Semiconductor capex super-cycle (new) Contested — near-term capacity sold out, but Chinese memory now threatens the "shortage through 2030" consensus before the 2028 supply tsunami

Shaulov (Sohn, May 12): after a decade of capex discipline the chip complex flips from "one spender to multiple spenders, all of which underspent" — Taiwan Semi "underspent significantly" (and "made a mistake… they will have to rectify"), while Intel's foundry is now "picking up customers" and Samsung is announcing. Forward-capex indicators (customer profitability) flash, wafer-fab equipment (WFE) goes from a perceived ~$120-130B toward ~$300B over 3-4 years, and because customers run 70-80% margins the equipment makers have pricing power on top of unit growth — "the estimates are 50 to 70% too low," LTAs make the names less cyclical (a multiple re-rate), and "the next 50 to 100 is up." His picks: Lam Research (memory-levered, his #1), the memory complex (Micron, SK Hynix, SanDisk — a NAND cycle "we haven't talked about in a decade," ~80% margins), and analog (Texas Instruments, Renesas, which "could look like memory" on pricing). Sacerdote (same panel): the complementary "decommoditization" / "golden age of hardware" leg — AI workloads (tokens ~14× a year vs ~2× chip efficiency, so demand outruns efficiency) push once-commodity parts (PCBs 10→120 layers, networking 1→3,200 Gbps, power +50-125%/rack) into IP-rich, high-margin pinch points where only "two or three companies" qualify (TTMI); units +50%, ASPs +20-100%, margins +300-500bps, 3-4-yr visibility — "the moves we're seeing are justified," all earnings. App Economy (Jun 26, Micron): the memory leg is now visible in the numbers — Micron's fiscal-Q3 revenue +346% to $41.5B at an 85% gross margin, almost all price (DRAM +60% in a quarter) not volume. The structural twist is 16 take-or-pay Strategic Customer Agreements (~$100B minimum, ~$22B of customer deposits, running through 2030) that set a margin floor above the prior ~60% cyclical peak — an attempt to turn memory's boom-bust into a floor; CEO Mehrotra sees "no line of sight" to supply catching up, pushing the crunch past 2027. Early ceiling signal: the Q4 margin step-up shrank to ~+1.4pp as SK Hynix adds supply. Halftime committee (Jun 25): the secular-vs-cyclical debate went live the day after the print — Terranova calls memory "an irreplaceable resource," not a commodity (no price elasticity; 16 Strategic Customer Agreements lock pricing to 2029), Link notes DRAM ASPs +60% / NAND +80% forcing Apple into price hikes, Lebenthal buys it as a cheap (~7× fwd) cyclical in the middle innings with big buybacks coming, while Brown's bear case is it's all price not volume — extreme pricing power eventually makes customers engineer around it (the DeepSeek precedent); Micron now a ~$1.3T company (9th-largest US stock) "done on raising prices." Eisman (Jun 26): the print in numbers — EPS +1,215% / revenue +345% YoY, ~8.7× 2027 PE, supply "constrained past 2027"; SanDisk +200% the quarter. Singh (Jun 28): Micron out-earned every Mag 7 name (85% gross margin; a ~$50B next-quarter guide) and the 16 five-year take-or-pay agreements (~$22B of customer deposits, floor pricing above any prior peak) cover >half of revenue — but the cost is now passing through to consumers (Apple +20–28%, Xbox +$100–150) and forcing ~15–20% US auto-production cuts in H2-26; the bear tell is Chinese funds calling memory "no moat." Carlson (Jun 26): the bear tell inside the blowout — Micron's fiscal-Q3 (~$41B revenue, +346% YoY, ~81% gross margin, growing faster than Nvidia at its peak) is more than 90% price, not volume, so the boom rests on pricing power that is now becoming the consumer economy's inflation problem rather than durable unit demand. Halftime committee (Jun 29): Lebenthal bought Micron as a ~7x-forward 'value stock' with rising estimates — the Citi/Adam-Parker 'not owning it is like missing Nvidia in 2023' analogy; 'this isn't different, it's cyclical' but 'middle innings.' Terranova — memory is 'more secular than cyclical,' the strategic customer agreements validating an extended cycle; play the derivatives