Micron offered its Taiwan employees a NT$1 million cash bonus, about US$31,650, in September 2026. The unions rejected it. They demanded a permanent profit-sharing plan allocating 15% of operating profits directly to workers, and a one-off payment equivalent to 83 months of salary. The company had just reported $28.24 billion in net income for a single quarter.
That standoff captures the strange moment the memory industry finds itself in. Last week, this column examined HBM pricing and CXMT's commodity base. This week, the question is whether the supply discipline those stories assumed is now locked into contracts that a demand shock would break.
The mechanism is the long-term agreement. Samsung, SK hynix, and Micron have all signed multi-year supply deals—typically five years—with major customers. Micron disclosed 16 "Strategic Customer Agreements" covering about 20% of its DRAM shipments and a third of NAND. The company has collected over $22 billion in cash and financial commitments, including roughly $18 billion in cash deposits. Samsung projected that multi-year contracts will account for 60–70% of its total memory production capacity once negotiations conclude.
The mechanism is the long-term agreement. Samsung, SK hynix, and Micron have all signed multi-year supply deals—typically five years—with major customers.
These agreements work well when prices rise. SK hynix has eliminated price ceilings from its long-term contracts, allowing spot market increases to flow through. Micron has introduced price bands with upper and lower limits, and says the lower limit alone secures gross margins exceeding any past cycle peak. Samsung has received about a quarter of the advance payments included in its major contracts. Customers are contractually obligated to pay even if they don't take delivery—"take-or-pay" clauses that Micron's executive vice president called "non-cancellable."
But the contracts are starting to cut the other way. SK hynix's DRAM market share fell from 28.8% to 24.9% in the second quarter, even as its DRAM revenue rose 37.9% to $38.59 billion. Samsung and Micron grew faster—63.4% and 65.5% respectively—because they had less of their capacity locked into HBM long-term agreements and more exposure to soaring conventional DRAM prices. SK hynix, with 58% of global HBM revenue in the first quarter, found its strength became a constraint. The more HBM it makes, the more wafer capacity it consumes, and the less conventional DRAM it can sell into a rising market.
The opportunity cost is structural. Each generation of HBM requires more wafer area and more complex processes. Every unit of HBM capacity SK hynix adds displaces more conventional DRAM supply. When server DRAM prices are rising alongside HBM, the calculus of expanding HBM capacity becomes more expensive in ways that don't show up in HBM contract margins.
Now consider what could break the demand side. On September 10, 2026, DeepSeek released V4.1-Flash, which needs one-quarter as much HBM and one-eighth as much SSD capacity for its KV cache as the previous generation. That's a 75% reduction in HBM and 87.5% in SSD for one specific but crucial workload. DeepSeek didn't claim the entire data center uses less memory—the comparison covers KV cache only, not model weights or training. But inference is where sustained token growth turns architectural improvements into demand shocks.
The bull case is that cheaper tokens stimulate more usage. Micron's Cloud Memory revenue reached $13.77 billion with an 83% gross margin; Core Data Center revenue was $11.52 billion at 87%. Sandisk's data-center revenue rose 103% sequentially to $2.98 billion, with pricing generating roughly two-thirds of the increase. Those margins embed scarcity. If model designers broadly cut bytes required per token, demand can grow rapidly without matching today's memory-intensity forecasts.
Then there's the financing question. About 50% of the estimated $2.9 trillion in AI capital expenditure planned for 2025 to 2028 is expected to be funded externally, according to Kim Jeong-hun of Korea Ratings. That makes memory demand sensitive not just to AI usage but to interest rates, capital market liquidity, and investor risk appetite. Long-term agreements improve visibility, Kim said, "but they do not eliminate customer credit risk or financing risk."
The history of these contracts is not encouraging. During the 2017–2018 memory supercycle, Big Tech rushed into long-term agreements. When inventory adjustments began in the second half of 2018, contract fulfillment was delayed, and the industry faced a downturn in 2019. Earlier, AMD signed a take-or-pay wafer agreement with GlobalFoundries, then paid $320 million in 2012 to amend and terminate it when PC demand slumped. More recently, Microchip Technology's "Preferred Supply Program" required customer commitments during the 2021 shortage. By November 2025, CEO Steve Sanghi said, "We are not going to force customers to buy anything they don't need."
The legal enforceability of current contracts is limited in practice. It would be difficult for Samsung or SK hynix to charge liquidated damages to key Big Tech customers over an AI memory oversupply. Renegotiations would more likely involve sharing losses, delaying delivery schedules, or adjusting unit prices. Straining relationships with customers who control future supply opportunities is a cost few suppliers would willingly bear.
This is why the labor unrest at Micron matters beyond the picket line. The company's cumulative net income over four quarters exceeded $50.47 billion. Direct manufacturing workers were offered bonuses equivalent to 35 to 68 months of basic salary. They rejected it because the boom's gains are being distributed through discretionary bonuses rather than structural profit-sharing. SK hynix reached a settlement last September to allocate 10% of annual operating profit directly to employees for the next decade. Samsung workers held a demonstration attended by over 30,000 union members. The social contract inside these companies is being renegotiated in real time, even as the commercial contracts are being signed.
The contracts themselves are a bet that AI demand grows linearly for five years—an eternity in semiconductors. If it doesn't, the same provisions that provide revenue visibility become a mechanism for delaying the inevitable. Customers consume excess inventory before placing new orders, revenue is delayed rather than accelerated, and suppliers face the choice between enforcing contracts and preserving relationships.
The PC makers evaluating CXMT memory are watching this closely. If the incumbents' long-term agreements keep them locked into HBM while conventional DRAM remains tight, Chinese suppliers gain a window. CXMT's DDR5-8000 and LPDDR5X-10667 modules are JEDEC-compliant, built on a 16nm node three years behind the leaders, but available. The shortage is doing what export controls and years of development could not: pushing global OEMs to test Chinese DRAM.
Memory's long-term contracts were designed to smooth the cycle. Instead, they may be concentrating it—shifting risk from the spot market to the balance sheet, from quarterly negotiations to five-year commitments that no one can credibly enforce if the AI capex curve bends. The shield and the burden are the same instrument. It depends entirely on which direction prices move next.