The Numbers Got Bigger. So Did the Warning Signs.

When we last covered this trade, the DRAM ETF had doubled in 43 days. It hasn't stopped.

According to Business Insider, DRAM has now gained 88% since its April 2 launch — in roughly five weeks. Roundhill CEO Dave Mazza told Business Insider his firm built the fund because memory chips were "the most supply-constrained layer of the AI infrastructure buildout" with no dedicated ETF to capture the theme. Investors answered. Hard.

Micron surged 38% in a single week — its best weekly performance since 2008, according to Fortune. The Philadelphia Stock Exchange Semiconductor Index leaped 60% in six weeks. Retail traders piled in at the highest rate in a year that same week, per JPMorgan data cited by Fortune.

Sandisk is now the top S&P 500 performer year-to-date at 558% according to Business Insider. Seagate up 172%. Western Digital up 156%. Micron up 137%. SK Hynix up 186%. Samsung up 114%.

Those are NOT normal stock market numbers. That's mania territory.

What Changed: Google Just Threw a Wrench

On March 24, Alphabet's Google unveiled TurboQuant — a new compression method that Google claims could reduce the memory required to run large language models by six times, according to CNBC. The entire bull thesis for memory stocks rests on AI being permanently memory-hungry. TurboQuant directly attacks that assumption.

Deutsche Bank flagged this in a Tuesday note, warning investors to "continue to brace themselves" for compression and efficiency innovations that could structurally reduce memory demand. Deutsche Bank didn't say the rally is over. But they're not cheerleading either.

Efficiency gains in AI are real. DeepSeek already rattled chip stocks earlier this year by demonstrating that you can get serious AI performance with dramatically less compute. TurboQuant is the memory version of that same story.

Executives Say the Cycle Is Dead. A Harvard Professor Disagrees.

The bull case from inside the industry is loud and unified.

HPE CEO Antonio Neri told CNBC flat out: "We will continue to raise prices because the industry will continue to raise prices. There is not enough supply for demand." A Seagate executive told the South China Morning Post that price hikes are likely "the new normal" for the next few years. SK Hynix told CNBC the entire industry is undergoing "structural change," with hyperscalers moving to long-term multi-year supply contracts instead of annual deals. Broadcom CEO Hock Tan said on his earnings call he has locked in supply through 2028. Meta VP of Engineering Yee Jiun Song told CNBC: "We're absolutely worried about HBM supply."

That's a lot of executives saying the same thing. Which is either a genuine paradigm shift — or exactly what executives say at the top of every cycle.

Willy Shih, a Harvard Business School professor who has tracked semiconductor cycles since the 1980s, told Fortune he's not buying it. "Anytime people show me these curves that just go to the sky with no end, that never continues forever," Shih said. "This too will pass."

Shih's point isn't that AI memory demand is fake. It's that the market always convinces itself the cycle is gone — right before the cycle reasserts itself.

William de Gale, portfolio manager at BlueBox Asset Management, told CNBC the memory industry tends to have "enormous ups and downs." His assessment was blunt: "In the long run it's a pretty dreadful industry. I suspect that's still the case every time people make an argument that the memory cycle is gone, and it's now a long-term value-creating industry — just before it all goes horribly wrong."

What's Actually Driving Demand Right Now

The structural demand case has merit.

IDC, a tech market intelligence firm, told Business Insider that "the memory market is at an unprecedented inflection point, with demand materially outpacing supply." A single AI server requires roughly eight to ten times the DRAM of a traditional server, according to Fortune. That's not hype — that's physics.

Ruben Dalfovo, investment strategist at Saxo, told Business Insider that AI's shift from training to inference is making memory "more strategic, because inference needs speed, bandwidth and power efficiency, not just raw compute."

Demand is real. Whether current stock prices reflect realistic demand or fantasy-land projections is another question.

The Convergence Problem

Business media is treating this as a binary debate: either AI breaks the memory cycle forever, or it doesn't.

The actual risk is that the pace of efficiency improvement outstrips genuine demand growth. Google's TurboQuant, DeepSeek-style breakthroughs, and model optimization are all converging at the same moment that memory stocks are pricing in a decade of constrained supply. The market is not pricing in the possibility that AI gets dramatically more memory-efficient over the next 18 months.

Fortune gave Willy Shih real space to make his case. CNBC's coverage tilted heavily toward the executive consensus. Business Insider's piece read more like a fund prospectus than journalism. Coverage of TurboQuant and the 88% ETF gain in five weeks mostly failed to draw the connection between them.

The Bet

Retail investors are flooding into a sector that just went up 88% in five weeks, guided by a chorus of executives with billions of dollars of personal incentive to say "this time is different." A Harvard expert with 40 years of cycle data says it isn't. And Google quietly dropped a technology that could cut AI memory demand by 83% — the same week Micron had its best week since 2008.

Maybe the cycle really is dead this time. But "this time is different" are the four most expensive words in investing.