off the parabola — toolmakers KLAC/LRCX/AMAT and the ~$17B alpha-capture name ONTO. Link — SOXX +85% YTD with momentum intact; she owns only Marvell ($10 EPS power by 2027) and Broadcom (~$20 EPS power, the laggard she's adding). Singh (Jul 5): a violent memory selloff (SK Hynix −30% in a week, Goldman's high-beta momentum index −18%, its worst since COVID) turned Micron's historic beat-and-raise into a sell-the-news to ~$992 — but Samsung then hiked Q3 DRAM another 20% and UBS extended the DRAM/NAND shortage to 2028, framing the pullback as "positioning, not fundamentals"; China's CXMT is ASML-blocked and generations behind (its memory isn't even cheap), so it's a long-term structural competitor, not a near-term cycle-killer. App Economy (Jul 7): SK Hynix's ~$28B Nasdaq ADR (SKHY, ~Jul 10) — the largest ADR listing ever — brings the #1 HBM maker (~56% share, ~2/3 of NVIDIA's HBM4, sold out through 2028, record 72% operating margin in Q1 FY26 on revenue +198%) to US investors at parity with Micron (~7× fwd); every dollar goes to capacity (Yongin, HBM packaging, ASML EUV, a $4B Indiana plant). Framing: peak-cycle margins treated as a new baseline — bull case "AI broke the memory cycle" (multi-year contracts, structural demand), bear case "memory always reverts" (Samsung back in HBM4 at a projected ~28% of the 2026 market, China grinding up from the low end); the author would watch the ADR trade a quarter or two rather than pay up for permanence. Halftime (Jul 10, listing day): SKHY opened +16% over the $149 list (~$173) on the biggest-ever US ADR raise ($26.5B) — Partsinevelos's frame: "bigger, cheaper, closer to NVIDIA than Micron" (56–58% HBM share, the #1 NVIDIA supplier); the committee's memory split stayed live — Baruch keeps MU as his largest position (sold out through 2026, sell at a 20× rerating; support 850–900), Weiss took his profit (would rebuy ~925; "you'll never see 20× on a Micron — 10 is the historical peak, all roads lead to Taiwan Semi"), Harrington calls memory "tradeable, not investable" (cash flows too ambiguous; one efficiency-software breakthrough could break it); SMH context: only 4 of the past ~22 sessions moved less than ±1.5%, and JPM argues the 15% pullback sets up outperformance. Singh (Jul 12): the cycle isn't peaking — Samsung printed +19x profit ("40 years of profit in one year") yet fell 14% because "so much good news is priced in"; SK Hynix's NY listing (178M ADRs, $25B, 7x oversubscribed) debuted +14%, TrendForce lifted Q3 DRAM ASPs another +13–18%, BofA sees shortages into 2027 and Meta's memory orders are "stronger than ever" — he bought Micron in the flush; the caveat is Citi's consumer demand-destruction warning (PCs −15% / phones −12%). Pabrai (jun-30, rec. jun-8): memory is now a protected three-player oligopoly (Samsung, SK Hynix, Micron) — patents plus "black magic" fab know-how make a fourth entrant near-impossible (Micron told him a burnt-down fab might not be rebuildable at the same cost/yield), the three can't meet demand and chips are on allocation. But hold, don't chase: "if you already own it, don't sell it; if you don't own it, don't buy it — the party has only just started." He regrets selling his own SK Hynix and Micron. Finucane (Jul 16/17): the SOX's +230% in 14 months has one precedent — the final 14 months into March 2000; the 12–15 highest-quality semis screen to ~40% downside in a mere normal bear market vs his 20% max-downside entry rule, semis moved >5% on half of June's sessions ("do you actually think the value of these businesses changes 5% every other day?"), and an "accounting arbitrage" (100% of hyperscaler capex booked day-one as chip revenue vs slow depreciation on the buyer's books) has let the semis take the hyperscalers' free cash flow — the smartest management teams in the world won't allow that forever. Micron as the cautionary math: a bullish analyst's own numbers (EPS $250 in 2028 → $50 by 2030) imply ~$400 at a normal 8× vs a ~$1,200 peak — a two-thirds base-case drop. Singh (SSR, Jul 19): the memory peak call is contested — Morgan Stanley says memory pricing likely peaks, but BlueFin's opposing