The rerate case doesn't need LTAs to defend peak margins. It needs the mid-cycle price level to be structurally higher than before AI, and floor prices, customer-funded capacity and HBM eating wafers all point that way.
I agree LTAs won't protect 80% margins once demand turns. The onsemi and Microchip cases show suppliers will choose the customer over the contract. But a pushout moves earnings in time. It doesn't remove them. In the old cycle a downturn meant canceled orders, inventory dumped into spot, and losses. If the next one looks like volumes slipping a year at a floor price, the trough becomes a slow year instead of a loss year.
At 3 to 4x peak earnings the market is pricing a classic loss-making trough. That's the gap. The deposits are already in hand.
Where I'd get worried is if pushouts come with price cuts down to market. That lowers the level, not just the timing.
I'd like to believe LTAs are durable too. But take one example.
Analog chipmaker onsemi said this in 2023.
On January 5, 2023, at JPMorgan's CES forum, CEO Hassane El-Khoury was asked how much LTSAs would still hold once the shortage eased. His answer: “They are a legally binding agreement ’cause they’re a liability for the customer because they’re take or pay.” He also said he would not accept a customer reopening price talks just because it could get the same parts 10% cheaper elsewhere.
On October 30, 2023, on the Q3 earnings call, onsemi cut its SiC shipment forecast for the year from $1 billion to more than $800 million after one automotive OEM reduced demand. An analyst asked how this could happen when the company had been saying its LTSAs were legally binding. The same CEO reaffirmed that they were binding, but said he did not want to enforce the contracts just to push inventory onto a customer whose demand had dropped. If it worked for both sides, terms could be adjusted customer by customer, whether by a quarter, a year, or longer.
In its 10-K filings, onsemi also states that it amended LTSAs with some strategic customers by mutual agreement in 2024, and changed delivery schedules and committed volumes on some LTSAs in 2025 as demand shifted.
Microchip is another case. Customers in its Preferred Supply Program (PSP) originally had to place 12 months of orders and could not cancel or reschedule them unless Microchip raised prices. On its November 2025 earnings call, however, the CEO said the company had accepted some cancellations and pushouts so customers could adjust their backlog. On the same call, the CFO separately addressed the multiyear LTSAs the company had signed. Some of those contracts were still in effect, he said, but customers had been allowed to defer taking delivery of committed volumes by one to two years if they needed to.
Every time I see cases like these, I doubt how much memory makers' LTAs will protect earnings in a downcycle. Right now the most reliable thing is the prepayments and deposits themselves. That cash can go into capacity expansion today.
I can picture hyperscalers with no DRAM inventory continuing to buy at current prices. But once inventory has piled up, will they keep taking every committed unit at the same price? I have a hard time picturing memory makers holding customers to the high contract price all the way through.
The LTA itself may never be broken and may stay in place. But if both sides agree to push this year's volumes into next year or two years out, this year's earnings estimates come down. If price gets renegotiated too, margin estimates have to come down as well.
Again, I agree LTAs are sturdier than they used to be. What I question is whether, after demand turns, memory makers can still earn 80% margins on contracts signed during the upturn and keep collecting hundreds of billions of dollars in profit because of those LTAs.
1
183
When the market believes four contradictory reasons at once, the reasons aren't the reason.
Software gets replaced by AI, don't buy software. Hardware fundamentals are great but unsustainable, don't buy hardware. Hyperscalers are about to issue equity, don't buy hyperscalers. Neoclouds can't borrow, don't buy neoclouds. Every layer of the stack has its own sell case. The conclusion is always sell.
These four don't hold together. Software being replaced requires AI to actually work. If AI works, you need the hardware. If hardware demand is real, someone has to fund the datacenters. If AI doesn't work, the first two reasons vanish. If it does, the last two are about build speed, not demand. At least two of the four cancel each other out.
Believing all four at once has one explanation: the year's P&L is already made, risk needs to come down, but "I'm selling because I've made enough" isn't a professional thesis. So each layer gets its own fundamental story. One person, one action, four explanations.
This explains October. Good earnings sold off because earnings weren't the reason for selling. Bad news hit hard and good news didn't because sellers were looking for reasons to sell. Net exposure cut to a five-year low while gross stayed high because what got cut was this year's performance risk, not the view on the industry.
