Bernstein’s interpretation: $142B in long-term orders—can it hold up the memory cycle?

TL;DR

• Bernstein focuses on Micron and SanDisk’s new LTA, saying long-term purchase agreements are improving visibility into memory revenue.
• The two companies disclosed total RPO of about $142 billion and financial guarantees of about $33 billion, but this is far below the model’s protection baseline.
• LTA can raise the contract-termination cost for large customers, but consumer, China customer, and spot demand will still retain cyclical volatility.

In its latest report, Bernstein puts long-term purchase agreements back in the spotlight for the memory industry: Micron and SanDisk have already signed a batch of new LTAs with purchase commitments, minimum prices, and financial guarantees, aiming to put a floor under profitability over the coming years.

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

Based on publicly filed documents from Micron and SanDisk and conference-call commentary, Micron has signed 16 strategic customer agreements, of which 14 have cumulative minimum revenue of about $100 billion based on the minimum contract price. Related cash deposits and financial commitments are about $22 billion. SanDisk’s three-quarter contracts correspond to about $42 billion in minimum contract revenue, while five agreements total financial guarantees of more than $11 billion.

Together, the two companies’ guarantees of about $33 billion do make it more expensive for large customers to walk away. But Bernstein’s model scales the revenue that may need LTA protection over the next 3-5 years to about $5.2 trillion. Under the report’s framing, the current guarantees are only about 0.6%.

This is the disagreement the report is trying to highlight: LTA is changing the bargaining position of memory companies and large customers, but it’s more like adding padding to soften a downturn cycle—not converting DRAM and NAND into utilities.

Large customers are locked into long-term deals; guarantees start turning into real money

LTA isn’t complicated. Customers commit to purchase volumes in advance for the coming years, and suppliers provide supply assurance and pricing mechanisms. If a customer doesn’t buy, it could lose the paid-up guarantees or incur other economic costs.

This time, unlike past common procurement “intent” in the memory industry, the key change is that financial guarantees enter the contract structure.

As of June 2026, Micron has signed 16 strategic customer agreements, including 4 mega customers and 3 mid-sized customers. The 14 agreements have cumulative minimum revenue of about $100 billion based on minimum contract prices, and the company expects to receive cash deposits and related financial commitments of about $22 billion. This definition includes both signed agreements and agreements signed after the quarter; it’s not identical to RPO at the end of the balance-sheet period.

SanDisk disclosed that as of April 3, 2026, RPO was $41.6 billion. The earnings call also mentioned that three quarter contracts provide about $42 billion in minimum contract revenue, five agreements total financial guarantees of more than $11 billion, and they cover over one-third of FY27 bit supply.

The mechanisms differ between the two companies. Micron’s guarantees emphasize weighting on the back end. As the contract progresses and the customer’s remaining purchase obligation declines, the guarantee as a proportion of RPO rises—making termination costs heavier later in the term. SanDisk is closer to fixed-amount guarantees, with the guarantee amount expected to remain relatively stable over the contract period.

Micron’s 16 agreements: RPO about $100 billion; guarantees about $22 billion; SanDisk’s 5 agreements: RPO about $42 billion; guarantees over $11 billion.

The bull case values this most. The biggest problem for the memory industry in the past was that when prices fell, profits collapsed too quickly. If large customers are willing to pay guarantees for long-term supply, suppliers can at least secure a clearer revenue floor, and capital expenditures and capacity planning don’t have to be driven entirely by spot pricing.

$33 billion in guarantees isn’t small, but it still can’t cover a deep downturn

The size of the guarantees and the size of the revenue that needs protection are not on the same scale.

Using its model, Bernstein estimates that if LTA needs to cover potential revenue over the next 3-5 years, the corresponding protection baseline is about $5.2 trillion. This figure is based on the report’s model framing; public company filings don’t directly disclose a similar all-industry revenue baseline, and it also needs to distinguish between total memory and semiconductor revenue versus supplier-sample revenue.

Even so, a 0.6% guarantee ratio still says one thing clearly: LTA can’t cover profitability across all price scenarios.

If spot prices only decline moderately, contract walk-away isn’t economically attractive for customers. Losing guarantees, damaging supplier relationships, and potentially not getting scarce capacity—these costs are enough to keep customers fulfilling the contracts. Demand for stable supply from AI servers, cloud providers, and data center customers is also stronger than for typical consumer electronics customers.

But when prices fall deep enough, customers will still do the math. As long as remaining purchase volumes are still large and the spot price relative to the contract floor is low enough, customers may decide it’s still cheaper to buy in the market even if they forfeit the guarantees.

A back-end-weighted mechanism helps mitigate this. As the contract moves further out, remaining RPO declines, and the guarantee-to-remaining-obligation ratio rises—so the cost for customers to abandon contracts becomes higher. The protection strength may be greater later in the contract term, and memory cycles often need more protection later on.

