Behind this role reversal is the structural rise of memory demand in the AI computing power arms race. Cloud service providers are beginning to realize that DRAM and NAND are no longer readily available commodities, but the chokehold of AI infrastructure.
In order to lock in critical capacity for the next few years, they have to trade upfront payments and long-term commitments in exchange for safeguards – which are long-term agreements (LTAs).
This change in business model is preciselyJPMorgan ChaseThe core proposition of the latest research report of semiconductor research analyst Jay Kwon:As the proportion of LTA in the shipment structure of storage vendors rises, P/B (price-to-book ratio), a valuation tool that has been used for decades, is outdated and must be switched to P/E (P/E) framework.
Based on this logic, JPMorgan Chase willSamsung ElectronicsThe target price was raised to 480,000 won (8x P/E at the expected EPS from 2026 to 2027),SK HynixThe target price jumped sharply from 1.8 million won to 3 million won,ArmorThe target price doubled directly from 38,000 yen to 80,000 yen, becoming the highest in the whole market.
The underlying logic that really supports this round of revaluation is a gap between supply and demand that is extremely difficult to reverse.The calculation shows that even under the assumption of radical expansion, the supply of AI storage from 2026 to 2030 still can't fill the demand pit of cloud vendors — —This gap is roughly equivalent to the capacity of 450,000 wafers per month.The buyer had no way back, which allowed the seller to really get the pricing power for the first time.
From "building and waiting for customers to buy" to "taking money and working again", the storage industry is moving closer to the "production-to-order" model of Foundry.The profit visibility brought by LTA is the core scalpel to tear off the label of "strong cycle".The recent EPS increase is just an appetizer. What really determines the height of the stock price of storage giants is the fundamental qualitative change brought by these long-term contracts.
The gap of 450,000 wafers/month completely turned the bargaining chips over
Every round of price increases in the storage industry in the past has proved to be a flash in the pan-expansion inevitably pushes prices back. But this time is different, the constraint comes from the combined force of both sides.On the supply side, big storage manufacturers that have just climbed out of the last round of bleeding downward cycle generally suffer from "phobia of production expansion". Without a firm order, no one wants to smash capital when the marginal returns are unknown.
This year, Kioxia's capital expenditure accounted for only 5% of revenue, compared with the average of more than 20% in the past five years. This restraint is not accidental, but a collective choice of the whole industry.SamsungThe proportion of capital expenditure of SK Hynix in the next two years is also controlled in the mid-single digit range.
The demand side is reverse acceleration.The requirements of generative AI for server storage throughput have increased exponentially, and the proliferation of agentic AI has made the demand for enterprise-level solid-state drives explode in an all-round way. Kioxia's single-season ASP increased by more than 100% month-on-month, and the overall price of DRAM has tripled in the past year.
As a result of the two-way squeeze of supply and demand, cloud vendors have lost all confidence at the negotiating table.Micronetc. have taken the lead in disclosing the progress of the long-term agreement. Among them, five LTA contracts signed by an American NAND manufacturer cover more than one-third of its bit demand in fiscal year 2027. It is expected that Samsung and SK Hynix will soon announce the largest long-term deal in semiconductor history.
Great changes in valuation system: From "cyclical stocks" to "quasi-OEM", LTA reshapes pricing power
For the past decade or so, the capital market's valuation of memory chips has been firmly anchored to the price-to-book ratio (P/B). The reason is simple:The products are highly homogeneous, the bargaining power is weak, the capital expenditure is huge, and the performance shows drastic cyclical fluctuations with the macro economy.But in the AI era, this logic is being completely subverted. JPMorgan throws a highly forward-looking view in the report:The storage industry is undergoing a structural shift "from cyclical to secular". The catalyst at the core is precisely the full spread of long-term agreements (LTAs).
With the blowout of demand for HBM (high bandwidth memory) and enterprise-grade SSD by AI servers, cloud service providers (CSPs) are actively signing LTAs with storage vendors for a term of up to 3 to 5 years for the consideration of supply chain security.
According to JPMorgan Chase's observation, there have been many LTA cases in the industry at present, and it is even expected that Samsung and SK Hynix will disclose the largest LTA contract in semiconductor history soon. Some LTAs not only lock in prices (or set price floors), but also include advance payment mechanisms.
This "Make-To-Order" model makes storage manufacturers more and more likeTSMCSuch wafer foundries.The jump in earnings visibility and certainty prompted Wall Street to switch the valuation system of storage stocks from traditional P/B to P/E (P/E).
JPMorgan Chase bluntly said,LTA is a core factor driving industry re-ratings, and based on this, a valuation hub based on 8x P/E of expected earnings from 2026-2027 (mostly 6x in historical downcycles) is given.
Giant Differentiation: Hynix Leads, Samsung "Rich Fire", Kioxia Breakthrough
Despite the overall upward trend of the industry, in this feast, the seats of each family have quietly changed.Samsung: After the compensatory increase, special dividend is a prerequisite for the completion of the valuation
Samsung's share price has soared 143% this year, but its technological laggard on HBM is a clear shortcoming. At present, its improvement in HBM verification is more about catching up with the industry tailwind, rather than really regaining the technical moat.