This state has an expiry date. January 1 locks the year. The "made enough" motive disappears and only the parts with actual data survive. The debate narrows from "is the cycle ending" to "how much does 2028 decline." It's already happening. This week not a single sell-side model disagrees on 2027.
One layer deeper: the market always moves the disagreement to wherever you can't verify it. You win 2026, it argues 2027. You win 2027, it argues 2028. The prize for winning isn't a re-rating, it's the next exam starting early. Earnings multiply and the multiple compresses, because the multiple prices the probability that this is the peak, and the better earnings look, the more they look like a peak.
So the edge isn't knowing fundamentals better. 2027 being strong is public information. The edge is holding through two years of multiple compression while the argument changes shape again and again.
Marginal buyers price short-term volatility and they do it well. They just don't price the long term, because nobody pays them to. That bill you pay yourself, in patience.
3
157
At 6x forward, $MU is priced for earnings to fall by more than half after FY28. That requires either 2028 supply arriving a year early or SCAs being torn up at scale, and neither is visible before late 2028. The stock will not re-rate on good fundamentals; it will move on two things: rate volatility collapsing, and the answer on 2028 supply. The first is a Q4 macro question. The second starts with Korean earnings in late October and resolves with next year's capacity data. Until then, hold it, let buybacks retire roughly 10% of the share count a year and do the work on per-share value, and don't pay 50x for derivatives of the same thesis. De-speccing and land/power/shell bottlenecks are real risks, but they cap the amplitude, not the direction, and they are already in the 6x.
$SPY How Rate Volatility Compresses Multiples: The Transmission Chain and Where We Are
The 10-year Treasury yield has risen from 4.17% to 5.27% this year. It reaches equity prices through two channels.
The first is the valuation benchmark. The 10-year now yields 5.27%, roughly 19x in P/E terms. The S&P 500 forward P/E also sits at 19x, so the index earns the same yield as Treasuries while carrying more risk. Every company is being re-examined on one question: where is the excess return over Treasuries. Companies dependent on far-out earnings have the hardest time answering. Companies with current cash flow and buybacks have the easiest.
The second channel is positioning, and it carries more weight. Fund risk limits are set on volatility. When rate volatility rises, equity volatility follows, the same position breaches its limit, and risk management forces de-grossing. No judgment is involved. The first positions cut are the most crowded and highest beta. Multiple funds sell the same names, prices fall, remaining longs lose, limits tighten further, selling continues. The loop reinforces itself.
The result is a disconnect between price and fundamentals. This earnings season, many companies posted strong results with estimates revised up, and prices did not respond. The market does not doubt the earnings. No participant is in a position to add. Short-term prices are set by forced sellers, not by whoever has the right call.
We are in the later stage of the loop. US L/S fund net leverage has fallen three weeks straight to the 2nd percentile over five years. Most of what could be cut has been cut. Index-level shorts have started to cover, core inflation and payrolls came in soft, and hike expectations have been pushed out. But rates themselves have not eased. The long end rose 11bp on the week, credit is starting to show strain, and the Iran talks and midterms remain unresolved.
The loop has two exits. One is exhaustion: positioning reaches a floor and selling pressure fades on its own. That condition is close. The other is rate volatility falling, risk limits loosening, shorts covering and longs rebuilding. That condition has not appeared. How it ends is unclear, but downside is now smaller than upside.
Two takeaways. The market is pricing rate uncertainty, not the rate level. And a price set by forced sellers carries no information about the company itself.
Personal notes. Not investment advice.
1
152
$MU reports FQ4 on 9/30 after close.
The quarter will most likely beat, driven almost entirely by pricing with bits roughly flat. GM should land around 87%. Consensus for the FQ1 revenue guide is near $58B.
That said, most semis have beaten this earnings season and traded flat on it. Over half of MU's revenue is now under long-term agreements, roughly 40% with price floors. Upside sensitivity has narrowed; downside is better protected.
The first-half narrative was EPS revisions plus multiple expansion. After July's drawdown the market shelved the second part. Estimates are still rising, which on its own supports the stock, but multiple expansion needs several quarters without bad news to be validated, not one print. The debate has shifted to duration.
