TL;DR

• Bernstein focuses on Micron and SanDisk's new LTA, believing that long-term purchase agreements are improving the visibility of memory revenues.
• Both companies disclosed a total RPO of about $142 billion and financial guarantees of about $33 billion, but this is still far below the model protection scale.
• LTA can increase the cost for major customers to breach contracts, but consumer, Chinese customers, and spot demand will still retain cyclical fluctuations.

Bernstein has brought the long-term purchase agreements in the memory industry back to the forefront in its latest report: Micron and SanDisk have signed a batch of new LTAs that include purchase commitments, minimum prices, and financial guarantees, attempting to establish a revenue floor for the coming years.

This floor does not appear to be as thick as it seems.

According to documents and conference call disclosures from Micron and SanDisk, Micron has signed 16 strategic customer agreements, with 14 of these agreements calculating a cumulative minimum revenue of approximately $100 billion based on minimum contract prices, along with cash deposits and financial commitments of about $22 billion. SanDisk's three contracts for the quarter correspond to about $42 billion in minimum contract revenue, with five agreements collectively having financial guarantees exceeding $11 billion.

The total guarantees from both companies of about $33 billion indeed make breaching contracts more expensive for major customers. However, Bernstein's model estimates that the revenue scale that may require LTA protection over the next 3-5 years is about $5.2 trillion. Based on this report's calculations, the existing guarantees only account for about 0.6%.

This is precisely the divergence that the report aims to express: LTA is changing the negotiation position of memory companies and major customers, but it is more like providing a cushion during a down cycle rather than turning DRAM and NAND into public utilities.

Major customers locked into long-term agreements; guarantees start turning into real money

LTA is not complicated. Customers commit in advance to future purchase volumes, and suppliers provide supply guarantees and pricing mechanisms. If customers do not buy, they may lose their pre-paid guarantees or incur other economic costs.

This time, compared to past common purchase intentions in the memory industry, the focus is on financial guarantees being integrated into the contract structure.

As of June 2026, Micron has signed 16 strategic customer agreements, including four ultra-large customers and three medium-sized customers. The cumulative minimum revenue of 14 agreements, calculated based on minimum contract prices, is about $100 billion, with expected cash deposits and related financial commitments of about $22 billion. This figure includes signed agreements and agreements signed after the quarter, and does not exactly equal the ending RPO on the balance sheet.

SanDisk disclosed that as of April 3, 2026, its RPO was $41.6 billion. The company also mentioned on the conference call that three contracts for the quarter provide about $42 billion in minimum contract revenue, and that five agreements collectively have over $11 billion in financial guarantees, covering more than one-third of bit supply for FY27.

The mechanisms of the two companies are different. Micron’s guarantees place more emphasis on back-end weighting. As the contracts progress and the customers' remaining purchase obligations decrease, the ratio of guarantees relative to RPO will rise, making the cost of breaching contracts heavier in later stages. SanDisk, on the other hand, is closer to fixed amount guarantees, with the guarantee amounts expected to remain relatively stable during the contract period.

Micron has 16 agreements, with an RPO of about $100 billion and guarantees of about $22 billion; SanDisk has 5 agreements, with an RPO of about $42 billion and guarantees exceeding $11 billion.

The bullish side sees this as the most important point. The biggest problem in the memory industry in the past was the rapid collapse of profits when prices fell. If major customers are willing to pay guarantees for long-term supply, suppliers can at least secure a clearer revenue baseline, and capital expenditure and capacity planning do not have to be entirely dictated by spot prices.

$33 billion in guarantees is substantial but cannot withstand deep downturns

The scale of guarantees and the revenue scale that need protection are not on the same level.

Bernstein calculates that if LTAs are to cover potential revenue over the next 3-5 years, the corresponding protection scale is approximately $5.2 trillion. This figure is part of the report's model calculations; publicly available company documents do not directly disclose such an industry-wide revenue scale, and it is necessary to distinguish between memory, total semiconductor revenue, and supplier sample revenue.

Even so, the 0.6% guarantee ratio still indicates one thing: LTAs cannot guarantee profits under all price scenarios.

If spot prices merely decline moderately, customers breaching contracts might not make economic sense. Losing guarantees, damaging supply relationships, and potentially missing out on scarce capacity in the future would be sufficient costs for customers to continue fulfilling contracts. The demand for stable supply from AI servers, cloud vendors, and data center clients is also stronger compared to ordinary consumer electronics customers.

However, when prices drop sufficiently low, customers will still consider the economics. As long as significant remaining purchase volumes exist and spot prices are low enough relative to the contract floor price, customers may find it cheaper to buy in the market, even if they lose their guarantees.

The back-end weighting mechanism can alleviate this issue. As the contract progresses, the remaining RPO decreases, increasing the relative proportion of guarantees to remaining obligations, hence increasing the costs for customers to breach contracts. The protective force may be stronger in the later stages of the contract, and memory cycles usually require more protection in the later phases as well.