The logic on the financial side is clearer.The hybrid ASP of DRAM and NAND will end 2026 with year-over-year gains of 293% and 234%, respectively, with operating profit estimates raised by 2% to 11%, but the final EPS estimate was only fine-tuned by 1% to 5% due to strike-induced higher labor costs and tax pressures.
The bigger catalyst lies in cash distribution.Samsung's pool of funds available for shareholder returns from 2024 to 2026 is close to 160 trillion won, with more than 115 trillion won available for special returns after deducting regular Dividend. The market favors direct dividend payouts over share buybacks, and expectations to deliver special Dividend after Q3 2026 have been fully priced. If management can't deliver on schedule, valuation reshaping is missing the last piece of the puzzle.
SK Hynix: It rose by more than 200% during the year, but the compound growth rate of EPS of 31% indicates that it has not reached the top
SK Hynix is the most fierce target in this round of market, and the target price jumped directly from 1.8 million to 3 million won. Underpinning this number is a substantial improvement in earnings quality – EPS is expected to grow at a 31% CAGR from 2026 to 2028.
ASP negotiation in HBM business has entered the fast lane, and the increase in the proportion of long-term associations has not only stabilized the basic price market, but also made Hynix take the absolute initiative in the pricing game with cloud vendors. Even if the follow-up new capacity plan is rolled out, the overall DRAM bit supply growth rate from 2027 to 2028 will still be limited to 20%, and the shortage pattern in the era of high gross profit will continue.
The cash return path is equally clear:From 2025 to 2027, up to 240 trillion won of free cash flow will all be distributed to investors (14 trillion in 2025, 66 trillion in 2026, and 160 trillion in 2027), with total shareholder return yields reaching 5% and 12% in 2026 and 2027, respectively.
Kioxia: After 326% Gain, Investor Day A Shareholder Return Roadmap Determines Whether the Valuation Ceiling Can Be Broken
The upward logic of Kioxia is the most direct: in Q4, ASP soared by more than 100% quarter-on-quarter, operating profit quadrupled quarter-on-quarter, and capital expenditure accounted for 5%. The share of high-margin SSD and storage products in total sales has risen from about 50% to 60%, and the marginal contribution rate will remain at an ultra-high level of 85% to 90% through FY2027.
The LTA has landed – Kioxia finalized a mid-term contract covering 2027 to 2028 in Q4. Combined with the extremely restrained pace of production expansion, 17 trillion yen of free cash flow will be intensely released in the next three years.
At present, 7 times P/E calculates the target price of 80,000 yen, which is still 10% discount to the 15-year historical average (8 times) of the storage industry.
There is only one trigger for the elimination of this discount:A clear and robust shareholder return plan on Investor Day. The pending reduction expectation of the Bain Capital-led consortium is the biggest external variable to suppress the valuation premium. Making money is only the first step. How to distribute profits to shareholders is the key to decide whether the miracle of 326% skyrocketing can continue.
Effective boundaries for the new framework: Can you really say goodbye to "cycles"?
This set of deduction logic is not without failure boundaries. The biggest risk is embedded in the starting point of the logical chain: if the progress of LTA negotiations is opaque, or the AI capital expenditure of cloud vendors is obviously low, the P/E valuation framework based on "profit certainty" will immediately lose support. Specifically:The first is the ultimate torture of AI realization.
JPMorgan hit the nail on the head by pointing out that investors are still concerned about a "disconnect between the AI capex race and the profitability of the semiconductor ecosystem." If the large AI model of cloud vendors fails to bring matching commercial returns for a long time, or there is a breakthrough in the underlying technology (such as reducing the memory consumption of AI servers), the current extremely elevated demand is expected to face the risk of shrinking.
Secondly, the "contract spirit" of LTA still needs to be tested.
It is not without precedent in history. During the DRAM shortage period in 2017, manufacturers also signed forward purchase agreements, but with the slowdown of demand and high inventory, these agreements were easily overturned in the subsequent downward cycle.
This time, whether storage manufacturers can truly strengthen the execution of contracts by introducing third-party financial derivatives guarantees determines whether the logic of "getting rid of the periodic law" can stand.
Finally, the structural evolution of the global capacity map cannot be ignored.
In addition to the restrained expansion of traditional giants at new nodes such as BiCS 8/10, the trend of diversification in global supply chains is accelerating. With the gradual rise and release of emerging markets (such as China's local NAND production capacity), the structure of global storage supply side is becoming more abundant.
In the foreseeable future, this increase in regional production capacity will pose a new test to the existing global supply and demand balance system, and the competitive landscape of the industry will thus add more dimensional variables.
JPMorgan Chase believes,AI has indeed sent a brand-new "armor" to the storage industry, and the implementation of LTA has also made the moat of the industry wider than ever. However, the cycle may be stretched and ironed, but it is difficult to completely eliminate.
In the frenetic capital feast, who can continue to lead in technology, who can exercise restraint in production expansion, and who can truly enjoy the long-term compound interest from "cyclical stocks" to "growth stocks".
Comments