The Trillion-Dollar Memory Sales Frenzy, and the Memory Buyers’ Profits Slashed in Half
TechFlow Selected TechFlow Selected
The Trillion-Dollar Memory Sales Frenzy, and the Memory Buyers’ Profits Slashed in Half
AI’s demand for computing power and storage may indeed be structural; LTA may truly rewrite industry rules, and a trillion-dollar market cap may be just the starting point.
Navigating Web3 tides with focused insightsAuthor: Xiao Jing, Tencent Technology
Two events occurred simultaneously on the evening of May 26.
Xiaomi released its Q1 2026 financial results. Total revenue amounted to RMB 99.1 billion, down 10.9% year-on-year; adjusted net profit stood at RMB 6.07 billion, plunging 43.1% YoY. Revenue from smartphone operations totaled RMB 44.3 billion, down 12.5% YoY; gross margin fell to 10.1%, a decline of 2.3 percentage points from the prior-year period.
During the earnings call, Xiaomi Group President Lu Weibing cited one key figure: memory prices for identical configurations surged nearly fourfold year-on-year. For a smartphone equipped with 12GB LPDDR5 RAM and 512GB UFS storage, memory costs alone increased by approximately RMB 1,500. He stated that Xiaomi “will not pass on rising memory costs to consumers,” while also forecasting that this price-increase cycle will extend into 2027—and possibly even 2028. To survive, Xiaomi proactively discontinued its entry-level models, causing quarterly shipment volume to drop to 33.8 million units.
The second event: Micron Technology surged over 19% in a single day, pushing its market capitalization above $1 trillion. UBS raised its Micron target price from $535 directly to $1,625—a staggering 204% increase—making it the highest target price among all 46 brokerages covering Micron.
Just days earlier, Citigroup had lifted Micron’s target price from $425 to $840, while HSBC raised its target from $750 to $1,100. Wall Street has rarely shown such unanimity on a cyclical stock. Twelve months ago, Micron’s share price was under $110—representing an eightfold gain within one year.
On the same day, memory sellers celebrated a trillion-dollar狂欢 (carnival), while memory buyers suffered a halving of profits.
Goldman Sachs played a particularly intriguing role in this carnival. In December 2025, Goldman assigned Micron a neutral rating with a $205 target price. In Q1 2026, it reduced its Micron position by nearly 20%.
On March 19—the day Micron reported earnings—Goldman raised its target price from $360 to $400 but maintained its neutral rating, even though the stock price had already far exceeded $400. Micron then surged 40% in one week, and Goldman missed the move precisely.
On May 17, Goldman published a memory industry report concluding there is “the most severe supply shortage in 15 years” and upgraded its overall rating for the memory sector. Yet it kept its rating on Micron neutral and retained the $400 target price. Goldman stands apart—either the last sober participant in this carnival, or the one who missed the rally most spectacularly.
But such stark divergence merits serious reflection.
01 Why the Frenzy? A New Story Called LTA?
In his May 26 research note, UBS analyst Timothy Arcuri argued that Long-Term Agreements (LTAs) are fundamentally eliminating the cyclical nature of the memory industry.
Memory chips represent the semiconductor sector’s most commodity-like product. For four decades, DRAM and NAND prices have followed a brutal rhythm: up for two years, down for two years—with price collapses never failing to occur. Profits at Micron, Samsung, and SK Hynix have resembled electrocardiograms; markets have never dared value these firms based on “steady-state profitability.” Over 40 years, cyclical stocks’ valuation multiples have generally fluctuated between 8x and 15x P/E.
Figure: Micron’s financial data—EKG-style volatility
UBS’s narrative is that these companies’ “cyclical curse” is breaking—and AI is the protagonist.
Cloud providers—including Microsoft, Google, Amazon, and Meta—are signing fixed-price, three-to-five-year LTAs with memory manufacturers to secure supplies of HBM and DDR5 amid the AI arms race—and many contracts include upfront payments. These are not traditional semiconductor “letters of intent”; they are binding procurement commitments locking in volume, price, and even wafer capacity.