It’s still not unconditional insurance. The degree of LTA protection depends on three numbers: how far spot prices fall, how much purchase obligation the customer still has, and how much guarantee balance remains.

RPO declines over time, and the guarantee/RPO ratio rises; if spot ASP falls below the contract floor by too much, customers may still choose to terminate.

This is the core of the bull-bear disagreement. The bulls see that memory companies finally have large customers’ “real money” long-term commitments. The bears worry that the scale of these commitments may still be insufficient to protect peak earnings. Once the downturn cycle is deep enough, customers will act based on cost.

Not all memory demand is willing to be locked by LTA

There’s also a real-world limit: not all customers are suitable for long-term agreements.

US cloud providers are the ideal target. Their demand is large, credit is strong, they’re sensitive to supply stability for AI infrastructure, and they have more incentive to lock in supply through long-term agreements. Micron has largely completed negotiations with US CSPs; the follow-up is still underway with China CSPs, enterprises, and some other customers.

Consumer businesses are different. SanDisk’s CFO previously said the consumer segment is “more transactional,” and LTA “does not apply.” Mobile phones, PCs, and consumer storage channels are more accustomed to buying based on price and inventory-cycle timing. Once prices fall, customers naturally want to retain flexibility rather than be locked into multi-year floor pricing.

China customers may also not become stable LTA buyers. On one hand, China cloud providers and end customers may be more inclined toward local suppliers. On the other hand, local DRAM and NAND supply expansion adds uncertainty to long-term purchase commitments.

Bernstein estimates that in the overall DRAM and NAND end-market, 30%-50% of shares may be difficult to cover with LTA. Even if top suppliers lock in large US customers, a significant portion of the market will continue to operate based on spot prices, short orders, and cycle expectations.

A split of the DRAM/NAND end market shows that besides US CSPs, there are also China CSPs, enterprise servers, consumer/PC, smart phones, and other demand; roughly 30%-50% of the market may be difficult to cover with LTA.

As long as a sufficiently large share of demand stays in the spot or short-term order system, price signals won’t disappear. As long as price signals exist, capacity expansion by suppliers, inventory reductions by customers, and channel order cuts will all amplify cycle volatility.

AI demand supports valuation, but peak earnings can’t be extrapolated directly

The market is willing to give higher valuations to memory companies, and one reason is that AI demand has changed the bottom shape of this cycle.

On the DRAM side, HBM demand remains strong. Bernstein’s Asia team predicts that HBM prices in 2027 could rise by 2-2.5x versus 2026, while regular DRAM commercial prices have already risen sharply and may remain high over the next 12 months. Even though HBM is more stable than ordinary memory, it shares some production capacity with conventional DRAM, and how capacity is allocated will affect other product lines.

On the NAND side, AI inference and longer context windows also bring a new demand narrative. Early AI training consumed mostly HBM and DRAM, but as inference, agentic AI, and long-context applications increase, storage demand could continue to climb. It’s important to note that statements about capacity related to Vera Rubin shouldn’t be simply written as “NAND capacity for GPUs.” NVIDIA’s official page discloses 20.7TB of HBM4 GPU memory.

In this environment, the value of LTA looks more like fixing a portion of high-demand revenue. If AI demand stays strong, suppliers can lock in some large-customer purchases via long-term agreements. If prices fall back, the guarantees and floor prices can again delay earnings decline.

SanDisk’s pressure-test results point to a similar conclusion. Bernstein’s model shows that under more stringent assumptions, LTA still allows FY29-FY30 EPS to be higher than the no-LTA case across most penetration scenarios, with stronger protection especially later. But the same set of stress tests also indicates that peak earnings can’t be extrapolated simply. Under lower operating-margin scenarios, EPS could be materially below current run-rate levels.

SanDisk FY29-FY30 EPS sensitivity charts show that EPS varies widely across different ASP and LTA penetration rates; LTA improves the downside but can’t lock in peak earnings.

The most important takeaway in this report isn’t “the memory cycle is over,” but “the downturn may be softened.”

Micron and SanDisk securing long-term agreements and financial guarantees shows that large customers are willing to pay for supply certainty in the AI era. For memory companies, that improves revenue visibility over the coming years and makes the capital market more willing to believe that the earnings floor is higher than in the past.

The constraints are equally clear. Guarantees of $33 billion can only provide partial buffering. Consumer segments, China customers, and some transactional demand won’t all move into long-term agreements. Bernstein also estimates that China DRAM share could rise from about 8% to 16% in the coming years, and NAND may face stronger supply pressure after 2028.

What LTA truly needs to prove isn’t whether deals can be signed in a good period—it’s whether customers will fulfill in the next downturn cycle, whether the guarantees will be painful enough, and whether suppliers will continue to maintain production discipline. Until those questions play out, it’s a new buffer for the memory industry—not a cycle-ending “button.”