The biggest two weeks of the cycle so far. Micron printed a blowout that ended the cyclical argument for now, Korea crashed and recovered in 48 hours, the makers announced a trillion dollars of spending and the market hated it, and Apple started lobbying Washington to buy memory from a blacklisted Chinese company. None of it moved my tripwire. Here is what actually happened, the way I always do it. What is noise, what is signal, and what I am watching.

1. Micron ended the cyclical argument, at least for now

I covered the setup last week and the print deserves its own piece, so I will keep this tight. Micron reported revenue of 41.46 billion against about 35.6 expected, EPS of 25.11 against about 20.60, and then guided next quarter to 50 billion in revenue and 31 dollars of EPS, roughly 7 billion above where the street was. The stock jumped about 14.5 percent.

The number that matters is not the beat. It is the backlog. Micron disclosed around 100 billion dollars of remaining performance obligations across 16 strategic customer agreements, with minimum committed volumes and minimum pricing, take or pay and price floors. Mehrotra said tight conditions persist beyond calendar 2027 and that he has no line of sight on when supply catches demand. That is the structural transformation in one line: customers are now funding capacity and locking price, which is the literal mechanism that removes the cycle aspect for now in the business.

The honest counter, which I will not skip. The multiple stayed around 9 to 10 times forward earnings even after the blowout, so the market itself still prices this as cyclical. Capex is ramping to about 10 billion next quarter, which is the supply response the bears are waiting for. And the price floors have never been tested through an actual glut. The print extends the runway. It does not repeal the cycle. More on this in the standalone piece.

2. Korea crashed, and it was positioning, not fundamentals

The day before Micron, Korea's KOSPI closed down 10 percent, SK Hynix and Samsung both off more than 12 percent, with the exchange tripping circuit breakers twice in a single session. Kioxia fell more than 15 percent in Japan. Memory names in the US dropped 7 to 10 percent from record highs. It felt like the top.

It was not. No demand cancellation, no pricing crack, no hyperscaler capex cut. The triggers were pre earnings profit taking, a broad global AI valuation unwind, and investor Dan Niles publicly cutting semiconductor holdings and warning of a near term speed bump. A leveraged long unwind in the most crowded trade on Earth, not a thesis break. My tripwire was never the stock price, a scary headline, or a China announcement. It is hyperscaler capex guidance turning down, and it did not. I held, the way I held through the June drawdown, and 48 hours later Micron's print snapped the whole sector back. File the crash under noise. The thing it actually did was raise the stakes on the print, which then resolved in the bulls' favor.

3. The makers blink toward DDR5, and HBM shows its first crack

This is the two sided one, and I will not pretend it only cuts one way. Chosun Biz reported, citing industry sources, that SK Hynix is delaying the conversion of some HBM3E lines to HBM4, on the view that current capacity is enough to meet Nvidia's Rubin demand and there is no need to rush. The company is reportedly rotating resources back toward general purpose DDR5, where operating margins now rival or beat HBM and where the shortage is even more acute, to take profit Samsung has been making there.

The bullish read is that the tightness has spread from the apex all the way down to commodity DRAM, with general purpose margins reportedly heading toward 90 percent. The shortage is no longer just HBM, it is the whole stack.

The crack is the other half. The HBM leader saying it is adequately supplied and in no rush on HBM4 is the first dent I have seen in the sold out as far as the eye can see story that the bull case rests on. Treat it as a named rumor, not a disclosure, nobody at SK Hynix has confirmed it. But here is the important nuance for what I own: Micron is positioned differently. As the HBM challenger, not the leader, Micron cannot ease off the next gen transition the way a satisfied leader can, and its mix is already more weighted to conventional DRAM. On the call, Micron sounded like the hungry challenger pushing HBM4, which is the opposite of the SK Hynix signal and reassuring against it. Watch whether HBM tightness genuinely loosens or just rotates. So far it is rotating, which is on thesis.

4. The customers are now funding the supply, which is the real story

Three separate data points this fortnight all point the same way, and together they are the structural case made concrete. Micron's 100 billion of contracted backlog with floors. A report that SK Hynix customers are offering to buy its EUV machines and fund new fab lines outright as capacity hits zero. And Micron signing a multi year supply agreement with Anthropic across HBM, DRAM and SSD, tied to a strategic investment in Anthropic's Series H, where all three makers, Samsung, SK Hynix and Micron, came in as strategic investors at a 965 billion valuation.

When customers take equity stakes in their memory suppliers and offer to buy their lithography tools, memory is the binding constraint, full stop. A lab does not become a shareholder in its DRAM vendor unless allocation is the thing keeping it up at night. This is losing memory halts AI faster than losing Nvidia, showing up as corporate structure. The honest discount: the Anthropic deal is not exclusive to Micron, all three makers are in it, so it validates the sector more than it picks a winner. And for full transparency, since I should say it, Anthropic makes the AI I use to help research and draft these roundups, so I am flagging the conflict rather than burying it.

