The Memory Hole
The Memory Hole
April 1, 2026
Somewhere in Mountain View, a team of Google researchers published a paper about memory compression, and in the eight trading days since, Micron Technology has lost roughly a third of its value. This is not a story about Google's paper. The paper is almost certainly fine. This is a story about what happens when you build a $471 stock on a single premise — that AI will consume memory faster than anyone can make it — and then the market decides, with the patience of a mayfly, that the premise might have an asterisk.
Micron reported fiscal Q2 2026 revenues of $23.86 billion. A year earlier, the same quarter produced $8.05 billion. Gross margins guided to approximately 80% next quarter — above Nvidia, the company whose chips created the demand that Micron is now scrambling to supply. CEO Sanjay Mehrotra told investors that key customers are receiving only 50% to two-thirds of their memory requirements. The company signed its first-ever five-year supply agreement. Supply constraints are expected to persist past 2026, with meaningful new capacity not arriving until 2028 at the earliest. None of this mattered. The stock fell 10% on the Monday after earnings, and kept falling.
What the market latched onto: TurboQuant. Google Research published a compression algorithm achieving a 6x to 8x reduction in KV cache memory size for large language models, with zero accuracy loss. The bears seized on this like it was a verdict. If AI can do more with less memory, the entire supercycle thesis — the one that drove MU up 300% in a year — rests on shakier ground than assumed. SK Hynix and Samsung fell 6% and 5% respectively in Seoul. Kioxia dropped nearly 6% in Tokyo. The whole sector repriced in 48 hours based on a research paper that even the skeptics acknowledged was evolutionary rather than revolutionary.
Here's the thing about that framing. When you address a bottleneck, you enable more capable hardware. More capable hardware supports more powerful models. More powerful models demand more of everything downstream, including memory. Ray Wang at SemiAnalysis made exactly this point in real time: improving memory efficiency doesn't reduce memory demand, it raises the ceiling on what models can do, which creates new demand at a higher level. Jevons' Paradox has been restated about ten thousand times in the context of AI efficiency gains, and markets keep forgetting it every few months when a new compression paper lands.
But none of that touches the real story, which is: Micron had gone up 300% in twelve months. It didn't need a bad reason to sell off. It needed any reason at all.
This is the structural problem with the AI trade in Q1 2026. The narrative is real, the revenue is real, the supply constraints are real — and the stocks are so far ahead of even the best-case fundamental scenarios that they've become pure momentum vehicles. When momentum reverses, the debate about whether TurboQuant is existential or evolutionary is mostly beside the point. Citi analyst Atif Malik flagged concerns about rising capital expenditure in fiscal 2027 and the possibility that margins are peaking. Micron's planned capex for fiscal 2026 exceeds $25 billion, with construction-related spending projected to rise by another $10 billion in fiscal 2027. Samsung has committed north of 110 trillion won toward R&D and new production facilities. SK Hynix is exploring a U.S. listing to raise up to $14.4 billion, alongside a $7.97 billion order for ASML equipment.
The entire memory oligopoly is spending simultaneously, on the same bet, at the same time. Three companies. Same thesis. Billions in overlapping commitments racing to capture the same AI-driven demand pool. The history of the memory industry is a history of exactly this dynamic, except that every prior cycle was eventually resolved by a brutal, margin-destroying glut. The bull case is that this cycle is different — AI demand is structural, supply is physically constrained, fabs take years to build. The bear case is that it always looks different from the inside of the cycle.
What's changed for the broader market is more interesting than the Micron debate itself. The Nasdaq fell into correction territory last week. The Dow followed. The S&P 500 is roughly 9% off its all-time high, sitting precariously below its 200-day moving average — a level the technical crowd treats as a meaningful threshold. The VIX averages 17 when it's above its 200-day moving average; it averages 26 when the index is below it. The S&P is below it now.
The FOMC met on March 18 and voted to hold rates at 3.5–3.75%. One dissent: Stephen Miran, who preferred a 25-basis-point cut. That dissent is worth noting not because it changes anything in the near term, but because it tells you something about the internal pressure building inside the Committee. The hawks aren't winning an ideological debate anymore — they're being held in place by an oil shock that won't allow the obvious move. February nonfarm payrolls came in at minus 92,000. Economists expect March to recover to something like plus 57,000. If Friday's print surprises to the downside again, the gap between Miran and the rest of the table gets harder to explain as a principled disagreement. It starts to look like the Committee knows what it should do and is refusing to do it because imported inflation from a war it didn't start has cornered its options.
Nike, for what it's worth, beat estimates Tuesday evening. EPS of $0.35 against a $0.31 consensus. Revenue of $11.28 billion versus $11.24 expected. China EBIT of $467 million crushed the $269.5 million estimate by a factor of almost two. This is the company that was supposed to be terminally broken — market share lost to On, Hoka, New Balance, the whole narrative of a brand that had lost the cultural thread. A China beat of that magnitude at least complicates the obituary. The stock is up, but quietly, buried beneath the oil headlines and the tech unwind.
And then there's the $3 trillion figure. Bloomberg reported that Microsoft is in talks with Chevron and Engine No. 1 over a roughly $7 billion natural-gas power plant in West Texas — 2,500 megawatts, one of the largest of its kind in the U.S. — to supply the electricity AI infrastructure requires. That single deal, taken at face value, tells you more about the state of the AI build-out than any PMI print. The technology companies are now negotiating directly with energy majors to secure power generation at scale, bypassing utilities and grid operators, building what are essentially captive energy systems. The capex figures discussed in these contexts — $3 trillion across the AI infrastructure ecosystem over the next several years — make Micron's $25 billion look like a rounding error.
The irony is that the same war driving oil above $100 and threatening to tip the economy into stagflation is also making the power economics of AI data centers significantly more complicated. Natural gas prices in Europe are up more than 70% since late February. U.S. industrial electricity costs are moving. And the companies planning to spend $3 trillion on data centers are now looking at a world where the energy inputs for that investment are materially more expensive than they were four months ago.
The AI supercycle is not over. But it is being stress-tested simultaneously by a compression algorithm, a global energy shock, a Fed that cannot cut, and a set of quarterly earnings that were so good they generated selling pressure. In the history of bull markets, that combination — a compelling structural story under serious near-term macro pressure — has a way of separating the investors with conviction from the ones who were riding momentum and calling it the same thing.
The 200-day moving average is a line on a chart. But right now, it's doing a lot of explanatory work.