Samsung Locks 70% of Memory Capacity Into 2031 Contracts as Chip Industry Rewires Its Economics
Samsung Electronics has allocated roughly 70% of its memory production capacity to long-term supply agreements stretching as far as 2031. Nvidia, Microsoft and Google are among the named customers. The contracts carry:
- Rolling five-year horizons
- Partial prepayments (reportedly around a quarter of contracted volumes)
- Binding purchase commitments
- Price-floor mechanisms
Samsung is pricing predictability. SK Hynix, which completed about 10 multi-year agreements with deposits and binding terms, is deliberately keeping part of its output uncommitted so it can capture higher prices during shortages. The two largest memory producers are optimizing for different payoffs against the same scarcity.
Nvidia's $279 Billion Receipt
The clearest confirmation arrived in Nvidia's own filings. Supplier commitments jumped from $119 billion to $279 billion in a single quarter, an increase CFO Colette Kress attributed primarily to memory procurement. The payment schedule:
| Period | Amount |
|---|---|
| Remainder of FY2027 | ~$92 billion |
| Following year | ~$87 billion |
| Year after that | ~$88 billion |
This describes a company reserving physical manufacturing capacity years before the related accelerators ship. Nvidia's management continues to identify supply, not demand, as the binding constraint on growth, and still guides for approximately 70% revenue expansion after incorporating those supply limits.
Scarcity Priced in Real Time
Korean trade data makes the rationing visible. Between May and July:
| Metric | Change |
|---|---|
| DRAM export volumes | -13.2% |
| DRAM export value | +18.5% |
| Average unit prices | +36.6% |
Fewer chips, more revenue—textbook scarcity extraction.
Channel reports quote scarce HBM3E modules at four to five times contracted prices. Those figures deserve caution: the HBM "spot market" is thin, fragmented and populated by brokers handling emergency procurement. It should not be mistaken for a deep institutional benchmark. The direction, though, is unmistakable.
Unreserved capacity commands an enormous availability premium, and inventory itself has changed character. Samsung semiconductor inventory rose about 32% in the first half, with work-in-process up roughly 35%. Industry sources attribute part of that build to production being prepared against contracted LTA demand—wafers spoken for before they are stacked—which makes raw inventory figures unreliable as a standalone cycle-warning signal.
The Chinese Clock
CXMT, China's ChangXin Memory Technologies, has reportedly begun small-volume HBM3E production, with Alibaba's T-Head and Cambricon testing samples and a broader 2027 ramp targeted. CXMT also announced LPDDR6 mass production for a Xiaomi device.
| CXMT position | Detail |
|---|---|
| Current standing | Behind the big three in HBM generations |
| Constraints | Yields, advanced packaging access, export controls |
| Mizuho estimate | ~8% global DRAM share by 2027 |
CXMT does not need to match SK Hynix on frontier HBM to damage oligopoly pricing. It only needs to serve Chinese accelerators and mid-tier workloads competently enough to free domestic demand from Korean suppliers and add marginal capacity into adjacent DRAM categories.
If that coincides with the Korean fab expansions scheduled for 2028–29 and any deceleration in hyperscaler AI spending, the familiar memory problem returns through the contract channel: customers renegotiate or forfeit deposits, new capacity arrives anyway, and pricing resets.
The Capacity Optionality Thesis
Current profitability is extreme:
| Company | Q2 result |
|---|---|
| SK Hynix | 76% operating margin |
| Samsung Memory Business | All-time highs in revenue and operating profit |
These are scarcity rents, and LTAs will compress peak margins in exchange for higher troughs and greater visibility. The Samsung structure, by contracting 70% of output, effectively sells customers insurance against future shortages—surrendering enormous call-option value if scarcity worsens further. SK Hynix's refusal to do the same, reportedly even removing price caps from certain agreements, shows management pricing that optionality explicitly.
The scarce asset is not "HBM" as a product category. It is:
- Flexible DRAM wafer capacity
- Plus downstream packaging and test throughput
- Plus the contractual freedom to reprice across memory tiers
TrendForce calculated that DDR5 64GB RDIMM per-wafer revenue overtook HBM in Q1 2026, with HBM profitability also falling below high-end DDR5 as annual HBM contracts lagged the violent run-up in conventional DRAM prices. A wafer start that can be redirected between HBM, high-capacity server DDR5 and future custom products may be worth more than fully committed HBM capacity—precisely because the supplier retains the right to follow the margin.
The investors who understand this will position across the memory scarcity hierarchy rather than crowding a single-product HBM thesis. The investors who do not will discover, probably around 2028, that they owned the wrong bottleneck.
Not investment advice.