Figure: AI CapEx by major tech firms (2022–2026E): Combined projected 2026 CapEx reaches $725 billion. Breakdown: Amazon ($200B), Microsoft ($190B), Alphabet ($190B), Meta ($145B). 2026 figures reflect latest upper-end guidance as of April 29; Microsoft’s figure aggregates quarterly data into calendar-year totals.
Reports in April indicated Microsoft and Google were negotiating three-year DRAM LTAs with SK Hynix—including advance deposits. Previously, memory makers begged customers for orders; now, customers pay deposits to lock in capacity. The balance of power across the supply chain has flipped.
UBS’s modeling shows that incorporating LTAs into Micron’s earnings forecasts implies that even if spot DRAM prices crash 50% in fiscal 2029, Micron’s annual EPS would still exceed $100. LTAs can narrow DDR price volatility—from peak to trough—by roughly 50%. By 2027, 20%–30% of the industry’s total DDR bit shipments will be locked in via fixed-price LTAs; for top-tier hyperscalers, 60%–70% of DDR5 purchases may already be covered by fixed contracts.
From a valuation perspective, if cyclicality vanishes, memory stocks should no longer trade as cyclical equities—but rather as infrastructure utilities, shifting from 8x–15x P/E to 20x–30x P/E.
JPMorgan Chase issued a similar report in mid-May titled “LTAs Are Eliminating Memory Industry Cyclicality.” Citigroup’s logic centers on HBM production squeezing standard DRAM wafer capacity, thereby triggering long-term shortages across general-purpose memory.
Micron’s soaring share price reflects a classic “dual Davis effect”: simultaneous expansion in both earnings and valuation multiples.
02 This Memory Is Not That Memory
Wall Street’s “memory supercycle” narrative presents a unified bullish story. But “memory” is not monolithic—“memory” differs radically from “memory.”
The 2026 memory market exhibits three distinct tiers of divergence.
Tier 1: AI Memory—HBM, server DDR5, enterprise SSDs. Here, prices surge, shortages mount, and long-term contracts lock in capacity simultaneously. TrendForce forecasts Q2 2026 DRAM contract prices rising 58%–63% quarter-on-quarter, and NAND Flash contract prices climbing 70%–75% QoQ; Kioxia publicly stated its 2026 capacity is virtually sold out. This tier tells the story behind Micron’s trillion-dollar market cap.
Tier 2: Mobile & Embedded Memory—mobile DRAM and smartphone NAND. Prices here are also surging sharply. Counterpoint data shows Q1 2026 DRAM prices rose over 50% QoQ, and NAND Flash prices jumped over 90% QoQ. TrendForce reports indicate memory previously accounted for ~10%–15% of a smartphone’s bill of materials (BOM), but now comprises 30%–40%—placing particular strain on budget models.
Left chart – DRAM (RAM) trend: lowest-tier devices show steepest gains—rising from initial lows to a projected 35% in Q2 2026; high-end devices reach 23%; mid-tier hits 20%. Dashed line (post-Q1 2026) indicates forecasted values.
Right chart – NAND (flash) trend: prices remained relatively stable across all tiers through Q3 2025, then spiked sharply starting Q4 2025.
Xiaomi resides squarely in this tier. Its pain stems from “AI siphoning off capacity, leaving less for smartphones—so phone makers must pay higher prices for the remaining supply.”
Memory manufacturers prioritize AI customers’ capacity allocation. Smartphone makers have little choice but to accept new contract pricing. If you want to ship devices, you must buy at the new contracted rate; if you don’t, your production lines and product launches suffer.
Tier 3: PC Retail Spot Market—DDR5 modules, consumer SSDs. Here, the trend runs counter. TrendForce reported that by end-March, China’s channel price for a 32GB DDR5 module fell from near RMB 3,000 to RMB 1,050–1,500, with some clearance deals dipping to RMB 1,950; Tom’s Hardware noted certain DDR5 products dropped 25%–30% from peaks in both Chinese and overseas retail markets.
This divergence primarily reflects the split between retail spot markets and contract procurement. PC distributors hold inventory and can liquidate aggressively; smartphone makers procure under contract and lack the flexibility to dump stock.
