伯恩斯坦解讀：1420億美元長期訂單，能否托住記憶體週期？

TL;DR

• Bernstein focuses on Micron and SanDisk's new LTAs, believing long-term procurement agreements are improving revenue visibility for the memory sector.
• The two companies disclosed a combined RPO of approximately $142 billion and financial guarantees of about $33 billion, but this is far below model-protected exposure levels.
• LTAs can increase the cost for major customers to walk away, but consumer, China-based clients, and spot demand will still retain cyclical volatility.

In a new report, Bernstein puts long-term procurement agreements in the memory industry back in the spotlight: Micron and SanDisk have signed a new batch of LTAs with purchase commitments, minimum prices, and financial guarantees, aiming to provide a floor for earnings over the next few years.

That floor, however, is not as thick as it seems.

According to public filings and conference call data from Micron and SanDisk, Micron has signed 16 strategic customer agreements. Of these, 14 have cumulative minimum revenue of around $100 billion based on minimum contract prices, with related cash deposits and financial commitments of approximately $22 billion. SanDisk's three current-quarter contracts correspond to about $42 billion in minimum contract revenue, and five agreements total over $11 billion in financial guarantees.

Combined, the roughly $33 billion in guarantees does make it more expensive for major customers to breach contracts. However, Bernstein's model estimates that the revenue scale potentially needing LTA protection over the next 3-5 years is around $5.2 trillion. By this measure, existing guarantees represent only about 0.6%.

This highlights the core divergence the report aims to convey: LTAs are changing the negotiating position between memory companies and their large customers, but they serve more as a buffer for the eventual downturn rather than transforming DRAM and NAND into utilities.

Major Customers Locked into Long-Term Agreements, Guarantees Turn into Real Money

LTAs are not complicated. Customers commit to purchasing volumes over several years in advance, and suppliers provide supply assurance and pricing mechanisms. If a customer doesn't buy, they risk losing prepaid guarantees or incurring other economic costs.

Unlike standard purchase intentions common in the past memory industry, the key difference this time is the integration of financial guarantees into the contract structure.

As of June 2026, Micron had signed 16 strategic customer agreements, including 4 hyperscale customers and 3 mid-sized customers. Cumulative minimum revenue valued at minimum contract prices for 14 agreements is approximately $100 billion, with expected cash deposits and related financial commitments of about $22 billion. This figure includes signed and post-quarter agreements, not necessarily matching balance sheet ending RPO exactly.

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

The two companies have different mechanisms. Micron's guarantees are more back-end weighted. As contracts progress and customers' remaining purchase obligations decrease, the ratio of guarantees to RPO rises, making breach more costly later on. SanDisk's approach is closer to fixed-amount guarantees, expected to remain relatively stable over the contract period.

Micron: 16 Agreements, RPO ~$100B, Guarantees ~$22B; SanDisk: 5 Agreements, RPO ~$42B, Guarantees >$11B.

Bulls value this most. The biggest past issue for the memory industry was the rapid collapse of profitability during price declines. If major customers are willing to pay for guarantees for long-term supply, suppliers can gain clearer revenue baselines, avoiding complete reliance on spot prices for capex and capacity planning.

$33 Billion in Guarantees is Not Thin, but Insufficient for a Deep Downturn

The scale of guarantees and the revenue needing protection are not in the same league.

Bernstein's model estimates that if LTAs need to cover potential revenue over the next 3-5 years, the corresponding protection scale is about $5.2 trillion. This figure is based on the report's model scope. Public company filings do not directly disclose similar industry-wide revenue numbers, requiring differentiation between memory, total semiconductor revenue, and supplier sample revenue.

Even so, the 0.6% guarantee ratio illustrates one thing: LTAs cannot guarantee profitability under all price scenarios.

If spot prices only decline moderately, it's not worthwhile for customers to walk away. Losing guarantees, damaging supplier relationships, and potentially missing out on scarce future capacity are costs sufficient for customers to honor contracts. The demand for stable supply from AI server, cloud, and data center customers is also stronger than typical consumer electronics clients.

However, if prices fall deeply enough, customers will still do the math. As long as remaining purchase volumes are large and the spot price falls sufficiently below the contract floor, customers might find it cheaper to buy from the market even if they lose their guarantees.

The back-end weighted mechanism can help with this. As contracts progress, remaining RPO declines, and the ratio of guarantees to remaining obligations rises, increasing the cost of abandoning the contract. Protection may be stronger later in the contract lifecycle, when the memory cycle often needs protection most.

