The AI frenzy is driving a memory chip supply crisis
"The memory shortage has now graduated from a component-level concern to a macroeconomic risk," said Sanchit Vir Gogia, CEO of Greyhound Research
Highlights:
- Memory shortage could delay AI projects, productivity gains
- SK Hynix predicts memory shortage to last through late 2027
- Smartphone makers warn of price rises due to soaring memory costs
An acute global shortage of memory chips is forcing artificial intelligence and consumer-electronics companies to fight for dwindling supplies, as prices soar for the unglamorous but essential components that allow devices to store data.
Japanese electronics stores have begun limiting how many hard-disk drives shoppers can buy. Chinese smartphone makers are warning of price increases. Tech giants including Microsoft, Google and ByteDance are scrambling to secure supplies from memory-chip makers such as Micron, Samsung Electronics and SK Hynix, according to three people familiar with the discussions.
Keep updated, follow The Business Standard's Google news channel
The squeeze spans almost every type of memory, from flash chips used in USB drives and smartphones to advanced high-bandwidth memory (HBM) that feeds AI chips in data centers. Prices in some segments have more than doubled since February, according to market-research firm TrendForce, drawing in traders betting that the rally has further to run.
The fallout could reach beyond tech. Many economists and executives warn the protracted shortage risks slowing AI-based productivity gains and delaying hundreds of billions of dollars in digital infrastructure. It could also add inflationary pressure just as many economies are trying to tame price rises and navigate US tariffs.
"The memory shortage has now graduated from a component-level concern to a macroeconomic risk," said Sanchit Vir Gogia, CEO of Greyhound Research, a technology advisory firm. The AI build-out "is colliding with a supply chain that cannot meet its physical requirements."
This Reuters examination of the spiraling supply crisis is based on interviews with almost 40 people, including 17 executives at chipmakers and distributors. It shows industry efforts to meet voracious appetite for advanced chips — driven by Nvidia and tech giants like Google, Microsoft and Alibaba — created a dual bind: Chipmakers still can't produce enough high-end semiconductors for the AI race, yet their tilt away from traditional memory products is choking supply to smartphones, PCs and consumer electronics. Some are now hurrying to course-correct.
Details of the global scramble by tech firms and price increases described by electronics retailers and component suppliers in China and Japan are reported here for the first time.
Average inventory levels at suppliers of dynamic random-access memory (DRAM) — the main type used in computers and phones — fell to two to four weeks in October from three to eight weeks in July and 13 to 17 weeks in late 2024, according to TrendForce.
The crunch is unfolding as investors question whether the billions of dollars poured into AI infrastructure have inflated a bubble. Some analysts predict a shakeout, with only the biggest and financially strongest companies able to stomach the price increases.
One memory-chip executive told Reuters the shortage would delay future data-center projects. New capacity takes at least two years to build but memory-chip makers are wary of overbuilding for fear it could end up idle should the demand surge pass, the person said.
Samsung and SK Hynix have announced investments in new capacity but haven't detailed the production split between HBM and conventional memory.
SK Hynix has told analysts that the memory shortfall would last through late 2027, Citi said in November.
"These days, we're receiving requests for memory supplies from so many companies that we're worried about how we'll be able to handle all of them. If we fail to supply them, they could face a situation where they can't do business at all," Chey Tae-won, chairman of SK Hynix parent SK Group, said at an industry forum in Seoul last month.
OpenAI in October signed initial deals with Samsung and SK Hynix to supply chips for its Stargate project, which would require up to 900,000 wafers per month by 2029. That's about double current global monthly HBM production, Chey said.
Samsung told Reuters it is monitoring the market but wouldn't comment on pricing or customer relationships. SK Hynix said it is boosting production capacity to meet increased memory demand.
Microsoft declined to comment and ByteDance didn't address questions about the chip strain. Micron and Google didn't respond to comment requests.
'Begging for supply'
After ChatGPT's release in November 2022 ignited the generative AI boom, a global rush to build AI data centers led memory makers to allocate more production to HBM, used in Nvidia's powerful AI processors.
Competition from Chinese rivals making lower-end DRAM, such as ChangXin Memory Technologies, also pushed Samsung and SK Hynix to accelerate their shift to higher-margin products. The South Korean firms account for two-thirds of the DRAM market.
