Micron, SK Hynix and Samsung Electronics are trying to move the memory business away from the old pattern of violent upswings followed by equally sharp declines. An ABMedia analysis argues that this cycle is being reshaped by new contract structures, equity stakes in AI companies and capital-market moves, rather than by demand alone.
Citing P Equity Research and AI infrastructure analyst Muhammad Zuhair, the report lays out four changes across suppliers, buyers and capital markets that, in its view, help explain why the memory market is not following the same script as before.
Micron’s results are being treated as proof that demand is real
The starting point is Micron’s Q3 FY2026 earnings. The company posted $41.46 billion in quarterly revenue, up 74% from the prior quarter and 346% from a year earlier. GAAP gross margin came in at 84.9%.
The article says the jump from an $1.1 billion loss in FQ1-24 to $33.3 billion in FQ3-26 has gone a long way toward ending doubts about the strength of this DRAM upcycle.
JPMorgan’s supply-demand model, as cited in the piece, projects DRAM bit demand growth of 32.5% in 2026 and 34.4% in 2027. Over the same period, bit shipment growth is expected at just 22.9% and 21.7%. That would leave the supply-demand gap at roughly 13 percentage points in 2027 before the market reaches its first point of supply overtaking demand in 2028.
SCAs are replacing older long-term agreements
The report describes Strategic Customer Agreements, or SCAs, as the most important structural change in this cycle. Micron has reportedly signed 16 of them. Based on minimum committed shipment volumes and minimum contract prices, those agreements represent more than $100 billion in remaining performance obligations.
ABMedia contrasts SCAs with traditional LTAs, or long-term agreements. Under the older setup, suppliers committed to ship when customers asked for product, but buyers were not forced to take volumes. If spot prices fell, customers could move away and suppliers absorbed the damage.
SCAs are presented as a direct answer to that problem. The article breaks the design into three parts:
- pricing bands with a floor and a ceiling, intended to preserve margins across stronger and weaker market conditions;
- financial commitments, including cash prepayments or letters of credit held through third-party accounts, giving suppliers capital they can still draw on if a customer exits early;
- longer duration, usually three to five years, with Micron’s larger SCAs often extending from 2026 to 2030.
On coverage, the 16 signed SCAs account for about 20% of Micron’s DRAM shipments and 33% of its NAND shipments, according to the article. The company’s goal is to push that structure to cover more than 50% of total revenue.
The analysis says that marks a first for the memory industry: suppliers are no longer carrying the full burden of capacity expansion and demand volatility on their own.
Memory suppliers are buying into AI labs as well as selling to them
The second change is ownership alignment with major AI customers. The report says Micron, SK Hynix and Samsung Electronics have all taken sizeable stakes in top AI labs, including participation in Anthropic’s Series H round, in exchange for long-term supply commitments covering HBM, DRAM and SSD products.
At the same time, Anthropic and OpenAI are both working on in-house ASICs. As HBM demand rises with each new chip generation, the article argues that an equity stake changes the nature of the supplier-customer relationship. It is no longer just procurement. A supplier that also owns part of the customer may gain earlier visibility into the chip roadmap and tie its financial outcome more closely to the customer’s commercial success.
Memory is taking a bigger share of AI infrastructure spending
The third change is the growing weight of memory in AI capital spending. The article says memory accounted for around 7.7% of total capital expenditure at hyperscalers in 2023 and 2024. That share is put at 13.5% for 2025, 30.0% for 2026 and 36.2% for 2027.
Goldman Sachs estimates global AI infrastructure capex will total $7.6 trillion from 2026 through 2031. Of that, 2026 alone would account for $765 billion, with chips making up 64.5%.
At the system level, Morgan Stanley estimates the memory bill of materials for Nvidia’s VR200 NVL72 rack is 435% higher than for the prior-generation GB300, rising from $374,000 to $2 million. Memory’s share of total rack BOM would climb from 9% to 26%. The bank expects that upward move to continue after the Rubin architecture.
Mizuho extends that point to custom chips. Meta’s MTIA 500, also referred to as Astrid, is described as carrying 512GB of HBM4 with an estimated ASP of $16,000. Google’s TPUv9p, or Humifish, is listed with 432GB of HBM4E and an estimated ASP of $13,000.
Even if hyperscalers shift toward in-house silicon to avoid Nvidia’s gross margins of more than 70%, the report says they still face the same memory constraint: there is no low-cost substitute for HBM.
The article attributes HBM’s high cost to physical limits in manufacturing. The process is more complex, yield is lower, and it consumes more wafer area than conventional DDR. Bernstein estimates that, at current pricing, allocating capacity to conventional DRAM produces more than twice the wafer revenue of HBM and nearly three times the gross profit.
The risks have not gone away
The article does not frame this as a risk-free reset. It lists several points that could still pull the industry back toward a more familiar cycle.
SCAs only work if memory stays scarce
The first is that SCA prepayments depend on one condition holding: memory must remain in persistent short supply. The report says those customer prepayments are not permanent guarantees. They are rolling commitments reviewed each year. If supply-demand conditions normalize, or if spot and contract pricing move back toward balance, customers may lose the incentive to keep paying upfront. At that point, growth in SCA signing could stall.
In other words, the article sees the moat around SCAs as time-limited rather than permanent.
HBM pricing negotiations are still unresolved
A second risk lies in HBM pricing. The piece says average conventional DRAM pricing rose 4.5 times between Q3 2025 and Q2 2026, while HBM pricing was largely locked in under annual contracts and barely moved. That creates a margin gap and leaves suppliers with a difficult wafer-allocation decision.
According to the report, suppliers are seeking price increases of 2x to 2.5x for certain HBM products to close that gap. If negotiations fail, or land below target, margins on a mixed product portfolio could deteriorate.
2028 could mark a supply-demand turning point
The third risk is the supply-demand convergence projected for 2028. JPMorgan’s model shows global DRAM bit supply growth overtaking demand growth for the first time then. As new fabs move into volume production, total capacity would expand to about double current installed capacity.
The article adds that the balance point could arrive sooner if AI infrastructure capex runs into spending limits or if memory-efficiency technologies make real progress. One example it gives is Qualcomm’s HBC architecture, described as a near-memory computing approach based on vertically integrated LPDDR stacking and claimed to improve bandwidth-per-watt by 6x relative to HBM.
Chinese suppliers could add pressure in commodity segments
The final risk comes from China’s memory producers. The report says CXMT and YMTC remain focused on commodity DRAM and standard NAND, and do not pose a near-term challenge in HBM. But if their shipment volumes build quickly in general-purpose DRAM and fill the gaps left in consumer and mobile demand as the major suppliers prioritize HBM, the competitive picture could begin to change.
The central point of the analysis is not simply that memory demand is strong. It is that contract terms, customer alignment and AI infrastructure spending have all shifted at the same time. Whether that is enough to break the old cycle for good will depend on how long SCAs retain their force, whether HBM pricing can catch up, what happens to supply in 2028, and how far Chinese producers advance in commodity memory.