MR. Reports

July 5th, 2026

Weekly report · 3:22 PM ET. Each edition has its own page.

Weekly report 3:22 PM ET

Memory Report — 2026-07-05

Micron poured concrete on the AI-memory supercycle this weekend, and that gesture, capital committed to capacity that will not ship a wafer for two years, is the cleanest tell of what actually changed this week: the entire industry now treats the memory shortage as a multi-year regime, not a spike, and it is spending, contracting, and litigating on that conviction even as the stock market briefly lost its nerve.

The story of the week: a shortage getting priced in for years

Start with the money on the table. Micron held a groundbreaking ceremony on July 4 for a major expansion of its Higashi-Hiroshima facility in western Japan, a roughly 1.5 trillion yen, about $9.3 billion, project focused on advanced memory chips, particularly high-bandwidth memory essential for AI processors. Shipments from the expanded lines are expected to begin around summer 2028. Japan's Ministry of Economy, Trade and Industry has confirmed it will provide subsidies of up to 500 billion yen, bringing cumulative support to date to 775 billion yen. This is not a bet on a moment. It is a bet that tightness persists into the back half of the decade, and it slots into a global buildout that already spans two leading-edge fabs in Idaho with first DRAM output targeted for 2027, up to four fabs in New York, Singapore, and Taiwan.

Micron's spending does not sit alone. It lands on top of the largest state-backed capacity commitment of the cycle: Seoul's roughly $520 billion push for new Korean fabs, and the 392 trillion won, about $252.5 billion, Chungcheong package aimed squarely at the HBM packaging bottleneck, where Samsung and SK Hynix are building dedicated packaging and NAND fabs that begin operations in 2029. Read together, three companies that control nearly 90% of global DRAM are all pouring capital into capacity that arrives late-decade. Firms bracing for a near-term glut do not do that. Firms convinced the shortage has years to run do.

The pricing data underneath the spending still says shortage, just at a gentler slope. TrendForce's latest survey reveals the DRAM market will remain extremely tight in the third quarter of 2026, but weaker demand from consumer applications and the impact of a higher comparison base are expected to moderate contract price increases to 13 to 18% QoQ. With contract prices already at record highs and consumer price tolerance reaching its limit, NAND Flash contract prices are projected to increase by 10 to 15% QoQ, a noticeably slower pace than in previous quarters. That deceleration is the single most important shift in the week's data, and it is worth naming plainly: it is a slowdown in the rate of increase off record highs, not a reversal. Where AI demand concentrates, the house actually raised its view, lifting its Q3 PC DRAM forecast to 15 to 20% from 8 to 13%, projecting Q4 growth of 3 to 8%, with server DRAM expected to rise 13 to 18% on strong demand and tight supply. The split is the whole thesis in miniature: server and enterprise SSD demand keeps climbing while PCs and phones hit the wall of what consumers will pay.

The scare that wasn't

The week's loudest noise had nothing to do with supply. Investment sentiment chilled sharply after Meta mentioned entering the cloud business by utilizing surplus computing resources, raising concerns about Big Tech's AI capex slowdown, compounded by Apple's product price increases due to rising memory costs. The tape convulsed. SK Hynix logged its worst single-day drop in years Thursday, down more than 14%, Samsung fell around 9%, and the two dragged the Kospi down nearly 8%, before a violent Friday rebound reclaimed most of it. The Kospi closed the week at 8,088.34, down 3.84% from the prior week, having fallen intraday to the 7,370 level.

Strip away the whipsaw and the fundamentals never flinched. The demand wobble was a narrative, not a number, and the supply side stayed locked: the big three keep steering advanced-node wafers toward HBM and server DRAM, starving DDR4 and mature nodes and cascading the shortage down into DDR3 and even DDR2. The clearest rebuttal to the peak-out crowd is contract structure, and this is where suppliers pressed their advantage all week. Micron continues to lock customers into non-cancelable, take-or-pay Strategic Customer Agreements, now 16 of them, while SK Hynix has gone the other way and stripped price caps out of its long-term agreements entirely to keep full upside as prices climb, and both Korean makers are extending contract terms from one year to three to five. And the earnings back the setup: Micron reported operating profit of $33.32 billion for its fiscal third quarter, a 15.4-fold increase from a year earlier.

The number that settles it arrives Tuesday

Everything now points at two dates. Samsung discloses its preliminary second-quarter results on July 7, with the securities-firm consensus compiled by FnGuide pointing to revenue of roughly 169.4 trillion won and operating profit of about 85.5 trillion won, approximately $55.9 billion. That would be an 18-fold surge from a year earlier and a record high, while SK Hynix is expected to post around 64 trillion won in operating profit. Massive performance-bonus provisions remain a wild card, but the companies' actual profit-generating capacity is estimated to have exceeded the 100 trillion won threshold. Then July 10 brings SK Hynix's ADR listing on Nasdaq, an issuance of roughly 45.5 trillion won, about $29.7 billion, representing about 2.5% of total outstanding shares, the largest ADR debut in recorded market history and a direct dollar channel into the AI-memory trade at the very top of the cycle.

What comes next

Watch the July 7 print for one thing above all: whether server DRAM pricing is still outrunning expectations, or whether the export-price wobble that spooked Seoul was the first real crack. A blowout backed by strong memory ASPs quiets the peak-cycle chorus into late July; a provision-heavy or soft number hands the bears their first hard data point. Global investors are closely watching whether these earnings can quell peak-cycle concerns and serve as a turning point for sentiment recovery. July 10 then tests how much AI-memory premium public markets will underwrite in a pure-play maker, and whether SK Hynix's roughly 60% HBM share rotates growth money out of Micron. Note too that TSMC will release its June revenue data on July 10, a barometer for AI accelerator and foundry demand, ahead of its Q2 earnings on July 16.

Further out, three threads deserve real attention. First, the deceleration story: the question for next quarter is whether double-digit server and enterprise growth can offset single-digit consumer segments as record prices exhaust PC and smartphone buyers. Second, the glut clock: Micron's Japan fab, SK Hynix's M17, and the Korean and Chungcheong megaprojects all cluster their output into 2027 to 2029, precisely when the industry's simultaneous capex could tip tightness toward normalization. And third, the wildcard worth tracking, still unconfirmed: Korean media report that CXMT is testing a pilot bonded DRAM line in Hefei, aiming to achieve high-performance DRAM without EUV by fabricating the memory array and peripheral circuitry on separate wafers and bonding them, enabling ultra-high-density DRAM using only DUV lithography with multi-patterning. Samsung is developing its own bonded DRAM under the B1b project and SK Hynix is pursuing similar technology, but some Korean assessments suggest CXMT may currently hold an edge in both the technology and its development speed. Treat that as an unconfirmed report. But if it has legs, and if Washington grants Apple a waiver to buy Chinese memory, China's re-entry into the DRAM supply chain becomes the story that outlives this supercycle. For now, the concrete is being poured, the contracts are non-cancelable, and the shortage is a regime, not a spike. Tuesday tells us how durable.

Sources

← July 5th, 2026 (morning) All briefs