5. The trillion dollar capex bombshell, and why the stocks fell on it

On Monday, at a briefing chaired by Korea's president, Samsung and SK Hynix are set to unveil investment plans of up to 2,000 trillion won, about 1.3 trillion dollars, over ten years. Samsung's portion alone is reportedly around 646 billion across fabs, AI data centers, advanced packaging, batteries and displays. And the stocks fell on the news, Samsung down 4.7 percent, SK Hynix down 3.1 percent.

That reaction looks backwards until you see it. A massive capex plan is the supply response, and the supply response is the thing that eventually ends the pricing power this whole run is built on. The market read a trillion dollar spend as more capacity, more capacity as less scarcity, less scarcity as the toll booth closing. So a giant investment number made the producers fall.

Here is why the reaction is overdone on timing. Fabs are not a vegetable garden. A leading edge memory fab is a two year plus sequence, order EUV tools with a year plus lead time, build the cleanroom, install and qualify, then ramp yield. Micron said it plainly, new fabs deliver no meaningful output until fiscal 2028. So a number announced in June 2026 is a 2028 and 2029 supply event, and it cannot touch my window, which runs through 2027 and is already tight and already spoken for. The market sold a headline that does not change the next eighteen months of fundamentals. The supply response clock has visibly started, which matters, but it does not ring until 2028 plus, which my brief already flags as genuine uncertainty. The only near term supply variable that matters is how the makers rotate the capacity they already have, see item 3, not fabs that do not exist yet.

One more from the same week worth a line: Samsung also announced a buyback near 59 billion dollars. Read it with a discount, a chunk is mechanical, it owes employees a special bonus in stock and has to buy shares to cover the shortfall. Confidence signal, partly, obligation, partly. Sentiment tailwind into a nervous tape, helped the bounce, changed no fundamentals.

6. Apple is lobbying to buy blacklisted Chinese RAM

The single most thesis relevant China item to date, and the source is the Financial Times, not an aggregator. Apple is reportedly lobbying Washington for clearance to buy memory from CXMT. Apple just raised prices across its lineup, lost about 265 billion in market cap in a day, and Cook called it a hundred year flood he has not seen in 40 years.

Read it the right way, because it is bullish before it is bearish. Apple, the most powerful consumer hardware buyer on Earth, is begging to buy from a Chinese military designated company because the three makers cannot supply it enough memory at a price it can stomach. The lobbying is a symptom of how absolute the shortage is, not a solution to it. And the wall is holding, the House China committee chair called partnering with CXMT a grave mistake. CXMT sits on the 1260H military list, not the Entity List, so this is legally allowed but reputationally radioactive, which is why Apple wants political cover before it touches it. Pillar two, China blocked, is intact and currently enforced by politics rather than law.

The bearish seed, which I will name. A reputational wall can fall faster than a legal one, because there is no statute to repeal, only a political mood to shift. Apple lobbying at all is the first real sign that American corporate interest now wants the China wall lowered. It is not lowered. But the lobbying is the leading indicator to watch, and it is now on the board. My reassessment trigger sharpens: not Apple buying CXMT, but Washington granting any clearance or carve out, even China market only. That is the domino. We do not have it. We have Apple asking and Washington refusing, loudly.

7. The SK Hynix US listing got bigger and closer

Quick update to last fortnight. The ADR is moving forward, with reporting now pointing to a raise as large as roughly 29 billion dollars, materially bigger than the 14 billion first floated, though that larger figure is worth verifying as details firm up. The thesis is unchanged: it unlocks the wall of US money mandated to hold only US listed stocks, and hands American investors the actual HBM leader directly for the first time. Mild headwind to Micron's scarcity premium, bullish for the sector.

8. The pros caught up, and the froth got leveraged

Two sentiment markers I am reading as late stage, not bullish. First, IO Fund's Beth Kindig, a credentialed analyst, published a note saying essentially everything this newsletter has been saying, supply constrained until 2028, memory shifted from cyclical to structural bottleneck, AI systems getting more memory hungry each generation. Nice to be in good company. But three months ago a pro was not publishing this thesis on a retail platform next to an extremely bullish sentiment banner and a stock up 822 percent on the year. When the pros and the crowd all agree with your bullish call at all time highs, the trade is consensus, not contrarian. That is a marker that I am late, not wrong.

Second, Roundhill launched a 2X leveraged version of the DRAM ETF, ticker RAM, after the unleveraged fund pulled in more than 20 billion in assets and returned over 150 percent since its April launch. Honestly this is the most bearish thing I have seen in a while, and I hold the underlying. Leveraged single theme ETFs are a top of cycle product. They get launched when retail appetite for torque peaks, which is late, and the daily reset guarantees the buyers get ground up in any drawdown on an instrument that already swings 13 percent in a day. It is a sentiment signal, not a fundamental one. It says the crowd is euphoric and levered. It does not say demand is rolling. But it is exactly the kind of late stage marker worth respecting.

