The trillion-dollar frenzy of selling memory, profits from buying memory are halved

Author: Xiao Jing, Tencent Technology

Editor: Xu Qingyang

On the evening of May 26, two events occurred simultaneously.

Xiaomi released its Q1 2026 financial report. Total revenue was 99.1 billion yuan, a year-on-year decrease of 10.9%; adjusted net profit was 6.07 billion yuan, a year-on-year plunge of 43.1%. Revenue from the mobile phone business was 44.3 billion yuan, down 12.5% year-on-year, with gross margin falling to 10.1%, a decrease of 2.3 percentage points compared to the same period last year.

During the earnings call, Xiaomi Group President Lu Weibing mentioned a figure: the price of the same version of memory has skyrocketed nearly four times compared to the same period last year. For a phone with 12GB LPDDR5 + 512GB UFS configuration, the memory cost alone has increased by about 1500 yuan. He stated that Xiaomi "will not pass on the increased memory costs to consumers," but also predicted that the price increase cycle will continue until 2027 or even 2028. To survive, Xiaomi has actively cut entry-level models, with quarterly shipments dropping to 33.8 million units.

The second event was that Micron Technology's stock surged over 19% in a single day, pushing its market value past 1 trillion dollars. UBS raised Micron's target price from $535 directly to $1625, a one-time increase of about 204%, making it the highest target price among the 46 brokerages covering Micron.

A few days earlier, Citigroup had just raised Micron's target price from $425 to $840, and HSBC also raised it from $750 to $1100. Wall Street has not seen such a unified opinion on the same cyclical stock in a long time. Twelve months ago, Micron's stock price was less than $110. It has increased eightfold within a year.

On the same day, a trillion-dollar frenzy for selling memory occurred, while profits for buying memory were halved.

Goldman Sachs played a thought-provoking role in this frenzy. In December 2025, Goldman Sachs gave Micron a neutral rating with a target price of $205. In Q1 2026, Goldman Sachs reduced its position in Micron by nearly 20%.

On March 19, the day Micron released its earnings report, Goldman Sachs raised its target price from $360 to $400 but maintained a neutral stance, even though the stock price had already far exceeded $400. Then Micron's stock surged 40% in a week, and Goldman Sachs missed out.

On May 17, Goldman Sachs released a report on the storage industry, concluding that it was "the most severe supply shortage in 15 years," raising the overall rating for the storage industry. However, it remained neutral on Micron, with a target price still at $400. Goldman Sachs, in this alternative stance, is either the last sober person in this frenzy or the one who has missed out the most.

But this strong divergence is also worth serious consideration.

01 Why the Crazy Surge? A New Story Called LTA?

UBS analyst Timothy Arcuri's report on May 26 argues that Long-Term Agreements (LTA) are fundamentally eliminating the cyclicality of the storage industry.

Storage chips are the most commodity-like products in the semiconductor industry. The prices of DRAM and NAND have followed a brutal rule for forty years: they rise for two years, fall for two years, and price collapses have never been absent. The profits of Micron, Samsung, and SK Hynix resemble an electrocardiogram, and the market has never dared to value these companies based on "steady-state earnings." For forty years, the valuation range for cyclical stocks has been approximately 8 to 15 times earnings.

Figure: Micron's financial data shows electrocardiogram-like fluctuations

UBS's narrative is that the "cycle curse" of these companies will be broken, with "AI" as the main character.

Cloud providers like Microsoft, Google, Amazon, and Meta, in order to secure HBM and DDR5 supply in the AI arms race, have begun to actively sign fixed-price long-term contracts with storage manufacturers for 3 to 5 years, often with upfront payments. These contracts are not the traditional "intentional" agreements in the semiconductor industry; they are binding procurement commitments that lock in volume, price, and even wafer production capacity.

Figure: AI capital expenditures of major tech companies (2022---2026E): The total for the four companies is expected to reach $725 billion in 2026. Individually, Amazon $200 billion, Microsoft $190 billion, Alphabet $190 billion, Meta $145 billion. The 2026 data is based on the latest upper guidance from each company as of April 29.

