The current global chip shortage story does not start in a Taiwanese fabrication plant or a Silicon Valley boardroom. Strangely enough, it starts in Icheon, a sleepy South Korean city an hour southeast of Seoul, with a ringing phone. The second-largest chipmaker in the nation, SK Hynix, has been answering calls from almost every significant tech company in the world for the past few months. People close to the company claim that the offers are unprecedented in the memory chip industry.
These businesses have straightforward demands that are nearly impossible to fulfill. The type of guaranteed memory chips that power ChatGPT, Gemini, and the data centers being built in Texas, Arizona, and the Gulf are what they want. In exchange, they are making an unusual offer. Some are willing to pay for whole production lines. Others have offered to fully finance ASML’s extreme ultraviolet lithography machines, which cost hundreds of millions of dollars each. It’s the kind of suggestion that used to sound ridiculous. It sounds like Tuesday now.
| Company Profile: SK Hynix | Details |
|---|---|
| Headquarters | Icheon, Gyeonggi Province, South Korea |
| Founded | 1983 (originally as Hyundai Electronics) |
| Industry | Semiconductor manufacturing — DRAM, NAND flash, HBM |
| Stock Listing | KOSPI: 000660 |
| Key Customers | NVIDIA, Microsoft, Google, Meta, Apple |
| Flagship Product | HBM3E memory used in AI accelerators |
| Major Equipment Supplier | ASML extreme ultraviolet lithography machines |
| Approximate Workforce | Over 31,000 globally |
| CEO | Kwak Noh-jung |
| 2025 HBM Market Share | Estimated above 50% |
As you read this story, you get the impression that the balance of power has changed. The memory industry was harsh and cyclical for many years. Layoffs, boom, glut, repeat. Engineers in Korea used to make dark jokes about the industry devouring itself. However, the playbook has been altered by the AI boom. Artificial intelligence’s largest bottleneck is now high-bandwidth memory, the particular type that SK Hynix controls. Without it, Nvidia is unable to ship its GPUs. Without it, Microsoft cannot grow Azure. As of right now, only SK Hynix is producing enough of it at a high enough quality to be significant.
Interestingly, the company is exercising caution. It has reportedly been reluctant to accept the offers despite being flush with cash from its best year ever. In actuality, collecting money from a single client entails offering that client chips at a discount for years. It also entails being somewhat ensnared by them. It is said that executives are cautious about that trade. It’s difficult to ignore the irony as you watch this play out. For ten years, tech companies put price pressure on suppliers. They are essentially pleading to overpay now.
The overall picture is even more disorganized outside of the supplier narrative. The biggest grid operator in the US, PJM Interconnection, recently failed to obtain enough electricity for summer 2027 for the first time. AI data centers accounted for 94% of the anticipated increase in demand. Purchasing transformers takes two and a half years. Up until the late 2020s, gas turbines are sold out. The bottleneck starts with the chips. Next is the grid. And water presumably appears somewhere after that.
Eventually, SK Hynix might give in and accept one of these agreements. Another possibility is that the business maintains its position, sells to everyone at market rates, and rides the cycle upward. In any case, something has altered. In actuality, the pace of the AI era is currently being determined by a factory in a small Korean city. People in Redmond and Mountain View should be alarmed by that. Most likely, it already does.
