South Korea's two biggest chipmakers make almost all of the world's high-bandwidth memory, the special stacked memory chips that sit next to Nvidia's AI processors and feed them data fast enough to keep up. Right now, one of those companies is paying its workers so much more than the other that entire teams are quitting to join the rival down the road.
SK Hynix is handing out an average bonus of about $476,000 per employee this year, mostly in cash, funded by a formula that pays out 10 percent of the company's operating profit to staff. Samsung struck a similar deal with its union, but Samsung splits its bonus pool by division rather than company-wide. Workers in its memory division, which also makes HBM chips, are getting around $400,000. Workers in its foundry division, which manufactures chips designed by other companies and has been losing money, are getting roughly $135,000.
That gap is why engineers describe entire 30-person teams applying to SK Hynix at once, sometimes with their own managers pointing the way out. A Samsung union survey found that 81.5 percent of foundry employees and close to half of all semiconductor staff want to leave within two years. Samsung has already gone to court over it, winning an 18-month injunction against two former employees who moved to SK Hynix, arguing that chip technology counts as a national asset worth protecting.
The roots of this go back to a decision Samsung made in 2019, when it shrank its high-bandwidth memory team on the bet that the market would stay small. SK Hynix bet the other way. When ChatGPT and the AI boom that followed sent demand for these chips soaring, SK Hynix was positioned to capture it, and Samsung was not. By one industry tracker, Samsung's share of the HBM market fell to 17 percent this year from 41 percent a year earlier, while SK Hynix holds well over half the market.
The stakes go beyond bragging rights between two Korean companies. Samsung is the only company in the world that makes both memory chips and the advanced logic chips that sit underneath them in the newest HBM designs, a combination that could matter a lot for the next generation of AI chips. If Samsung keeps losing the engineers who make that combination work, it risks falling permanently behind, leaving SK Hynix and Nvidia's other suppliers with even more pricing power.
For any business outside the chip industry, the practical takeaway is simpler. Memory chips are getting more expensive because of the same AI demand driving this talent war, and that cost eventually lands in the price of servers, laptops, and cloud computing bills. South Korea's own industry group estimates the country will be short about 54,000 semiconductor workers by 2031. When the companies making a critical input for nearly all computing hardware are this short-staffed and this willing to poach from each other, expect the memory shortage, and its price tag, to stick around longer than most people assume.