Memory's Power Reversal: When the AI Cycle Flips Back

Wall Street used to call the seven hottest companies in the world the Magnificent 7 (M7). Lately it has another name for them: the Laggard 7. Over the past month the M7 ETF fell 12 percent, while a DRAM ETF launched only three months ago is up 176 percent.

They fly the same "AI" flag, so why does one soar while the other lags? The answer fits in one sentence. The side that sells the infrastructure now holds the power, and the side that buys it has lost it. This piece walks through how that reversal happened, in the order of past, present, and future. The back half focuses on which industries lead over the next 5 to 10 years, when Big Tech wins back pricing power, and what signals to watch for.

Part 1. The Past: 30 Years of Memory as the Weaker Hand

Memory (DRAM and NAND) was the most looked-down-on business in all of semiconductors. It was a fight over who could stamp out the same standardized part cheaper and in greater volume, a pure commodity.

The Chicken Game: Lose Money on Purpose to Kill Rivals

The history of the memory industry is the history of the chicken game. When a boom arrives, everyone expands their fabs. Supply then overtakes demand and prices collapse. At that point the well-capitalized leader (Samsung) deliberately ran at a loss and raised output to starve weaker rivals. The results were brutal.

The dozens of DRAM makers that once existed were ground down to just three. Today Samsung, SK Hynix, and Micron control roughly 95 percent of DRAM.

A Brief Glory: The 2016 to 2018 Supercycle

There were booms too. In 2016 to 2018, higher smartphone capacity, early cloud investment, and DRAM supply constraints from the NAND transition all overlapped, and memory shone for a moment.

But that was it. Everyone expanded again and prices broke in 2019. Booms of 4 to 7 quarters, busts of 4 to 8. Revenue down 25 to 40 percent, margins from the 50s into the 20s (or losses), share prices down 50 to 60 percent. That pattern repeated for 30 years without ever breaking.

So in that era the buyer always held the power. When Apple, Google, Microsoft, or Dell said "cheaper, or we buy elsewhere," memory makers had to comply. Memory was an industry that took prices, not one that set them.

Part 2. The Present: AI Flipped the Board, and Memory Took the Power

In the second half of 2025, that formula broke for the first time in 30 years.

HBM, the Game Changer

AI accelerators (Nvidia GPUs and the like) require HBM (high-bandwidth memory). And HBM shares the same production resources as ordinary DRAM. The problem is that HBM devours those resources at an enormous rate.

With capacity piling into HBM, commodity DRAM supply dried up. The result was an explosive jump in prices.

And the Power Flipped

Memory makers no longer even take every order. They allocate, prioritizing only hyperscalers and large OEMs. It is an era where you have to wait in line even when you want to buy.

The Scream From the Buyers (Big Tech)

On the other side the opposite is unfolding. AI is not install-once-and-done like software. As users and tokens grow, spending on GPUs, memory, power, and data centers (CapEx) grows with them.

The "low CapEx, high cash flow" formula that justified Big Tech's premium valuations is crumbling. On top of that, depreciation on GPUs that go obsolete in 3 to 5 years chews through profits fast. So they pass the cost on to consumers, a phenomenon called chipflation.

Part 3. The Future: How the Next 5 to 10 Years Play Out

This is the real heart of it. There are three core questions. Which industries lead, when Big Tech wins back pricing power, and what the signals are.

The Premise: The Cycle Did Not Die, It Just Got Longer

The most common delusion is "this time is different, the cycle is gone." Wrong. The mechanism that ends every memory boom is always the same.

Record margins → everyone expands → supply floods 18 to 36 months later → collides with slowing demand → prices collapse → margins compress

This gear turns again this time too. It just turns over a much longer cycle with a higher trough, for two reasons. First, supply is structurally blocked. Leading-edge process difficulty has spiked, power, labor, and permitting bottlenecks pile up, and the HBM-versus-commodity-DRAM trade-off adds on top, so even pouring in money does not produce capacity within a year or two. Second, demand surges in stages. From inference to edge to self-driving to humanoids, memory demand grows 10x and 100x along the way. So it is an extension of the cycle, not its end. Micron's CEO put it bluntly: there is no sign the shortage clears before 2028.

