SK Hynix is funding the bonus from record high bandwidth memory profits tied to Nvidia's AI accelerators, the same concentration Fitch just named a top global economy risk.
Samsung engineer "Lee" clocks out on time these days, and his colleagues notice. He uses the rest of his evening to draft applications to SK Hynix, where colleagues say the average bonus is set to clear $476,000 this year. A recent MIT Technology Review feature by Michelle Kim describes the same pattern across Samsung's chip engineering ranks.
The $476,000 figure is not a labor-market abstraction. SK Hynix funds it out of record high-bandwidth memory (HBM) profits. HBM is the stacked memory chip that sits next to Nvidia's AI accelerators. It sells for multiples of ordinary DRAM, and only a handful of suppliers can hit its yields. SK Hynix's Q2 2026 results put it on the tape: 79.3 trillion won (roughly $61 billion at prevailing rates) in revenue, up 257% year on year, with a 76% operating margin, and HBM4 now in mass production (SK Hynix Q2 earnings recap, HBM4 mass production note).
The same profit pool is now too large for the rest of the system to ignore. Bloomberg estimates Samsung and SK Hynix will pay roughly $42 billion in combined bonuses to AI chip workers this year. KED Global frames SK Hynix's bonus rate at 2,964% of base, a ratio rather than a per-head dollar amount. The Bank of Korea has flagged chip-worker bonuses as an inflation risk.
SK Hynix's bonus pool is the visible price of HBM profit concentration. Tom's Hardware puts per-employee bonuses near $447,000 this year, with projections of roughly $900,000 next year. When that much money moves between two named employers, it stops being a retention problem and starts being a supply-chain signal.
HBM is sold on multi-quarter contracts, and it sits in the most concentrated AI capex line item on the market, the high-end Nvidia accelerators. The chip sits between the AI model and the data center, which is why a yield slippage at one supplier shows up at the AI accelerator layer. Losing a senior process engineer at SK Hynix is not like losing a generic factory worker. A bad HBM yield affects every accelerator downstream, which is the reason the bonus is being paid in the first place. Samsung's semiconductor engineers are the upstream variable for that line item, and they are following the bonus.
That concentration is the same one Fitch named this week as one of the biggest risks to the global economy. Fitch's warning came alongside a global tech sell-off, investor anxiety over AI earnings, and an AI Hype Index that points the other way. The same dollar amount shows up in the bonus pool, in Samsung's morale problem, and in Fitch's correction risk.
The deflating hype half of this story is not that AI investment will disappear. It is that the AI capex cycle has become narrow enough to be repriced on a single product line. When Fitch names that as a top global risk, the rating agency is not saying the chips stop selling. It is saying the chips depend on a few companies, a few engineers, and a few contracts. Repricing on a single product means the cycle is no longer diversified across the same number of buyers, suppliers, and engineers that absorbed the last two semiconductor downcycles. The bonus gap is now a forecast for both the labor market and the equity market, and the same dynamic that drew Fitch's attention is the one paying Lee's bonus.
SK Hynix says HBM4 mass production is on track. Samsung has not closed its bonus gap, and its engineers are still clocking out early. The next leg of the AI infrastructure cycle will be decided in those two rooms before it is decided on a tape.