The main thing to watch is HBM. 2027 contracts are wrapping up and take effect in January. Over the past year conventional DRAM pricing ran far ahead of HBM, so HBM's share of DRAM revenue and its margins fell relative to the rest of the portfolio, leaving room for catch-up. Industry expectations sit around +50-60%; some channels say part of the negotiations are done at +80% to 2x. Management typically won't give a number, so it comes down to wording. New pricing doesn't hit until January, so nothing shows up in this quarter's numbers.
Also worth listening for: FY27 capex and the timeline for new capacity, any additions to the long-term agreements, and whether contract ceilings have been adjusted. Buyback restrictions lift in December, so expect only a brief mention this time.
PCE and NFP land in the week after the print, so the stock reaction will carry some macro noise.
Overall this looks like a confirmation quarter. Clarity on HBM pricing should come by late October, and on buybacks in December.
1
1
244
$SNDK If true, NAND Q4/1Q27 go from flat to double digits and every deceleration model gets rewritten. Hyperscaler eSSD is doing the lifting while consumer fades. The plateau just got pushed out two quarters.
Summary of : NAND Pricing Breakout Driven by eSSD
☕️ According to TrendForce, the leading third‑party memory research firm, NAND demand is tracking materially stronger than expected.
📈NAND demand tracking much stronger than expected — driven by eSSD as U.S. CSPs aggressively scale AI infrastructure
• WE RAISING NAND blended price outlook meaningfully:
3Q26 / 4Q26 / 1Q27 QoQ: +21% / +17% / +15%
(vs prior +18% / +2% / 0%)
• Inventory at U.S. hyperscalers expected to fall further by 4Q26 (from ~14 weeks in July vs ~15-week normal)
• Supply remains disciplined: bit growth +15% / +21% / +20% YoY in 2025/26/27
• eSSD demand from U.S. hyperscalers has nearly doubled vs prior projections → demand growth outpacing supply near quarters
• enterprise (especially AI-related eSSD) px/demand far outpacing consumer
• AI is reshaping the memory hierarchy → new stack: DRAM → SCM → TLC/QLC NAND → HDD
• Still bullish DRAM/HBM (SK Hynix, MU) and Now also constructive on NAND names (Samsung, SK Hynix, MU, SanDisk, Phison, etc)
#Trendforce
#Memory #DRAM #NAND #MU #SKHY #SNDK
1
317
Narrative Seeker retweeted
Summary of : NAND Pricing Breakout Driven by eSSD
☕️ According to TrendForce, the leading third‑party memory research firm, NAND demand is tracking materially stronger than expected.
📈NAND demand tracking much stronger than expected — driven by eSSD as U.S. CSPs aggressively scale AI infrastructure
• WE RAISING NAND blended price outlook meaningfully:
3Q26 / 4Q26 / 1Q27 QoQ: +21% / +17% / +15%
(vs prior +18% / +2% / 0%)
• Inventory at U.S. hyperscalers expected to fall further by 4Q26 (from ~14 weeks in July vs ~15-week normal)
• Supply remains disciplined: bit growth +15% / +21% / +20% YoY in 2025/26/27
• eSSD demand from U.S. hyperscalers has nearly doubled vs prior projections → demand growth outpacing supply near quarters
• enterprise (especially AI-related eSSD) px/demand far outpacing consumer
• AI is reshaping the memory hierarchy → new stack: DRAM → SCM → TLC/QLC NAND → HDD
• Still bullish DRAM/HBM (SK Hynix, MU) and Now also constructive on NAND names (Samsung, SK Hynix, MU, SanDisk, Phison, etc)
#Trendforce
#Memory #DRAM #NAND #MU #SKHY #SNDK
$AVGO FQ3: One Chip, Two Buyers, Three Discounts
On 9/2 Hock Tan gave two-year numbers for the first time: FY27 AI $115B, FY28 $230B, doubling twice, FY28 EPS >$30, supply locked, no more quarterly updates. The $100B nobody believed is finally gone, but $115B came in below a lot of buy-side expectations.