The real risks I am actually watching

Not the stock price, not scary headlines. The genuine ones, all on the demand side architecture, because that is the only thing that breaks this thesis structurally.

The named one this fortnight is Google's TurboQuant, which compresses key value cache data to cut memory needs, flagged by Kindig herself as a risk. That is the real category, memory efficiency, software that lets the same model run on less HBM. Her own rebuttal is fair, Google keeps adding HBM capacity in its latest chips anyway, so the efficiency gains are being eaten by bigger models, Jevons again. But KV cache compression is the vector to watch, and I am watching it.

Alongside it, SanDisk's High Bandwidth Flash, stacking NAND to reach HBM like bandwidth at far higher capacity, scaling toward 4TB versus HBM's 32 to 64GB per stack. It is the cleanest example of the substitution tail risk in my brief. But it is patent stage, it currently complements HBM rather than replacing it, the design keeps HBM on the interposer for the latency critical work, and it is a NAND maker talking its book. Watch the day a major accelerator ships with materially less HBM because flash took some of the load. That is not today.

And a non risk worth naming so you do not confuse it for one: the Tensordyne pitch making the rounds, claiming 17x tokens per watt over Blackwell, is a power efficiency story, not a memory efficiency one. Its own spec calls for large memory capacity. An efficient inference chip that serves more tokens needs more memory to feed it, not less. It is a Nvidia margin challenger, not a memory threat. Not my trigger.

One more rung I am building out: packaging. Intel hired the former CEO of SK Hynix to run advanced packaging and stood it up as its own unit pushing EMIB and HBI to volume, the same week Apple tied up with Intel on US manufacturing. The industry racing to build HBM packaging capacity beyond TSMC's CoWoS confirms packaging is where the bottleneck is migrating, the layer the whole HBM story runs through. Logic foundry story, does not move the three makers, but it tells you where the smart money sees the choke point.

Memory pricing update

The latest, and there is a real divergence worth sitting with. The stocks fell this week on the capex scare, but the actual price tape shows zero easing. Both things are true and they point at the same conclusion.

Contract. TrendForce's second quarter forecast is 58 to 63 percent quarter on quarter for conventional DRAM and 70 to 75 percent for NAND. Still enormous, but the read is deceleration, Q1 DRAM came in at a record 90 to 95 percent, so the rate of DRAM price growth is slowing even as NAND accelerates past it for the first time this cycle. PC DRAM contract momentum is cooling as high prices choke PC sales, with notebook shipments now guided down about 15 percent for 2026, while server DRAM keeps rising on depleted inventories and stays guided up into Q3 and Q4. Meaningful new capacity does not arrive until the second half of 2027 at the earliest.

Spot. Still firm to rising as of the last few days, no rollover. DDR5 spot is holding up with buyers accepting higher quotes, and DDR4 remains inverted, trading at or above DDR5 per gigabyte in some configs because it is being end of lifed with no new production. The cheapest 32GB DDR5 kit in the US has been tracked around 375 dollars, with daily volatility and outlier listings far higher. That DDR4 inversion persisting is the tell that the tightness is structural, not speculative.

Net: still a shortage, still guided up, DRAM growth decelerating off the Q1 peak, NAND carrying the baton, PC versus server split hardened. And critically, the physical price tape shows no easing at all, which means the trillion dollar capex scare that knocked the stocks is a 2028 story the actual prices have not begun to reflect. Sentiment flinched. Fundamentals did not.

Where that leaves me

The biggest fortnight of the cycle, and the thesis came out the other side stronger, not weaker. Micron's print gave the structural case its first hard balance sheet evidence. Customers are funding capacity and taking equity in their suppliers. The Korea crash was positioning and reversed in two days. The trillion dollar capex scare is a 2028 event that cannot touch the through 2027 tightness. Apple lobbying for blacklisted Chinese RAM is the shortage confirmed in the most dramatic way possible, with the political wall still standing. And the pricing tape shows no easing.

The two honest cautions I am carrying. One, the positioning is now crowded and consensus, the pros publish my thesis, retail is extremely bullish, and there is a 2X ETF, all late stage tells that say I am late even if I am right. Two, the real risks are the efficiency ones, TurboQuant and HBF, and the China wall, which is reputational and can move in a headline. Those are what I watch, not the candles.

My tripwire was never the stock price, a scary headline, or a China announcement. It is hyperscaler capex guidance turning down, and across the biggest two weeks of news this cycle, it did not. So I am holding.

I hold a long position in DRAM ETF (ticker: DRAM) and Micron (ticker: MU). Not investment advice. Do your own research.