SK Hynix headquarters in Icheon, Gyeonggi Province. (Yonhap)
Semiconductors haven’t succumbed to the winter everybody anticipated. Removed from frosty, the sector is as scorching as midsummer. Morgan Stanley, the funding financial institution whose pessimistic predictions about chipmakers have pushed volatility within the Korean inventory market, has quietly modified its outlook.
Morgan Stanley had predicted a chill for semiconductors in stories final yr titled “Reminiscence – Winter At all times Laughs Final” and “Winter Looms.”
However just lately, the financial institution printed a report titled “Reminiscence Supercycle – Rising AI Tide Lifting All Boats.” The financial institution retracted its forecast of a DRAM value correction by the top of 2025 and predicted that costs could rise by 2027.
For a mea culpa, that comes slightly late. The typical mounted value of DRAM (DDR4 8GB) already started to rebound in April 2025. The value rose steadily every month, greater than quadrupling from US$1.35 in March 2025 to US$6.30 as of early October.
The confounding issue for semiconductor value predictions is synthetic intelligence. As parts in each digital machine, chips are an integral a part of the fashionable financial system.
To make their value forecasts, semiconductor specialists have usually made use of the manufacturing PMI (standing for “buying managers’ index”), an index tracked by the Institute for Provide Administration. When the financial system is robust, chip costs rise; when the financial system is poor, they fall.
Undersupply persists regardless of poor financial system
The PMI has solely been above 50 in three months over the previous three years. Given previous traits, chip costs should fall. However international Massive Tech corporations decided to set future traits have poured large sums of cash into AI whatever the present financial local weather.
A big chunk of that cash has been used to purchase chips. That has produced an undersupply in bizarre DRAM, high-bandwidth reminiscence (HBM), solid-state drives (SSD) for servers and even good old school onerous disk drives (HDD).
This undersupply is because of underproduction, which itself is because of forecast failures. That’s, the chipmakers would have made extra chips in the event that they’d identified the chips would promote.
Deciding how a lot to provide is not any simple job. In an off-the-record dialog that left an enormous impression on me, a senior supervisor at SK Hynix as soon as remarked, “On the introduction of the cellular period, we made an enormous funding primarily based on our prediction {that a} huge market was opening up. Provide shortages solely final two or three years. Nevertheless it’s inconceivable to foretell how far the AI period will go. And that’s true for each consumers and sellers.”
SK Hynix is mainly the only real provider of HBM chips to Nvidia, which has been a runaway success within the AI ecosystem.
It’s tough to outline the parameters of the AI business. One may name the profitable chatbots put out by Massive Tech, corresponding to ChatGPT and Gemini, AI. Folks use them for translation, doc group, and varied inquiries. The extra just lately revealed reasoning fashions carry out the identical duties however require extra computing energy. As an alternative of simply spitting out a solution, it questions and evaluates its reply earlier than providing a response. Creating presentation supplies, drawing photos, making movies and packages — all of those duties are actually in AI’s area. AI will be put to make use of nearly anyplace; the one restrict is a scarcity of creativity.
Nvidia CEO Jensen Huang views the evolution of AI as following a course that goes from notion to reasoning to agentic AI to bodily AI, and argues that every part of this evolution requires a geometrical development of computing energy.
An AI may talk with different AI (A2A) to resolve issues whereas making lodge and restaurant reservations or buying merchandise. Self-driving vehicles could reign over the roads, and robots do the work for people. It doesn’t take a really lively creativeness to consider the sort of human work that AI robots will quickly do. Not solely will they do every little thing a human can — from house responsibilities to manufacturing — they are going to do issues that people can’t.
All of those tasks require reminiscence chips. Reminiscence chip producers like Samsung Electronics and SK Hynix can not construct vegetation in anticipation of conditions that mimic science-fiction eventualities set within the far future. But additionally they can not ignore calls for for merchandise. The numerical figures representing such demand are already past creativeness.
Throughout OpenAI CEO Sam Altman’s current go to to South Korea, he introduced plans to construct AI knowledge facilities with Samsung in Pohang and SK Hynix in South Jeolla. Altman met with Samsung Electronics Chairman Lee Jae-yong and SK Group Chairman Chey Tae-won and introduced his intentions to position orders for DRAM chips to the tune of 900,000 wafers a month till 2029.
Samsung is presently the world’s prime producer of reminiscence chips. But Samsung’s present manufacturing capability for DRAM chips is 600,000 to 650,000 per 30 days. SK Hynix’s month-to-month capability is round 500,000 chips. These two corporations comprise round 70% of the worldwide DRAM market, but Altman mentioned he’s keen to purchase extra chips per 30 days than both firm can produce. However how severe is he, actually?
