Author: Xiao Jing, Tencent Technology
Editor: Xu Qingyang
Two events occurred simultaneously on the evening of May 26.
Xiaomi released its first-quarter 2026 financial results. Total revenue amounted to RMB 99.1 billion, a year-over-year decline of 10.9%; adjusted net profit was RMB 6.07 billion, a sharp drop of 43.1% year over year. Revenue from the smartphone business was RMB 44.3 billion, down 12.5% year over 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, Lei Jun, President of Xiaomi Group, cited a figure: memory prices for the same configuration have surged nearly fourfold compared to the same period last year; the memory cost for a phone with 12GB LPDDR5 and 512GB UFS has increased by approximately RMB 1,500. He stated that Xiaomi “will not pass on the memory price increase to consumers,” but also predicted that the price hike cycle will continue until 2027 or even 2028. To survive, Xiaomi proactively discontinued its entry-level models, resulting in a quarterly shipment volume dropping to 33.8 million units.
Second, Micron Technology surged over 19% in a single day, pushing its market capitalization past $1 trillion. UBS raised its price target for Micron from $535 to $1,625, an increase of approximately 204%—the highest target among the 46 brokerages covering the stock.
A few days ago, Citigroup raised Micron’s price target from $425 to $840, and HSBC raised it from $750 to $1,100. Wall Street hasn’t been this unified on a single cyclical stock in a long time. A year ago, Micron’s stock was under $110. It has risen eightfold in one year.
On the same day, the trillion-dollar frenzy for selling memory saw profits for buyers halved.
Goldman Sachs played a fascinating role in this frenzy. In December 2025, Goldman Sachs assigned Micron a neutral rating with a price target of $205. In the first quarter of 2026, Goldman Sachs reduced its position in Micron by nearly 20%.
On March 19, the day Micron reported earnings, Goldman Sachs raised its price target from $360 to $400 but maintained a neutral rating, by which time the stock had already far exceeded $400. Shortly after, Micron surged 40% in one week, causing Goldman Sachs to miss the move entirely.
On May 17, Goldman Sachs released a storage industry report concluding there was "the most severe supply shortage in 15 years," and upgraded the overall storage industry rating. However, it maintained a neutral stance on Micron, with the target price still set at $400. Goldman Sachs is either the last sober person in this frenzy or the one who missed out the most.
But this strong disagreement is also worth serious consideration.
01 Why is it skyrocketing? A new story called LTA?
In a research report on May 26, UBS analyst Timothy Arcuri argued that long-term agreements (LTA) are fundamentally eliminating the cyclical nature of the storage industry.
Memory chips are the most commodity-like products in the semiconductor industry. For four decades, the prices of DRAM and NAND have followed a brutal pattern: two years of price increases followed by two years of declines, with price collapses never missing a cycle. The profits of Micron, Samsung, and SK Hynix have fluctuated like ECG readings, and the market has never valued these companies based on "stable earnings." Over the past forty years, the typical valuation range for cyclical stocks has been 8 to 15 times earnings.
Chart: Micron's financial data showing ECG-like fluctuations
The story of UBS is that the "cycle curse" of these companies will be broken, with the protagonist being "AI".
Cloud providers such as Microsoft, Google, Amazon, and Meta are proactively signing three- to five-year fixed-price long-term contracts with memory manufacturers, including advance payments, to secure HBM and DDR5 supplies in the AI arms race. These contracts are not the typical “non-binding” agreements seen in the traditional semiconductor industry—they are binding procurement commitments that lock in volume, price, and even wafer capacity.
Chart: Large Tech Companies' AI Capital Expenditures (2022–2026E): Combined total projected to reach $725 billion in 2026. By company: Amazon $200 billion, Microsoft $190 billion, Alphabet $190 billion, Meta $145 billion. 2026 figures represent the latest guidance upper limits as of April 29 for each company; Microsoft's figures are annual totals based on quarterly data.