Within the same “memory” industry, three tiers move in three directions. At its core, this fragmentation reflects the three major memory manufacturers deliberately shifting wafer capacity away from consumer applications toward AI. HBM production crowds out standard DRAM wafers; enterprise SSDs displace consumer NAND supply—leaving less capacity for smartphones and PCs. Smartphone makers, compelled to ship, absorb price hikes; PC distributors, flush with inventory, slash prices to clear stock.
Image generated with AI assistance
Micron and peers have actively chosen to allocate capacity to AI customers willing to pay premium rates. In the short term, this represents a successful product mix upgrade. Yet it also means Micron is closing off its exit routes—if AI demand slows, reverting capacity back to consumer segments may prove difficult.
Micron’s earnings report shows DRAM bit shipments grew only in the low single digits sequentially, and NAND bit shipments rose only in the low single digits—revenue growth came overwhelmingly from ASP increases. Micron’s current story rests entirely on “extreme scarcity within the AI-memory subsegment.”
Micron has bet everything on this subsegment.
03 Can LTAs Truly Eliminate Cyclicality?
The LTA logic appears robust. Given AI’s spending cadence, memory chip supply elasticity is extremely low—HBM capacity takes 18–24 months from planning to ramp, and HBM production cannibalizes generic DRAM wafers. Cloud providers sign LTAs fearing “AI project delays.”
Yet LTAs eliminate cyclicality only if demand does not collapse.
Different institutions use varying methodologies to tally AI CapEx—but all point in the same direction: AI infrastructure investment is surging from the hundreds-of-billions to nearly the trillions of dollars. Some market models estimate this trajectory implies an annualized CapEx growth rate approaching 40%–50%.
However, nothing in the physical world grows perpetually at >40% annually. An AI bubble burst isn’t required—merely a slowdown from 45% to 20% growth could reverse supply-demand equilibrium within 18 months. All three memory makers are aggressively expanding capacity: Micron’s FY2026 CapEx stands at $25 billion; FY2027 will add another $10 billion.
Another unavoidable reality: when a company’s revenue growth depends entirely on pricing elasticity—not volume elasticity—the story becomes fragile. Micron’s unit shipments rose just 4%–6%, yet revenue surged 196%—driven almost entirely by price hikes. Prices can rise—and fall—just as rapidly, and often fall faster than they rise. That is the essence of cyclicality.
Let’s run a simple arithmetic exercise.
Micron’s current market cap stands at $1 trillion. It has raised FY2026 CapEx to over $25 billion and expects FY2027 CapEx to increase significantly—some market reports suggest an incremental $10+ billion.
Micron’s non-GAAP net income for Q2 FY2026 stood at ~$14 billion—annualizing to ~$56 billion, implying a ~18x P/E. Extrapolating further price increases and LTAs yields a P/E around 15x.
That may appear “cheap.” But this P/E denominator reflects earnings achieved at the absolute peak of a supercycle—where DDR4 contract prices surged tenfold over 15 months, HBM sold out entirely for the full year, and gross margin rocketed from 36% to 75%.
Multiplying peak-cycle earnings by a seemingly “reasonable” multiple to arrive at an apparently “undemanding” valuation is the textbook valuation trap that signals the top of a cyclical stock.
Cisco’s P/E in 2000 was “only” ~60x—but built on 15 consecutive quarters of >50% revenue growth. When growth slowed from 50% to 20% and ultimately to zero, EPS didn’t need to plummet much for the stock to drop 80%—as both the multiple and earnings contracted simultaneously.
From a dual Davis effect to a dual Davis squeeze.
History teaches us one thing: in commodity markets, LTAs are never one-sided “price floors.” They protect buyers during up-cycles and sellers during down-cycles—but only if both sides retain the ability and willingness to fulfill obligations. The moment LTAs become most needed is precisely when they’re most likely to fail.
This isn’t to claim Micron is necessarily a bubble. AI’s demand for compute and memory may indeed be structural; LTAs may genuinely rewrite industry rules; a $1-trillion market cap may merely mark the beginning.
But when the entire Wall Street chorus chants “this time is different,” it’s at least worth pausing to ask: what happened the last time everyone was equally certain?
In a sense, profiting from the bubble requires participating in the carnival.
Yet Cisco took roughly 25 years—until today’s AI era—to finally surpass its internet-bubble-era closing highs, even though the internet truly did transform everything.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