A Roadmap of Leading Sectors by Phase

Split the next decade into three phases and the strong sectors come into sharp focus.

When Big Tech Wins Back Pricing Power

Here pricing power splits into two. One is cost-negotiation power over memory suppliers (the input side); the other is power over one's own valuation and consumers (the output side). The input side returns when supply catches up with demand; the output side returns when AI monetization is proven and FCF recovers. Consensus points both to roughly 2028.

Bottom line. The earliest Big Tech grips the memory-cost knife again is 2028, and realistically 2028 to 2030. But if physical-AI demand runs hotter than expected, that moment slips into the 2030s.

What Are the Signals: A Three-Stage Leading-Indicator Checklist

A re-reversal (the memory peak) always leaks signals first. Everything below is a leading indicator. When these light up, the lagging indicators, the quarterly revenue, operating profit, margins, and share prices of the three makers, peak and roll over after a lag. Revenue and profit are left off the list because they are exactly the answer sheet we are trying to call in advance. Among leading indicators, the closer to the head (the cause), the earlier the warning but the noisier; the closer to the tail (the price), the more certain but the less lead time.

How to read it: when Stage 1 lights up there is still room, an early warning; when Stage 3 lights up the peak is right in front of you. When four or five light up across stages, the lagging indicators (memory revenue and margins) are about to roll over and Big Tech is about to win back cost power, the signal that Part 1's brutal cycle is returning in a milder form.

Closing: Same AI, Different Fates

The core of this story is simple. AI is one thing, but within it the fates of the sellers and the buyers have split. Right now (2026) is the time of the sellers, of memory and infrastructure. Around 2028, supply catches up and demand cools, and the center of gravity starts to shift. Then Big Tech tries to take back the cost knife, and value moves again to whoever monetizes. But the cycle did not vanish, it just got longer. The trough is certain to come again, only it will likely be less cruel than before.

This is the age of selling picks. But the moment the signals that picks are getting cheap (the checklist above) light up, remember that the center of gravity shifts back to those who mine the gold.

For why memory and semis all collapsed together in June 2026 even though this structure is intact, see the follow-up, Everything's Red: Why June 2026 Sold Off All at Once. The key is not to mix structure with liquidity.

Disclaimer

This article is for informational purposes only and is not investment advice or a recommendation to buy, sell, or hold any security. The companies mentioned (SK Hynix, Samsung Electronics, Micron, TSMC, Broadcom, and others) are analytical examples, and the author may hold some of these positions. All investing carries the risk of loss of principal, and past performance does not guarantee future results. Investment decisions are your own responsibility, and you should consult a qualified financial professional where appropriate.

Sources

• Every Memory Cycle Ends the Same. Until It Doesn't. · UncoverAlpha
• DRAM's Crunch Has No Quick Fix: Pricing Power Into 2027 · k4i.com
• AI memory boom squeezes legacy DRAM supply · S&P Global
• Beware the boom and bust cycle of memory stocks · CNBC
• Memory Wall Bottleneck: AI Compute Sparks Memory Supercycle · TrendForce
• AI to Consume 20% of Global DRAM Wafer Capacity in 2026 · TrendForce
• Global Memory Shortage Crisis · IDC
• Hyperscaler Capex to Exceed Cash Flow by Q3 2026 · Epoch AI
• AI Capex 2026: The $690B Arms Race and FCF Collapse · philippdubach.com
• Why did the memory chicken game keep repeating · Semiconsam
• Update: DRAM prices soar after Qimonda bankruptcy · InfoWorld
• Elpida files for bankruptcy · Network World
• Memory: Are challenges ahead? · McKinsey