It is still not unconditional insurance. The protective power of an LTA depends on three factors: where the spot price lands, the customer's remaining purchase obligations, and the remaining guarantee balance.

RPO declines over time, Guarantee/RPO ratio rises; if spot ASP falls too far below the contract floor, customers may still choose to walk away.

This is also the core of the bull-bear divergence. Bulls see that memory companies have finally secured real, long-term financial commitments from customers. Bears worry that the scale of these commitments remains insufficient to protect peak earnings, and customers will act rationally based on cost once a downturn is deep enough.

Not All Memory Demand is Willing to Be Locked into LTAs

LTAs have a practical limit: not all customers are suitable for long-term contracts.

US cloud providers are the most ideal candidates. They have large demand, strong credit, high sensitivity to stable supply for AI infrastructure, and greater incentive to lock in supply via long-term agreements. Micron has essentially completed negotiations with US CSPs and is continuing with Chinese CSPs, enterprises, and some other customers.

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

Chinese customers may not necessarily become stable LTA buyers either. On one hand, Chinese cloud vendors and end customers may prefer domestic suppliers. On the other hand, expanding domestic DRAM and NAND supply will add uncertainty to long-term purchase commitments.

Bernstein estimates that 30%-50% of the total DRAM and NAND end market may be difficult to cover with LTAs. Even if leading suppliers lock in US major customers, a significant portion of the market will continue to operate based on spot prices, short-term orders, and cyclical expectations.

A breakdown of the DRAM/NAND end market shows demand from Chinese CSPs, enterprise servers, consumers/PCs, and smartphones beyond US CSPs, with 30%-50% of the market potentially hard to cover with LTAs.

As long as a sufficient portion of demand remains within the spot or short-term contract system, price signals will not disappear. With price signals present, supplier expansion, customer inventory reduction, and channel order cuts will continue to amplify cyclical volatility.

AI Demand Supports Valuations, but Peak Earnings Can't Be Extrapolated Directly

The market's willingness to assign higher valuations to memory companies is partly backed by AI demand changing the bottom of this cycle.

On the DRAM side, HBM demand remains strong. Bernstein's Asia team predicts HBM prices might rise 2-2.5 times in 2027 compared to 2026. While conventional DRAM commercial prices have already risen significantly, they may remain high for the next 12 months. Although HBM is more stable than regular memory, it shares some capacity with conventional DRAM, and capacity allocation will affect other product lines.

On the NAND side, AI inference and longer context windows bring new demand expectations. Early AI training mainly consumed HBM and DRAM, but as inference, Agentic AI, and long-context applications increase, storage demand may rise further. It's worth noting that statements regarding capacities related to Vera Rubin should not simply be written as "GPU's NAND capacity"; NVIDIA's official page discloses 20.7TB of HBM4 GPU memory.

In this environment, the value of LTAs is more about locking in a portion of high-cycle revenue. If AI demand remains strong, suppliers can secure some major customer procurement through long-term contracts. If prices fall, guarantees and floor prices can help delay profit decline.

SanDisk's stress test points to similar conclusions. Bernstein's model shows that under severe assumptions, LTAs still allow FY29-FY30 EPS to be higher in most penetration scenarios compared to no-LTA situations, with protection especially stronger later on. However, the same stress tests also show that peak earnings cannot be extrapolated easily. Under lower operating margin scenarios, EPS could be significantly below current running levels.

SanDisk FY29-FY30 EPS sensitivity table shows a wide range under different ASP and LTA penetration scenarios; LTAs improve the downturn scenario but cannot lock in peak earnings.

The most valuable takeaway from this report is not "the memory cycle is over" 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 supply certainty in the AI era. For memory companies, this will improve revenue visibility over the next few years, making it easier for capital markets to believe earnings floors are higher than in the past.

The limits are equally clear. The $33 billion in guarantees only provides a partial buffer. Consumer customers, Chinese clients, and some transactional demand will not fully enter long-term agreements. Bernstein also estimates that China's DRAM share could rise from roughly 8% to 16% over the next few years, and NAND could face stronger supply pressure after 2028.

What LTAs truly need to prove is not whether they can be signed during an upcycle, but whether customers will honor contracts, whether guarantees will be painful enough, and whether suppliers will maintain capacity discipline during the next downturn. Until these questions are answered, LTAs remain a new buffer for the memory industry, not the button to end its cycle.