bull view sees HBM4 pricing doubling/tripling in 2027 (SK Hynix's HBM4-ramp delay read as a "veiled threat" to Nvidia margins); Micron guided +15% and sees the supply imbalance persisting beyond 2027, and Singh sides with the bulls because the data centers driving demand aren't even built yet. TSM and ASML both beat and raised yet sold off on capex fears; SNDK/MU tagged as short-term buys in the flush. Singh (Jul 26): Morgan Stanley forecasts 25% memory price increases on AI demand; Micron's HBM is sold out through 2027. BofA refutes the open-source-AI bear case on memory: closed models amortize demand across shared HBM pools, but 10,000 enterprises self-hosting the same open model replicate the weights across 10,000 separate pools — and at 1M-token contexts the KV cache alone can exceed 40 GB per session → more DRAM sockets worldwide. No cheap alternative: CXMT's 64GB DDR5 server module runs ~$1,240, and a Micron/Meta white paper shows a 38× slowdown once shuffle data spills from DRAM to SSD. Positioning washed out then turned: hedge funds cut momentum/semi longs ~5% of gross market value (one of the largest on record), reversed 80% of net buying since mid-June, and began re-buying last week. INTC guided Q3 to $15.8–16.8B vs $15.1B est on data-center CPU sales +59%; AMD signed a tens-of-billions AI-server deal with Anthropic; STM −17% as the non-AI analog/auto laggard. Hay (Haymaker Daily, Jul 29 — "Taking a Big Knife to the Semis"): the KOSPI blow-off's engine — "essentially two stocks: Samsung and Hynix" — has plunged 40% in five weeks despite ballistic earnings and modest P/Es, the same five weeks the SOXX began a descent "increasingly sickening for its formerly jubilant holders." MU lost a third of its market cap since Jun 24 despite blow-out earnings — cheapness on peak-cycle earnings is no floor once a mania unwinds; "no slice of the S&P 500/NASDAQ was more adored than the semis." Tactically: a bounce is "possible, even probable" and would be "an opportunity to once again reduce exposure… the bigger the bubble, the bigger the bust." The counterweight — Carlson (Jul 31): he keeps ASML as his one semiconductor leg but refuses to concentrate in the category: semis are "commodity-like… scarcity-driven," and much of the volume is "one-time scarcity-driven volume" supplied to the big capex aggregators — order books that are pull-forward, not run-rate. Luria (Eisman Ep 70, Jul 27) hunts the dislocation of two prices implying opposite cycle lengths: MU at ~6× earnings "as if the cycle is already over" vs INTC at ~100× (and Cerebras) priced "as if this cycle is continuing through 2030" — and today "the memory chip market is much better than the CPU market." Ives: NVDA is "their world, everyone else paying rent," a third-rate NVDA chip 18–24 months ahead of Huawei, ~$8–10 of spend across the rest of tech per NVDA dollar. Meanwhile the tape: Eisman (Jul 31) — SOXX −23% from its Jun-22 peak vs NASDAQ −7%, the selling concentrated in the AI supply chain. App Economy (Aug 1): equipment guided up into the chip selloff — Lam Research put the September quarter at $8.1B vs ~$7.1B consensus (+20% sequential), with NAND revenue doubling sequentially, advanced-packaging growth guided past 70% Y/Y, and its margin framework raised to mid-50s gross / mid-40s operating; KLA lifted its 2026 wafer-equipment market view to the low $150B range from "$140B plus," saying AI demand signals strengthened materially since March across logic, DRAM, HBM, NAND and packaging. The constraint is now delivery/absorption, not demand. 