If you need a hard disk or RAM, don’t put off your purchase. Because while the world marvels at the prowess of generative AI, a much quieter crisis is playing out in semiconductor factories: AI is now absorbing the bulk of the world’s hardware resources. GPUs, servers and, above all, memory – everything that enables these models to function now accounts for an overwhelming share of global production.
As a result, a severe memory shortage is already underway, and all indications are that it is set to worsen over the coming years. The market is dominated by two players, Samsung and SK Hynix, who between them control almost 70% of the world’s DRAM (the most common type of RAM memory chip). At the same time, Micron – Crucial’s parent company – is withdrawing from the consumer segment, a decision that makes a lot of sense: according to these manufacturers, all production capacity must now be redirected towards AI, even if this means leaving the classic consumer by the wayside.
According to technology company Calian, DRAM prices have quadrupled since January. One supplier cited by Calian, Coosea, shows the surge: the 4 GB DDR4x chip went from $7 for professionals to over $30 in mid-November. Flash memories follow suit: the 64 GB eMMC has risen from $3.2 to over $8.
A gluttonous HBM
Driving a model like GPT-4 doesn’t just require graphics chips (GPUs), it demands a colossal amount of HBM (High Bandwidth Memory), an ultra-fast memory stacked in multiple layers.
Samsung and SK Hynix have reoriented their factories towards HBM3 and HBM3e. As a result, a silicon wafer dedicated to HBM cannot be used to produce your DDR5 RAM or the NAND chips in your next SSD.
As a result, manufacturers have scaled back their consumer offerings to maximize profitability on the AI side, where companies are prepared to pay prices 5 to 10 times higher to secure their supplies.
The situation is similar for SSDs. Enterprise SSDs are now selling at a premium, and this crowding-out effect is having repercussions throughout the NAND chain (the type of flash memory used in SSDs). According to memory module manufacturers quoted by Commercial Times, current inventories will only last until the first quarter of 2026 (some could run dry as early as next March). The first signs are visible: 512GB SSDs downgraded to 256GB and 1TB models reduced to 512GB. Technical choices imposed not by product logic, but by scarcity.
According to TrendForce, contract prices for enterprise SSDs will rise by more than 25% in the fourth quarter of 2026, with the risk of an all-time revenue record for the sector.
And the pressure is also affecting smartphones: Avril Wu, VP Research at TrendForce, points out that demand from hyperscalers (cloud service providers such as AWS, Google Cloud, Microsoft Azure, IBM Cloud, Alibaba Cloud, Oracle, Meta…) is pushing Samsung and SanDisk to delay their NAND deliveries. As a result, the mobile industry is facing the same constraints. Xiaomi Chairman Lu Weibing has already warned that the shortage could push up the price of smartphones by 2026.
A situation that makes everything worse
Added to this pressure created by AI is a cyclical phenomenon: the renewal of post-Covid IT fleets, expected in early 2026. To put it plainly: rising structural demand is arriving at exactly the same time as supply is contracting. Players such as Dell and Lenovo have already warned their partners of imminent price hikes.
Commercial Times is now talking of a true NAND « dry year » in 2026, forcing major OEMs to reduce their SSD capacities to preserve end-user prices.
The idea that « technology always gets cheaper in the end » no longer applies. We’re moving into a period where SSDs won’t go down, entry/mid-range PCs will go up in price, supply lead times for businesses will lengthen, and smartphones will have to absorb the rising cost of memory components.
So why not just build more factories? The answer is simple: fear. Manufacturers – nicknamed the « Dramurai » (a contraction of DRAM and Samurai) – fear an abrupt turnaround in the AI market.
Building a factory costs tens of billions and takes 2 to 3 years. The « Dramurai » fear a downturn in the AI market and prefer to limit supply rather than risk ruinous overcapacity.
For companies, the advice is simple: renew your stocks now, as long as prices remain reasonable. The window could close quickly.
Written in French by Zakaria Choukrallah; edited in English by AngloMedia Group.