Samsung told customers in 2024 that it planned to end production of one type of DDR4 chips — an older variety used in PCs and servers — this year, according to a letter seen by Reuters. (The company has since changed course and will extend production, two sources said.) In June, Micron said it had informed customers it would stop shipping DDR4 and its counterpart LPDDR4 — a type used in smartphones — in six to nine months.
ChangXin followed suit in ending most DDR4 production, one source said. The firm declined to comment.
This shift, however, coincided with a replacement cycle for traditional data centers and PCs, as well as stronger-than-expected sales of smartphones, which rely on conventional chips.
In hindsight, "one could say the industry was caught off-guard," said Dan Hutcheson, senior research fellow at TechInsights.
Samsung raised prices of server memory chips by up to 60% last month, Reuters has reported. Nvidia CEO Jensen Huang, who in October announced deals and shared fried chicken with Samsung Electronics Chairman Jay Y. Lee during a trip to South Korea, acknowledged the price surge as significant but said Nvidia had secured substantial supply.
Google, Amazon, Microsoft and Meta in October asked Micron for open-ended orders, telling the company they will take as much as it can deliver, irrespective of price, according to two people briefed on the talks.
China's Alibaba, ByteDance and Tencent are also leaning on suppliers, dispatching executives to visit Samsung and SK Hynix in October and November to lobby for allocation, the two people and another source told Reuters.
"Everyone is begging for supply," one said.
The Chinese firms didn't address questions about the chip crunch. Nvidia, Meta, Amazon and OpenAI didn't respond to requests for comment.
In October, SK Hynix said all its chips are sold out for 2026, while Samsung said it had secured customers for its HBM chips to be produced next year. Both firms are expanding capacity to meet AI demand, but new factories for conventional chips won't come online until 2027 or 2028.
Shares in Micron, Samsung and SK Hynix have rallied this year on chip demand. In September, Micron forecast first-quarter revenue above market estimates while Samsung in October reported its biggest quarterly profit in more than three years.
Consultancy Counterpoint Research expects prices of advanced and legacy memory to rise by 30% through the fourth quarter and possibly another 20% in early 2026.
Smartphone sticker shock
Chinese smartphone makers Xiaomi and Realme have warned they may have to raise prices.
Francis Wong, Realme India's chief marketing officer, told Reuters the steep increases in memory costs were "unprecedented since the advent of smartphones" and could force the company to lift handset prices by 20% to 30% by June.
"Some manufacturers might save costs on imaging cameras, some on processors, and some on batteries," he said. "But the cost of storage is something all manufacturers must completely absorb; there's no way to transfer it."
Xiaomi told Reuters it would offset higher memory costs by raising prices and selling more premium phones, adding that its other businesses would help cushion the impact.
In November, Taiwanese laptop maker ASUS said it had about four months of inventory, including memory components, and would adjust pricing as needed.
Winbond, a Taiwanese chipmaker with around 1% of the DRAM market, was among the first to announce a capacity expansion to meet demand. Its board of directors approved a plan in October to sharply boost capital expenditure to $1.1 billion.
"Many customers have been coming to us saying, 'I really need your help,' and one even asked for a six-year long-term agreement," Winbond's President Pei-Ming Chen said.
In Tokyo's electronics hub of Akihabara, stores are restricting purchases of memory products to curb hoarding. A sign outside PC shop Ark says that since 1 November customers have been limited to buying a total of eight products across hard-disk drives, solid-state drives and system memory. Ark declined to comment.
Clerks at five shops said shortages had pushed prices sharply higher in recent weeks. At some stores, one-third of products were sold out.
Products such as 32-gigabyte DDR5 memory – popular with gamers – were over 47,000 yen, up from around 17,000 yen in mid-October. Higher-end 128-gigabyte kits had more than doubled to around 180,000 yen.
The hikes are driving customers to the secondhand market — benefiting people like Roman Yamashita, owner of iCON in Akihabara, who said his business selling used PC parts is booming.
Eva Wu, a sales manager at component trader Polaris Mobility in Shenzhen, said prices are changing so rapidly that distributors issue broker-style quotes that expire daily — and in some cases hourly — versus monthly before the crunch.
In Beijing, a DDR4 seller said she had hoarded 20,000 units in anticipation of further increases.
Some 6,000 miles away in California, Paul Coronado said monthly sales at his company, Caramon, which sells recycled low-end memory chips pulled from decommissioned data-center servers, have surged since September. Almost all its products are now bought by Hong Kong-based intermediaries who resell them to Chinese clients, he said.
"We were doing about $500,000 a month," he said. "Now it's $800,000 to $900,000."