2026-AUG-02 (Jay Singh SSR call): Nomura models DRAM revenue from ~$80B (2022) to >$2.06 trillion by 2030 — 5-7x mainstream forecasts — with production 29.3B GB → 128.4B GB, utilisation above 100% in later years, inventory turning negative, and ASPs $1.90/GB (2023 trough) → $13.70 (2026) → a $18.60 peak (2027) → $16-17 by 2030. "The old idea that DRAM is just a pure commodity cycle is being completely rewritten by AI." The July semi sell-off was positioning: SK Hynix printed a record 76% operating margin / 83% gross with DRAM ASPs +30% q/q and Samsung guided to a 19-fold profit jump even as the Kospi fell 44%. The CXMT Shanghai-IPO scare is misread — 6-9% of DRAM (Samsung 36%, SK Hynix 29%), commodity DRAM only (no HBM), >30% higher cost per bit, needs 15% share to matter (unreachable even by 2028), and trades at 30x versus incumbents at 4-6x. China's domestic DUV program targets ~5 systems this year and 20 next versus ASML's 131 shipped (BofA: ~2.4% of sales). At the lows Micron traded at 5x forward, SK Hynix 4x, SanDisk 5x, Nvidia 15x (from 25x). App Economy (Aug 4): AMD's Data Center crossed over — more than doubled to a record $6.7B and 58% of revenue (from 42%) at a 31% operating margin, with EPYC and Instinct both driving 107% growth: inference and agentic workloads need CPU capacity alongside accelerators, so the buildout lifts more than GPUs. Helios full-rack systems begin ramping at Meta and OpenAI (Microsoft, Oracle, Anthropic queued), moving the NVIDIA contest to the system level; the dated test is 2027, when the partnership list must convert into the tens of billions of Data Center AI revenue promised. 2026-AUG-02 — Mohnish Pabrai (New Money): grants the memory oligopoly's moat in full — Micron's CFO told him that with every patent, engineer and process person, a rebuilt fab might not hit the same throughput: "there's a part of this business that's black magic" — plus allocation-driven pricing power ("take a number"), the 4–5× price delta hyperscalers now pay. He still passes: "even better than pickaxe makers is put the whole thing in the too hard pile" — the unanswerable question is where the three players stand in three-to-five years. 2026-AUG-08 — App Economy (PRO, Q2 prints): memory is structurally short on the earnings tape — Sandisk revenue +372% to $9.0B at an 85% gross margin, datacenter 12%→38% of bits, NAND on allocation beyond 2027, eight multi-year "New Business Model" contracts backed by $16.5B of financial guarantees locking >half of FY27 and ~two-thirds of FY28 bits ("AI eats NAND"). Arista has secured its memory for 2026 while expecting shortages through 2028; Axon guides memory inflation into Q3 margins; Sony says it has locked enough memory for planned PS5 volumes. Singh (Aug 9): Micron, Samsung and SK Hynix have sold out all 2027 DRAM/NAND capacity; customers are being allocated only 60-70% of requested volumes and DigiTimes' insiders call 2027 "the most severe moment of memory shortage." Micron is expected to grow EPS 850% in 2026 and double again in 2027. The visible casualty side: Apple −7% on memory-constrained guidance, and VISN cutting Aurora EBITDA guidance $25M because memory inflation stopped it shipping coaxial upgrade boxes. 2026-AUG-07 — CNBC Halftime, the desk buys in: Simpson adds NVIDIA at 218 ("not a trade, an investment") and Link makes her first-ever NVIDIA purchase at ~18× forward — "cheaper than Colgate or Hershey" and the cheapest since 2019 — while Micron discloses $100B in bookings on take-or-pay contracts (visibility, not backlog; Lebenthal had bought at ~$1,000 avg on the elongated-cycle case) and Broadcom carries a ~$100B AI pipeline. The Aug-10 caveat from Amoroso stands: everybody is in the memory trade, bottlenecks eventually resolve and ASPs come down — "not great for the stock prices of semiconductors," great for the buyers (Apple) and the software layer. AUG-16 (Jay Singh, SSR): Morgan Stanley now models ~$400B of combined operating income for Micron across 2026 and 2027 on HBM, with gross margins of 80-89% and operating margins of 81-87%; Bank of America goes further at $236 of EPS by 2030 with $100B of revenue locked in by long-term contracts. The pricing-power tell is the smallest number on the page: Micron says data-centre customers "remain largely insensitive to memory pricing." Within the complex, Goldman notes optical is outperforming memory by 20 points to start August after both fell 30% in July, "because the market sees it as a bigger supply issue." 