Customer mix flipped:
- Anthropic becomes the largest customer next year (FY27 5GW, FY28 10GW)
- OpenAI second the year after (>5GW)
- $GOOGL drops to third, GW flat
GM went from 77% to a 73% Q4 guide. Management wants you to look at 66% op margin instead. The $20-30B/GW mentioned on the call was walked back on the sell-side callback after clients pushed for clarification. Misstatement. Guidance implies ~$11-12B/GW.
Splitting $230B by chip is more honest than by customer: roughly 70% is TPU, part bought by Google with cash, part bought by Anthropic with borrowed money. The other 30% comes from Jalapeño and programs that haven't scaled. Six customers sounds diversified. It's one proven chip sold to two buyers, plus a pile of things on paper.
Three risks driving the discount:
- Google share: MediaTek/MRVL/AMD eating into side programs. Mainline locked through 2031, but even the most bullish sell-side conceded down to two-thirds
- Anthropic credit: chip works, money isn't closed. XPV platform with Blackstone and Apollo targets 20GW by 2028. First-tranche residual guarantee exposure is $29B against a $35B first 1GW deployment. CFO declined to cap AVGO's total guarantee
- OpenAI/Other clients execution: zero track record scaling an ASIC
At $357, FY27 is fully priced, ~19x. FY28 looks like 12x but that's taking management's $30 at face value. The market discounted all three risks and only credits $130-150B of AI revenue, EPS $18-20, so you're back to 18-20x, same as FY27. The 12x is an illusion. The market is saying no growth for FY28.
Every sell-side kept Buy. Some cut the numerator, some cut the denominator, some changed methodology to hold the target. Same underlying call: the numbers are credible, but nobody will prepay for two labs' balance sheets.
The AVGO story is no longer about EPS beats. It's about any one of three discounts narrowing on evidence. The biggest is Anthropic's credit, and that doesn't get tested until 1H27.
Until then, cheap is real. So is the reason it's cheap.
1
180
$SNDK : what the market is still arguing about
Buybacks. Mgmt committed 100% of excess cash to repurchases at 50% FCF margin. That's $25B next year, 10% of market cap. Four years at current price retires 47% of shares. For scale: S&P avg buyback yield is 2-3%, Apple at peak was 3-4%. No $200B+ company has ever done this. Yet almost no analysts model declining share count in their PT. Every model locks 148M shares. Universal consensus positive, zero in valuation. Each quarter $5-6B gets executed, each quarter shows up as an "EPS beat," each quarter the market acts surprised. No new narrative needed, no multiple expansion needed. The denominator shrinks on its own.
Customer concentration. Eight customers cover two-thirds of bits, three are US hyperscalers. Everyone reads this as "locked in." Some wrote "does create risks" and moved on. It's both: locked in and exposed. Any hyperscaler shifting capex language and the market immediately asks about the contracts, no NAND price decline required. Take-or-pay exists, $16.5B guarantees exist, but guarantee coverage is 18%. In a deep downturn the likely outcome is renegotiation not default. This risk is priced at zero.
HBF. First die taped out, samples 2027, Meta joined the consortium. Not in the model, pure option. The math: HBF die holds a quarter of TLC capacity at similar area, so 4x wafer per exabyte. Cost ~$0.19/GB, pricing $1.25-2.50/GB, implied GM north of core NAND's 84.6%. Volume ramp means demand up, margin up, supply tighter, all at once. But this is post-2028. Market applies a time discount and that's correct.
Buybacks are the mispricing. Concentration is the ignored risk. HBF the market has right, just early.
1
5
445
The rotation you are reading about already happened. PB data shows memory got cut hardest in July, and in the first week of August money went back into semis, networking and optics, while memory saw zero re-gross.
The “educational” threads about pod constraints and positioning showed up after the positioning was done. The trade got built first, the logic got popularized after.
Why is he explaining it to you now?
Imagine yourself as a pod analyst with weeks to maximum 6 months for a position to yield alpha: in order to deal with memory underperformance, either you rotate into sectors with better narrative / catalyst support, or you hold like a LO investor and get fired.
L/S analysts can be bullish for the medium and long term but still be selling because of duration..
Therefore, a lot of outsized fluctuations are better explained by positioning these days, and the rotation has already begun early August when @SemiAnalysis_ posted the Rubin HBM note.
Would expect players like $SMTC $TSEM to outperform in this environment as performance loss are compensated with higher NPO content.