It is still not unconditional insurance. The protective force of LTA depends on three numbers: where spot prices fall, how much purchase obligation remains for customers, and how much guarantee balance is left.

RPO declines over time, and the guarantee/RPO ratio increases; if the spot ASP drops too low relative to the contract floor price, customers might still choose to breach the contract.

This is also the core of the bullish-bearish divergence. Bulls see that memory companies have finally obtained long-term commitments from customers in real monetary terms. Bears are worried that the scale of these commitments is still insufficient to protect peak profits, and that once the downturn cycle is deep enough, customers will act based on costs.

Not all memory demands are willing to be locked into LTAs

LTA also has a practical limit: not all customers are suitable for signing long-term agreements.

U.S. cloud vendors are the most ideal target. They have large demands, strong credit, sensitivity to the stability of supply for AI infrastructure, and are also more motivated to lock in supply through long-term agreements. Micron has essentially completed negotiations with U.S. CSPs and is still advancing negotiations with Chinese CSPs, enterprises, and some other customers.

Consumer business is different. SanDisk's CFO has stated that consumer business is "more transactional," and LTA "is not applicable." Mobile phones, PCs, and consumer storage channels are more accustomed to purchasing based on price and inventory cycles. Once prices fall, customers naturally wish to maintain flexibility rather than being locked into multi-year floor prices.

Chinese customers may also not become stable buyers of LTAs. On one hand, Chinese cloud vendors and end customers may prefer domestic suppliers. On the other hand, the expansion of domestic DRAM and NAND supply may also increase the uncertainty of long-term purchase commitments.

Bernstein estimates that approximately 30%-50% of the overall DRAM and NAND end market may be difficult to be covered by LTAs. Even if leading suppliers lock in large U.S. customers, there will still be a considerable portion of the market operating based on spot prices, short-term orders, and cyclical expectations.

Splitting the DRAM/NAND end market shows that, besides U.S. CSPs, there are also Chinese CSPs, enterprise servers, consumer/PC, smartphones, etc. Approximately 30%-50% of the market may struggle to be covered by LTAs.

As long as a sufficiently large proportion of demand remains in the spot or short-order system, price signals will not disappear. As long as price signals exist, suppliers will expand production, customers will deplete inventory, and channels will cut orders, still amplifying cyclical fluctuations.

AI demand supports valuations, but peak profits cannot be directly extrapolated

The market is willing to give memory companies higher valuations, partly because AI demand has changed the bottom shape of this cycle.

On the DRAM side, HBM demand remains strong. Bernstein's Asian team predicts that HBM prices may rise 2-2.5 times compared to 2026 by 2027. Regular DRAM commercial prices have previously increased significantly and are expected to remain high over the next 12 months. Although HBM is more stable than ordinary memory, it shares some production capacity with regular DRAM, and capacity allocation can affect other product lines.

On the NAND side, AI reasoning and longer context windows have also brought new demand expectations. Early AI training mainly consumed HBM and DRAM, but with the increase in reasoning, agentic AI, and long-context applications, storage demand may continue to rise. It should be noted that descriptions regarding Vera Rubin's related capacity should not be simply written as "NAND capacity for GPUs," as NVIDIA's official page discloses 20.7TB HBM4 GPU memory.

In this environment, the value of LTA is more about solidifying a portion of high-demand revenue. If AI demand continues strongly, suppliers can lock in certain large customer purchases through long-term agreements. If prices fall back, guarantees and floor prices can delay profit declines.

SanDisk's stress tests also point to similar conclusions. Bernstein's model shows that under stricter assumptions, LTAs can keep FY29-FY30 EPS higher than scenarios without LTAs in most penetration scenarios, with protection being especially stronger in later stages. However, the same set of stress tests also indicates that peak profits cannot be simply extrapolated. In scenarios with lower operating margins, EPS may be significantly lower than current operational levels.

SanDisk FY29-FY30 EPS sensitivity table shows that EPS spans widely under different ASP and LTA penetration rates, improving the down scenario but unable to lock in peak profits.

The most worthwhile judgment to retain from this report is not "the memory cycle is ending," but rather "the cycle downturn may be softened."

Micron and SanDisk securing long-term agreements and financial guarantees indicates that major customers are willing to pay for certainty in supply during the AI era. For memory companies, this will enhance the visibility of revenues over the coming years and make capital markets more inclined to believe that the profit baseline is higher than in the past.

The limitations are equally clear. The $33 billion in guarantees can only provide partial cushioning, and consumer, Chinese customers, and certain transactional demands will not all be entering long-term agreements. Bernstein also estimates that the Chinese DRAM share may increase from about 8% to 16% over the next few years, and NAND may face stronger supply pressure after 2028.

The true test for LTA is not whether it can be signed during an up cycle, but whether customers will fulfill contracts during the next down cycle, whether the guarantees will be painful enough, and whether suppliers will continue to maintain capacity discipline. Until these questions are resolved, it remains a new buffer for the memory industry, not a termination button for the cycle.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。