memory, they are almost the same components that are inside. So the factories can only produce a limited capacity in total, and an increase in demand in one of the sectors also affects the others,” specifies Philippe Charpentier. The company says prices for QLC drives, a type of flash memory, doubled in the last quarter alone, and the ceiling may not have been reached yet.
No rapid improvement on the horizon
However, where demand from Big Tech has accelerated strongly and rapidly, chip production is not very elastic. Increasing production requires making significant investments that take years to bear fruit. “Building a semiconductor factory represents billions of dollars of investment and takes years. Which means that for next year and the one that follows, the shortage is already recorded,” notes Philippe Charpentier.
Micron was allocated more than $6 billion in direct manufacturing subsidies from the US government as part of the Chips Act of 2022, to support its investments in new factories. Last month, the company hosted Commerce Secretary Howard Lutnick and other top U.S. officials at an event in Virginia, where it is starting to make some of its most advanced memory chips at its Manassas factory.
Samsung and SK Hynix are planning an expansion of their chip manufacturing capacities to the tune of $600 billion via new factories in South Korea, with the aim of doubling the quantity of DRAM chips produced by the country within five years.
A reversal likely to cause a systemic crisis?
Some are beginning to worry about the seemingly limitless rise in memory prices, which, combined with the AI bubbleforms a particularly dangerous combination. Indeed, the memory market is traditionally a cyclical market, marked by periods of high demand, which lead to an increase in production, which quickly leads to overproduction. And the rise in prices, passed on to consumer electronics goods, leads to a drop in consumption, and ultimately a drop in memory prices. Such a cycle occurred during Covid, leading Micron and SK Hynix to record the biggest losses in their history.
The problem, which notably prominently market expert James Mackintosh in his Streetwise column is that, unlike consumers who buy computers and smartphones, Big Tech does not allow themselves to be easily discouraged by rising prices: launched into a frantic race for the biggest models, with deep pockets and strong debt capacities, they can absorb the rise in prices, and therefore prolong the cycle… Until a market boom which would ultimately produce prices that were ultimately untenable for them and lead them to reverse their investments, causing them to burst the AI bubble. If we add the fact that these companies are at the same time ready to increase their losses rather than raise the price of their AI subscriptions, for fear of driving their customers away from the competition, we find ourselves facing a particularly dangerous spiral,
“Apple’s price increase sent a negative signal in this market that reflects a deeper concern: with the increase in computing and memory costs, is there a breaking point where companies should slow down these gigantic AI deployments? This ties in with some of the fears surrounding neoclouds and hyperscalers in general,” summarizes Dan Ives, analyst at Wedbush.
Another, more optimistic scenario would be that the bubble will never burst, with the rise of AI having effectively shifted memory from a cyclical market to a structurally rising market. Defenders of this thesis argue in particular that the memory chips used in servers and AI chips are more technical, more strategic and often sold via longer contracts. In a departure from the way memory supply contracts have historically been negotiated, Micron has entered into 16 long-term strategic agreements with some of its largest customers.
Note, however, that such assumptions have already been refuted in the past, as noted in the newsletter Vifargent.iodedicated to financial investment: “This is not the first super-cycle with a favorable market narrative: in 2017/18, the demand for smartphones, with increasing memory content, and data centers, as well as a supposed capacity discipline, had led to expectations of a change in status from a cyclical market to a structurally rising market. The rest was classic: when supply caught up with demand in 2019, the market turned around.” Be careful, therefore…