2026-AUG-15 — App Economy Insights PRO (Applied Materials Q3): the bottleneck has moved from demand to the toolmaker. Revenue +25% to a record $9.1B with Semiconductor Systems +30% and a record 34% operating margin; DRAM is now 26% of Semi Systems revenue on the HBM race and advanced packaging is guided to grow more than 70% this year. Customers file eight-quarter forecasts and are "pushing Applied to deliver equipment faster," so Applied has nearly doubled manufacturing space and plans to double quarterly system output by 2028. Q4 guided $10.25B vs $9.56B consensus — and the stock still slipped after nearly doubling this year: the constraint is real and already discounted. 2026-AUG-17 — Joseph Carlson (After Hours): the valuation side of the trade starts drawing sellers: Carlson is trimming ASML ("the valuation is very stretched today… despite the fact it's been one of my best performers over the past 2 years"), and Pat Dorsey trimmed too — tempered by the overshoot lesson (Dorsey's first trim at ~$1,300 before the run to ~$2,000: "you can be very slow to take gains in stocks like this"). Dev Kantesaria refuses to cut it at all (+62% YTD) for a portfolio-construction reason: ASML is "the one confounding variable in his portfolio." 2026-AUG-17 — Jay Singh (David Lin Report), NVDA (Aug 26) checklist: data-centre revenue (85–90% of the top line) and its sequential growth, Blackwell → Vera Rubin ramp, 2–4-quarter forward visibility, GAAP gross margin vs the 73–75% band (TSMC chips-on-wafer packaging and HBM costs the pressure — "which is what hurt Apple"), hyperscaler capex read-throughs, the Singapore shipment investigation, Spectrum-X/InfiniBand and CUDA licensing. Memory peers run 85% gross margins — "which is insane." 2026-08-18 (App Economy Insights, Q2 13Fs): institutional money rotated into the equipment and memory layers — AMAT a top-five buy at three funds, MU at Coatue/Altimeter/Sands, STX at Lone Pine/Sands/Tiger. Six of the twenty tracked funds bought memory or storage in one quarter: "as AI infrastructure scales, funds are increasingly looking beyond compute to the components that feed and store all that data." Astera Labs (ALAB) was Light Street's #1 buy and Whale Rock's #2 on the interconnect thesis — as clusters scale, the constraint migrates from the accelerator to connectivity. Counter-signal: Druckenmiller exited AVGO, INTC and MU. 2026-AUG-19 (Ted Oakley, Oxbow Advisors, The Real Story) — the bear case: most vulnerable group in the market. Pull every SMH constituent's chart — all at new highs, basket up 20%, "usually that's a sign that you probably don't have a lot more to go." AMD, Intel and Nvidia named; "all the new money went the last six months"; 46 years of feast-or-famine back to TI. 2026-AUG-23 — Jay Singh (SSR call, 2026-AUG-23): Nvidia is raising AI-server prices 15%+ on Vera Rubin and Grace Blackwell systems because memory costs are soaring — and "the inability of the industry's most dominant company to hold the line on prices or absorb the growing costs shows how much leverage makers of memory chips like Samsung, SK Hynix and Micron have," with NAND and DRAM "parabolic." Hence: a 75% gross margin is "likely a peak margin." Supplier confirmation — SK hynix's surprise 30 billion buyback (JPMorgan overweight) and Micron's $10B Boise research commitment inside $250B of planned US spend, with UBS reiterating buy "given price increases and future capacity additions." Counter-case from Kedrosky (Meb Faber #648, Aug 28): the super-cycle is the current "this time is different" narrative (RAM never comes back, GPUs a permanent duopoly), and he is watching "incredible and unprecedented cash inflows into the Taiwanese chip and Chinese chip manufacturers" that imply a tsunami of supply in early 2028. In the most capital-intensive boom-bust industry on earth, "once you lock in supply, prices are going to zero" — fixed costs force output out the door. Two supporting observations: GPU utilization at only 35–40% even at peak load in one large GPU-rental fleet, while used A100 prices are bid on scarcity — i.e. hoarding and double/triple ordering, a latent supply overhang rather than consumption; and AI-designed silicon ("vibe chipping") collapsing the moat from inside — the inference-ASIC startup Edged cleared design verification in 42 days versus a normal six or seven months, first design working, so "the notion that some kind of tribal knowledge protects you as a chip manufacturer is going away." His marker for the eventual top: "remember when Micron was at such and such a price." 