In April, Microsoft and Google were reported to be negotiating three-year long-term agreements with SK Hynix for DRAM, including advance payments. Previously, manufacturers sought orders from customers; now, customers are paying deposits to secure capacity. The power dynamics in the supply chain have reversed.
UBS’s model estimates that, if LTA is incorporated into Micron’s earnings forecast, even if DRAM spot prices drop by 50% in fiscal year 2029, Micron’s annual EPS would still remain above $100. LTA can reduce the price fluctuation range of DDR from peak to trough by approximately 50%. By 2027, 20% to 30% of total DDR bits shipped across the industry will be locked in under long-term fixed-price agreements. For leading hyperscalers, 60% to 70% of their DDR5 purchases may already be under fixed contracts.
From a valuation perspective, if the cyclical nature disappears, storage stocks should no longer be valued as cyclical stocks but rather as infrastructure utilities, with P/E ratios rising from 8–15x to 20–30x.
JPMorgan also released a report in mid-May with a similar conclusion, titled directly “LTA is eliminating the cyclical nature of the storage industry.” Citigroup’s logic is that HBM production will crowd out conventional DRAM wafer capacity, leading to long-term scarcity in general-purpose storage.
The surge in Micron's stock price brings a Davis double squeeze in profits and valuation.
02 This storage is not that storage
Wall Street’s use of "storage supercycle" refers to a unified bull market narrative. But "storage" and "storage" are completely different.
The storage market in 2026 exhibits a three-tier differentiation.
The first layer is AI storage: HBM, server DDR5, and enterprise SSDs. Price increases, shortages, and long-term supply agreements locking in capacity are all happening simultaneously. TrendForce predicts that in the second quarter of 2026, DRAM contract prices will rise by 58% to 63% quarter-over-quarter, and NAND Flash contract prices will increase by 70% to 75%; Kioxia has also publicly stated that its 2026 capacity is nearly fully sold out. This layer is the story behind Micron’s trillion-dollar market cap.
The second layer is mobile and embedded storage: mobile DRAM and phone NAND. Prices here have also surged sharply. According to Counterpoint data, DRAM prices rose more than 50% quarter-over-quarter in Q1 2026, while NAND Flash prices increased by over 90% quarter-over-quarter. Related reports from TrendForce indicate that memory traditionally accounted for about 10% to 15% of a phone’s BOM, but has now risen to 30% to 40%, with even more pronounced pressure on entry-level models.
Left chart DRAM (memory) trend: Entry-level devices show the steepest increase, rising steadily from low initial levels to a projected 35% by Q2 2026; high-end devices reach 23%; mid-range devices reach 20%. The dashed line (after Q1 2026) represents projections.
Right chart NAND (flash memory) trend: Prices remained largely stable across all tiers in the first three quarters of 2025, but surged sharply starting in Q4 2025.
Xiaomi is at this level. Its challenge is that "AI has taken away production capacity, leaving less for smartphones, forcing smartphone manufacturers to pay higher prices for the remaining capacity."
The manufacturer has prioritized capacity for AI customers, leaving smartphone manufacturers with few options for contracted purchases. If you want to ship, you must buy at the new contract price; if you don’t, your production line and new product timeline will be affected.
The third layer is PC retail spot: DDR5 modules and consumer-grade SSDs. Here, a reverse trend has emerged. According to TrendForce, in late March, 32GB DDR5 modules in China’s distribution channels dropped from nearly RMB 3,000 to between RMB 1,050 and RMB 1,950 for clearance sales; Tom’s Hardware also reported that DDR5 products in both Chinese and international retail markets have fallen 25% to 30% from their peaks.
But this is primarily a divide between retail spot and contract procurement. PC channels have inventory and can liquidate it; smartphone manufacturers procure under contract and have no liquidation option.
In the same "storage" industry, three tiers are diverging in three directions. This differentiation stems from the fact that the three major storage manufacturers are shifting wafer capacity from consumer-grade to AI-focused production. HBM manufacturing is consuming regular DRAM wafer capacity, enterprise SSDs are displacing consumer-grade NAND supply, leaving less capacity for smartphones and PCs. Smartphone manufacturers, forced to meet shipment targets, have no choice but to accept price increases; PC channels, with ample inventory, can lower prices and clear stock.