OpenAI CEO Sam Altman speaks to South Korean President Lee Jae Myung on the Seoul presidential workplace on Oct. 1, 2025. (Yonhap)
Orders piling up
Just a few days after stories of Altman’s pledge, American chipmaker AMD introduced that it’ll provide OpenAI with 6 gigawatts (GW) of GPUs, beginning within the second half of 2026.
OpenAI plans to construct an information heart requiring 1 GW of energy utilizing the MI450 GPUs that AMD expects to launch in 2026. Folks anticipate AMD to generate tens of billions of {dollars} in annual income. It goes with out saying that AMD’s AI semiconductors additionally require reminiscence chips.
In September, Nvidia introduced that it might make investments US$100 billion in OpenAI. Nvidia introduced its “letter of intent for a landmark strategic partnership to deploy not less than 10 gigawatts of NVIDIA techniques for OpenAI’s next-generation AI infrastructure to coach and run its subsequent technology of fashions on the trail to deploying superintelligence.”
To start out off, Nvidia will make investments US$10 billion within the second half of 2026 to construct infrastructure that employs Nvidia’s next-generation AI chip platform, Vera Rubin. Huang informed CNBC that the 10-gigawatt mission with OpenAI is equal to between 4 million and 5 million graphics processing models, across the whole variety of models Nvidia is predicted to ship in 2025 and “twice as a lot as final yr.”
In brief, the extra orders are equal to a whole yr of orders fulfilled by the biggest AI chip agency on the earth.
That’s not all. Bloomberg has reported that Nvidia will make investments US$2 billion in Elon Musk’s xAI. Nvidia’s funding is a part of the US$20 billion funding spherical that xAI is pursuing. xAI is constructing Colossus, the world’s largest supercomputer, in Tennessee. The mission is predicted to be outfitted with 100,000 GPU models. The mission is planning on boosting that as much as 200,000 models when extra investments are available in.
These just lately introduced investments in AI infrastructure will produce the equal of all AI infrastructure presently out there. OpenAI has no money. Its plan is to increase its lifespan by investments from Nvidia and AMD.
Suspicions of an AI bubble and cross buying and selling have been raised. These suspicions query the feasibility of taking cash acquired from Nvidia to buy Nvidia chips, and of utilizing cash raised from AMD to buy AMD semiconductors.
Concerning these qualms, Huang mentioned that the corporations don’t have any cash, so they have to “increase that cash by, initially, their revenues, which is rising exponentially, fairness or debt.”
“After we invested in OpenAI, early on, my solely remorse is that we didn’t make investments extra,” Huang has mentioned.
In different phrases, they’ve potential however no cash, so I’m investing in them.
Not like OpenAI, which has been actually loud about its funding plans, corporations with cash are securing investments with out making a ruckus.
Google introduced again in July that it might make investments US$85 billion in AI in 2025. This was up US$10 billion from preliminary plans. Microsoft pledged US$80 billion, and Meta US$72 billion.
“What’s happening on the earth versus what occurred in 2000 is simply dramatically completely different. Again then, there have been pets.com, hospitals.com. All the web corporations mixed had been, what, US$30-$40 billion in dimension,” mentioned Huang.
“In the event you take a look at the hyperscalers now, that’s the place the primary tranche of AI infrastructure is constructing. That’s about US$2.5 trillion of enterprise that’s already working at present. That enterprise, that US$2.5 trillion enterprise, and the capex that goes beneath that’s about, name it, US$500 billion,” Huang mentioned.
Know-how paradigm shift recreation
As demand for reminiscence chips exceeded expectations, costs shot up accordingly. Manufacturing must be elevated. But corporations can’t be anticipated to double or triple present manufacturing primarily based on figures in an announcement. In the event that they overreach of their investments whereas demand doesn’t meet expectations, the consequence can be utter catastrophe. This sudden inflow of demand has resulted in one thing of a cheerful dilemma for Korean chip corporations.
Pleased dilemmas, nevertheless, don’t at all times end in comfortable endings. If corporations under-invest, they might be surpassed by Chinese language corporations and different rivals or lose their alternative. In the event that they over-invest, they stands out as the ones left within the rubble when the AI bubble bursts.
Within the AI period, the chip business is not about enterprise cycles. It’s now a paradigm shift recreation. This comfortable dilemma will decide the way forward for South Korea’s semiconductor business.
By Kwon Quickly-woo, editor at 3ProTV
Please direct questions or feedback to [english@hani.co.kr]