Narrative Seeker retweeted
Replying to @DratchCap
Agree on all four. One more thing that hasn't been priced: HBM 2027 annual contracts are being negotiated right now, suppliers targeting +50-100% pricing on top of +50-60% bit growth, a product line more than doubling revenue next year. Contracts sign at year-end, so none of it shows up in monthly price trackers or any company guide yet. The market is busy calculating second derivatives on quarterly spot while the largest single repricing event of the cycle sits on the calendar, unprinted. Same mispricing as the LTA consideration, nobody's paying for what's already been exchanged.
Narrative Seeker retweeted
I think the mkt is confusing HBM per GPU with aggregate HBM demand. Reducing HBM for a certain SKU just spreads that bit-demand across more revenue generating systems. And at the system level, HBM content is going up due to scale-up / out networking.
On that latter pt, optics and memory are part of the same reflexive demand loop. Betterr optics (CPO etc) enables larger connected systems, which means more memory-bearing endpoints. More endpoints and more memory traffic pull in more optics. And round and round we go...
There is of course a difference in the expression between the two. Memory supply is tighter / more oligopolistic, multiples are way lower, expectations totally reset (sad), monster cash flow is here *now*. Optical stocks attract the fast money bc they are smaller (easier to double on account of less mkt cap), but competition is greater / capacity easier to add, multiples much higher and cash flow is further out. I like some optical to be clear, and it's not either/or imo, but we need to be intellectually honest about the difference.
You also can't have it both ways on LTAs. If they cap spot upside, they also lock in volume, px floors, duration, and pre-payments (!). I know the consensus playbook is sell the cyclical before prices and margins peak. Usually this is right.
But when the latter is happening bc the companies are choosing partnership mentality / pre-payments / duration, they shouldn't also trade at cyclical multiples (~2x eps, 1.3x fwd book) with 30% FCF yields. Frankly, I'm having scary flashbacks to my distressed / levered equity days just writing that, and yet the secular dynamics could not be more different.
Anyway, the mkt PAYS YOU to TAKE PAIN. And now it definitely owes me some money ;)
I think the market ultimately has no choice but to go sell memory, long optical in the "short term." Actually, some hedge funds already seem to have this position on.
There are three main reasons.
1. With Korean leveraged ETFs effectively dead, LPs are in a redemption rush, which could bring out additional sell on flow.
2. Nvidia is nerfing Rubin Ultra's HBM and responding with optics, tying multiple racks together, so that even if Rubin Ultra's per rack performance is not superior to Rubin, at the cluster level optics let the Rubin Ultra cluster hold an edge over the Rubin cluster. This holds even if Rubin Ultra's HBM nerf is a supply problem rather than a demand problem.
3. Consensus is forming that memory prices will peak within the next two quarters.
Medium to long term I am still a memory bull, but short term I am somewhat bearish on memory. I currently have no memory position.
$MU $SNDK $SKHY
The Three-Stage Earnings Engine in Memory
The growth-vs-cyclical debate on memory stocks carries a hidden assumption: earnings have peaked. Break down what’s actually driving earnings, and the current price surge is only stage one of three.
Stage one is pure pricing. Since 2H25, volumes barely moved while prices rose 4-5x from trough. The market pays 5-7x PE, which is its way of saying these earnings are temporary. The slope here is indeed flattening.
Stage two is volume times high prices. MU’s new fabs ramp through 2027-28 into a market where the deficit is widening, and all three major suppliers have already sold out their 2027 capacity. In a traditional cycle, adding volume means crashing prices. This time volume is coming and prices haven’t budged.
Supply up, prices flat sounds absurd, but it’s happening right now: capacity expanding, contract prices rising, 2027 output fully booked, prices still making new highs. The reason: demand is growing faster than supply, and LTAs have locked existing supply into high price bands. Earnings can keep expanding without price acceleration. “Slowing price hikes = peak earnings” breaks down here.
Stage three is the resonance of multiples and capital structure. Cyclicals get low multiples for exactly one reason: money earned in the upturn gets handed back in the downturn’s price war. LTAs change that expectation. When contracts cover two-thirds of shipments and trough-year EPS still exceeds the prior cycle’s peak, there’s no basis left for 5-7x. Meanwhile SNDK alone generates ~$70B of FCF over two years, enough buyback to retire a quarter of the share count. Even if total profits flatline, per-share earnings keep rising through the denominator.