Reports in April indicated that Microsoft and Google were negotiating a three-year DRAM long-term contract with SK Hynix, which included upfront deposits. Previously, manufacturers sought orders from customers; now, customers are paying deposits to secure production capacity. The power dynamics in the supply chain have reversed.

UBS's model shows that if LTAs are included in Micron's earnings forecast, even if DRAM spot prices plummet by 50% in fiscal year 2029, Micron's annual earnings per share can still maintain above $100. LTAs can narrow the price fluctuation range of DDR from the peak of the cycle to the trough by about 50%. By 2027, 20% to 30% of the total bit shipments of DDR in the industry will be locked in by fixed-price long-term contracts. Among the DDR5 purchases of leading hyperscalers, 60% to 70% may already be under fixed contracts.

From a valuation perspective, if cyclicality disappears, storage stocks should not be valued as cyclical stocks but rather as infrastructure utilities, shifting from 8 to 15 times PE to 20 to 30 times PE.

J.P. Morgan also released a report in mid-May with a similar conclusion, titled "LTA is Eliminating Cyclicality in the Storage Industry." Citigroup's logic is that HBM production will crowd out general DRAM wafer capacity, leading to a long-term shortage in general storage as well.

Micron's stock price surge has ushered in a switch in profit and valuation systems, a "Davis Double."

02 This Storage is Not That Storage

Wall Street speaks of a "storage supercycle" as a unified bull market narrative. However, "storage" and "storage" are completely different.

The storage market in 2026 shows three layers of differentiation.

The first layer is AI storage: HBM, server DDR5, enterprise SSDs. Here, price increases, shortages, and long-term contracts locking in capacity are happening simultaneously. TrendForce predicts that in Q2 2026, DRAM contract prices will rise by 58% to 63% quarter-on-quarter, and NAND Flash contract prices will rise by 70% to 75%; Kioxia has also publicly stated that its capacity for 2026 is basically sold out. This layer is the story of Micron's trillion-dollar market value.

The second layer is mobile and embedded storage: mobile DRAM and phone NAND. Prices are also rising sharply here. Counterpoint data shows that in Q1 2026, DRAM prices rose over 50% quarter-on-quarter, and NAND Flash prices rose over 90% quarter-on-quarter. TrendForce's related report indicates that memory used to account for about 10% to 15% of a phone's BOM, but now it has risen to 30% to 40%, with pressure more evident on low-end models.

Left chart DRAM (memory) trend: Low-end models have the highest increase, climbing from initial low levels to a predicted 35% in Q2 2026; high-end models to 23%; mid-range models to 20%. The dashed part (after Q1 2026) is a forecast.

Right chart NAND (flash memory) trend: All price levels remained stable in the first three quarters of 2025, but began to surge sharply from Q4 2025.

Xiaomi is situated in this layer. Its pain is that "AI has taken away production capacity, leaving less for mobile phones, and phone manufacturers must pay higher prices for the remaining capacity."

Original manufacturers have prioritized production capacity for AI customers, leaving phone manufacturers with limited choices in contract procurement. If you want to ship, you have to buy at the new contract price; if you don't buy, production lines and new product rhythms will be affected.

The third layer is PC retail spot: DDR5 modules, consumer-grade SSDs. Here, there are counter-directional fluctuations. TrendForce reports that by the end of March, the price of a 32GB DDR5 module in Chinese channels dropped from nearly 3000 yuan to 500 to 1050 yuan, with some clearance prices as low as 1950 yuan; Tom's Hardware also noted that some DDR5 products in the Chinese and overseas retail markets fell 25% to 30% from their peak.

However, this is mainly a split between retail spot and contract procurement. The PC channel has inventory and can sell off; phone manufacturers have to procure according to contracts and have no option to sell off.

In the same "storage" industry, there are three layers with three different directions. The essence of this differentiation is that the three major storage manufacturers are shifting wafer capacity from consumer-grade to AI. HBM production crowds out general DRAM wafers, and enterprise SSDs crowd out consumer-grade NAND supply, leaving less capacity for mobile phones and PCs. Phone manufacturers, needing to ship, are forced to accept price increases; PC channels, having sufficient inventory, can pressure prices down.

