SK Hynix just paid its employees a bonus worth 2,964% of their monthly base salary. The company's profit-sharing agreement, struck with its labor union in September 2025, allocates 10% of annual operating profit directly to workers with no cap. With analyst forecasts projecting $169 billion in operating profit for 2026, average employee payouts could reach $477,000 this year. South Korean matchmaking services have reported increased interest in chip industry employees following bonus announcements. Young South Koreans joke that the best outfit for a blind date is an SK Hynix uniform. Meanwhile, a significant portion of enterprise AI projects—the ones those chips are built for—are getting abandoned after proof of concept.
The surface story is appealing in its simplicity: the AI boom creates winners. Chip workers get rich because they make an essential physical thing. Software companies waste money because they chase hype. This framing lets everyone feel comfortable. The hardware people earned it. The software people didn't plan properly. Move along.
ERROR: VALUE_CAPTURED_AT_WRONG_LAYER
Here's the structural problem. The AI economy is concentrating extraordinary wealth at the physical substrate layer—the people who fabricate memory chips—while the organizations that actually buy those chips and try to use them are lighting money on fire at a historically unprecedented rate. Gartner found that by the end of 2025, at least 50% of generative AI projects had been abandoned after proof of concept due to poor data quality, inadequate risk controls, escalating costs, or unclear business value. Their forecast says 60% of AI projects unsupported by AI-ready data will be abandoned through 2026. A RAND meta-analysis across enterprise AI initiatives found that a majority delivered no business value at all. The failure is not at the model layer. It's upstream: data readiness, organizational maturity, use-case drift. Nobody who's been accountable for a customer ID across three enterprise systems in their life is surprised by this.
What makes this genuinely perverse is the incentive loop. SK Hynix's bonuses are funded by demand from companies building AI data centers. Those data centers exist because executives committed budgets to AI deployments. Most of those deployments will fail—not because the chips are bad, but because the organizations buying them have fragmented data, no governance, and success metrics they defined after the demo, not before it. The chips work beautifully. The systems that consume them are broken. But the demand signal doesn't care. Nvidia needs HBM chips, SK Hynix makes them, workers get substantial bonuses, and matchmaking services update their profiles. The value chain is functioning perfectly as a wealth-transfer mechanism from enterprise AI budgets to South Korean semiconductor workers. It is not, however, functioning as an intelligence-delivery mechanism for anyone.
This is what happens when an industry prices inputs by speculative demand rather than realized output. The chip factories don't need AI to work for their customers. They need AI to be purchased by their customers. As long as CIOs keep signing datacenter contracts based on board pressure and competitive fear, the hardware layer keeps printing money regardless of what happens one layer up. The workers at SK Hynix are not doing anything wrong—they're doing something extraordinarily difficult and valuable. But the economic signal they're responding to is being generated by an enterprise software market where, as one analysis put it, the underlying problem is not technology but leadership and process gaps. That's a polite way of saying: the people writing the checks don't know what they're buying.
The fix is unglamorous and structurally obvious: tie AI infrastructure procurement to demonstrated data readiness, not to strategic intent. No company should be signing a multi-year datacenter or compute contract without a completed data audit showing governed, labeled, pipeline-ready data for at least one production use case. Make the hardware purchase contingent on the homework being done. This would slow procurement cycles, irritate every cloud and chip vendor on earth, and immediately cut the failure rate. It would also, incidentally, make the bonuses that do get paid actually sustainable—because they'd be funded by customers who got value, not customers running on fumes and a slide deck.