No gap between stages. The quarter stage one decelerates, stage two’s volume takes over, and stage three’s buybacks and multiple migration run throughout. The market is pricing stage one as the whole story, and giving it one year.
One distinction matters: NAND and DRAM are different stories. NAND has zero greenfield capacity, grows bits through node migration, and for the first time ever grows slower than DRAM. Its path is volume locked, price floored, re-rated slowly on cash flow certainty. DRAM has real fabs under construction where volume and price move together. Lumping them into one “memory” debate blurs the picture.
Promises and proof are different things. In 2022-23 the HDD industry went through its deepest downturn ever, and the duopoly chose production cuts over price wars. Only after that live test did the multiples truly move. Memory’s test likely comes in some quarter of 2027 when pricing flattens. Before it, the debate won’t end. After it, the debate won’t be needed.
2
577
SNDK's gross margin plateau is a welcome move. It means they are not squeezing customers to the extreme, and that they are looking at a long term vision instead. The disciplined, balanced capex execution I have been arguing for all along, the kind that prevents a company from investing to extremes at the top of the cycle and taking severe damage in the downcycle, requires SanDisk's NBM model. There is no way around it.
So is now the right time to buy SanDisk? The bears argue that the gross margin plateau marks the peak of this cycle. That is also why the stock is falling right now.
As I wrote in my SK Hynix earnings review(You can find it in the Citrini Substack chat), this is the point where we should be looking at Q, not P. As long as you believe in the AI supercycle, as long as big tech keeps spending, as long as compute companies like Nvidia remain in short supply, there is no scenario where the shortage eases for memory alone. That is the corresponding effect of LTAs (NBM) and disciplined capex. Even if gross margins stall, as long as Q expands and the floor set by those LTAs prevents an excessive collapse, the total profit pool will grow enormously.
Layer massive shareholder returns such as buybacks on top of that, and memory companies should be rerated, and they will be rerated. From cyclical stocks to essential infrastructure of the AI world.
$SNDK
$SKHY Shawn Kim’s formal reaction to SK hynix’s record quarter, published today: Overweight, PT ₩2.6M (+68%), thesis unchanged, inventories “continue to DECLINE across servers, PC, mobile and NAND,” ~10 LTAs “materially improve profit and FCF visibility,” and the stock at 4.9x 2026 / 3.4x 2027 is “assuming a cyclical downturn, still undervaluing.” His own model has EPS rising through 2028. This from the same analyst whose bearish “NDR note” — 13 weeks of NAND inventory, “demand has evaporated,” sell DRAM too — recently tanked Korean memory stocks.
Same analyst, same firm — the bear note that moved a market exists only in screenshots; the published research says the opposite. Verbal NDR color stripped of its qualifiers isn’t research, it’s ammunition. Price what analysts publish, not what the telephone game delivers.
$SKHY
A note reportedly from Morgan Stanley’s Shawn Kim is circulating as a bearish call on memory. Having read his formal July 2 report, the points being attributed to him — module maker inventory buildup, consumer NAND pricing pressure, distributor channel stress — are already in that published note, word for word. But in the original, they sit inside a balanced framework: supplier inventory at historical lows, server demand still tight, distributors themselves saying demand isn’t fundamentally weakening.
What happened is textbook transmission loss. An NDR where an analyst discusses risk factors gets compressed into “Kim turned bearish.” The qualifiers disappear, the server-side counterpoints disappear, and “consumer pricing might hit a ceiling” becomes “demand has evaporated.” The formal report says dual-speed market. The whisper chain says sell.
Worth noting: his formal ratings haven’t changed, no price target has been revised, and no downgrade has been issued. If he had real conviction in a structural turn, the move is a published report with his name on it — not an NDR that leaks sideways.
The absence of a formal call is itself information.
The actual insight in his work is useful: consumer-grade NAND is hitting an absorption ceiling after three quarters of aggressive price hikes, while server/AI demand remains supply-constrained. That’s a channel-level observation about pricing friction, not a demand collapse thesis. Conflating the two is how noise gets manufactured.
8
808