Image generated with AI assistance

Micron and others have actively chosen to allocate capacity to AI customers who are more willing to pay. In the short term, this is a beautiful upgrade in product structure. But it also means that Micron is blocking its retreat; once AI demand slows down, production capacity may not be easily redirected back.

Micron's financial report shows that quarter-on-quarter, DRAM bit shipments only grew in the single digits, and NAND bit shipments only grew in the low single digits, with growth mainly coming from ASP increases. Micron's current story is solely based on the "extreme shortage of AI storage in this branch."

Micron has bet everything on this branch.

03 Can Long-Term Contracts Really Eliminate Cycles?

The logic of long-term contracts seems solid. Under the spending rhythm of AI, the supply elasticity of storage chips is extremely low; HBM capacity takes 18 to 24 months from planning to production, and producing HBM will crowd out general DRAM wafers. Cloud providers sign long-term contracts out of concern for "AI project delays."

However, for long-term contracts to eliminate cyclicality, there is a prerequisite: the demand side must not collapse.

Different institutions have different statistical criteria for AI CapEx, but the direction is consistent: investment in AI infrastructure is surging from hundreds of billions to nearly a trillion dollars. According to some market models, this is a capital expenditure curve approaching 40% to 50% annualized growth.

However, in the physical world, nothing grows at over 40% forever. It doesn't require an AI bubble to burst; a slowdown from 45% to 20% could reverse the supply-demand balance for storage chips within 18 months. All three storage manufacturers are currently expanding production crazily, with Micron's CapEx for fiscal year 2026 at $25 billion, and an additional $10 billion expected in 2027.

Another thing that must be faced is that when a company's revenue growth relies entirely on price elasticity rather than volume elasticity, the story is fragile. Micron's shipment volume has only increased by 4% to 6%, with revenue growth of 196% mainly driven by price increases. Prices can rise and fall, and they tend to fall much faster than they rise. This is also the essence of cyclicality.

Let's do a simple arithmetic problem.

Micron currently has a market value of $1 trillion. Micron has raised its CapEx for fiscal year 2026 to over $25 billion and expects significant increases in capital expenditures for fiscal year 2027, with some market reports mentioning increments possibly exceeding $10 billion.

Micron's non-GAAP net profit for Q2 of fiscal year 2026 is about $14 billion for the quarter, simply annualized to about $56 billion, corresponding to about an 18 times PE. If subsequent price increases and long-term contracts continue to be projected, the PE could be calculated to be around 15 times.

It seems reasonable to say "cheap." But this PE's denominator is based on a supercycle peak profit where DDR4 contract prices increased tenfold over 15 months, HBM was sold out for the year, and gross margins jumped from 36% to 75%.

Multiplying peak profits by a seemingly "reasonable" multiple to arrive at a seemingly "not expensive" valuation is precisely the classic valuation trap seen when cyclical stocks peak.

In 2000, Cisco also had a PE of "only" over 60 times, built on the foundation of 15 consecutive quarters of 50%+ revenue growth. When growth slowed from 50% to 20% and then to 0%, EPS didn't need to drop much for the stock price to fall 80%, as both the multiple and earnings contracted.

From a Davis Double to a Double Kill.

One thing history teaches us is that in the commodity market, long-term contracts have never been a one-sided "floor." They protect buyers in an upcycle and protect sellers in a downcycle, but the premise is that both sides have the ability and willingness to perform. The moment long-term contracts are truly needed is precisely when they are most likely to fail.

This is not to say that Micron is necessarily a bubble. The demand for computing power and storage from AI may indeed be structural, LTAs may have truly rewritten industry rules, and a trillion-dollar market value may just be the starting point.

But when all of Wall Street simultaneously shouts "this time is different," it is at least worth pausing to ask: what happened the last time everyone was so certain?

In a sense, enjoying the frenzy of a bubble is the way to make money.

However, it took Cisco about 25 years to finally surpass the closing high from the internet bubble era in today's AI age, and the internet has indeed changed everything.