Image generated with AI assistance
Micron voluntarily chose to allocate its production capacity to AI customers more willing to pay. In the short term, this is a successful product mix upgrade. But it also means Micron is closing off its escape route—should AI demand slow down, it may not be able to easily shift capacity back.
Micron's earnings report shows that DRAM bit shipments increased only in the low single digits quarter-over-quarter, and NAND bit shipments increased only in the low single digits, with growth primarily driven by higher ASPs. Micron's story today is solely about the extreme shortage in the AI storage segment.
Micron has bet everything on this branch.
03 Can long-term contracts really eliminate cycles?
The logic behind long-term agreements seems solid. Under the spending pace of AI, the supply elasticity of memory chips is extremely low; HBM production capacity takes 18 to 24 months from planning to deployment, and HBM manufacturing crowds out generic DRAM wafers. Cloud providers enter long-term agreements due to concerns about "AI project delays."
But long-term contracts eliminate cyclical fluctuations on one condition: demand does not collapse.
Different institutions use varying methodologies to measure AI CapEx, but the direction is consistent: AI infrastructure investment is moving from the hundreds of billions of dollars to nearly the trillions of dollars. According to some market models, this represents a capital expenditure curve with an annualized growth rate of nearly 40% to 50%.
However, nothing in the physical world grows at over 40% indefinitely. Without requiring an AI bubble to burst, simply a slowdown in growth from 45% to 20% could reverse the supply-demand balance for memory chips within 18 months. All three memory manufacturers are currently expanding production aggressively: Micron’s capital expenditure for fiscal year 2026 is $25 billion, with an additional $10 billion planned for 2027.
Another issue that must be acknowledged 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 increased by only 4% to 6%, while its 196% revenue growth was primarily driven by price increases. Prices can rise and fall, and they tend to drop much faster than they rise. This is the essence of cyclical dynamics.
Let's solve a simple arithmetic problem.
Micron's current market capitalization is $1 trillion. Micron has increased its fiscal year 2026 CapEx to over $25 billion and expects further significant increases in capital expenditures in fiscal year 2027, with some market reports suggesting the additional amount could exceed $10 billion.
Micron's non-GAAP net profit for the second quarter of fiscal year 2026 was approximately $14 billion, which annualizes to about $56 billion, implying a P/E ratio of around 18x. If future price increases and long-term contracts are extrapolated, the P/E ratio could be further reduced to approximately 15x.
It might seem "cheap," but this P/E ratio is based on earnings at the peak of a supercycle—where DDR4 contract prices rose tenfold over 15 months, HBM sold out entirely for the year, and gross margins jumped from 36% to 75%.
Multiplying profits at the cycle peak by a seemingly "reasonable" multiple to arrive at a valuation that appears "cheap" is the classic valuation trap seen at the peak of cyclical stocks.
In 2000, Cisco's P/E ratio was also “only” over 60 times, built on 15 consecutive quarters of revenue growth exceeding 50%. 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 simultaneously.
From the double squeeze to the double kill.
History has taught us one thing: in commodity markets, long-term agreements are never a one-sided "floor." They protect buyers in upcycles and sellers in downcycles—but only if both sides have the capacity and willingness to fulfill their obligations. The very moments when long-term agreements are most needed are often the moments when they are most likely to fail.
This is not to say that Micron is definitely a bubble. The demand for computing power and storage from AI may truly be structural, and LTA may have genuinely rewritten the industry rules—trillion-dollar market caps may just be the starting point.
But when Wall Street as a whole is shouting, “This time is different,” it’s at least worth pausing to ask: What happened the last time everyone was this certain?
In a sense, you can only profit by riding the euphoria of the bubble.
However, Cisco took about 25 years to surpass its closing high from the dot-com bubble era, and now, in today’s AI era, it has finally done so—while the internet has indeed changed everything.