2026-AUG-30 (Jay Singh, SSR call): Nvidia guided to 70% revenue growth through FY28 against 47% consensus and the semiconductor index still fell 2.3% on the week — the cleanest statement yet that the group is being de-rated rather than downgraded (SOX 24-month forward P/E 21-22× → ~15×). Marvell raised FY28 revenue to $18B (+50%) and "the stock still didn't rally given this was priced in." Memory split two ways: Apple evaluating DRAM from CXMT and NAND from YMTC took SanDisk −6% and Micron −5% on top of threatened US chip tariffs — "which is why memory underperformed AI overall" — against TrendForce's server DRAM +270% y/y, enterprise SSD +235%, HBM +70-140% by 2027 and the fact that Nvidia's commitments jumping $119B → $279B is "primarily related to procurement of memory." The reconciliation is which memory: contracted/qualified/high-bandwidth versus the commodity end a state-backed entrant can flood. Overhang: the White House is weighing semiconductor tariffs "that would extend beyond chips to products like servers, laptops, and gaming consoles" — timing Singh calls "horrible" right after Nvidia's print. 2026-SEP-01: Trader Ferg relays the Kedrosky supply call verbatim as this week's pick: "the incredible and unprecedented cash inflows into Taiwanese and Chinese chip manufacturers, with what looks like a tsunami of supply in early 2028. And you know this is a boom-bust industry. So once you lock in supply, my friend, prices are going to zero. Why? Because I've got to cover my fixed costs. So this stuff is all going out the door." The mechanism is fixed-cost recovery, not demand: capacity funded now prices product in 2028 and clears at any price. (Ferg's Finds, 2026-SEP-01, quoting Paul Kedrosky.) Slegers 2026-JUL-12: the memory leg framed as a commodity peak, not a growth story — Micron, Western Digital and SanDisk "all up +200 to +700% this year" while "memory is a commodity business. These companies don't have any pricing power." The supply chain of causation is traced: the 2023-24 price war (Micron CEO Sumit Sadana on customers "being very aggressive with pricing") shut down fab investment, AI demand then hit starved capacity, and gross margins now exceed the 2018 peak that preceded a margin collapse and a falling stock the following year. Grantham quoted for the mechanism — "if you make obscene profits, you'll get ferocious competition" — and the conclusion is a refusal rather than a short: "the market is clearly not expecting any mean reversion… that's not a bet that I want to make." Slegers 2026-AUG-13: Arka Bhattacharjee's twenty-year locked-portfolio list puts three of ten names directly on the semiconductor and electrification build-out — ASML at #4 (“the only company in the world that can build Extreme Ultraviolet (EUV) lithography machines”), Applied Materials at #7 (the systems used to make “nearly every new computer chip and advanced display in the world”) and Schneider Electric at #6 — with Microsoft and Alphabet as the demand side, so more than half a ten-name portfolio rests on AI capital spending continuing. The framing is explicitly picks-and-shovels: “they supply the machinery to all chipmakers,” a claim on total chip manufacturing rather than on any single chipmaker. Neither the violent cyclicality of equipment orders nor the political exposure of a single Dutch chokepoint is addressed. Slegers 2026-APR-21: frames ASML as the supplier-of-record on AI capex — 90% EUV market share and "literally zero real competitors", machines that weigh "as much as two blue whales", a 20-30 year service annuity after every sale, and "every dollar spent on AI eventually flows back to ASML." And then passes on price: 36.8x forward against a 32.0x ten-year average, with the stated entry at 25x = €745 against a €1,245 price. A monopoly identified and declined is a useful marker for where this cycle's valuations sit. 2026-AUG-26 (Dowd/Phinance, WTFinance): "You don't need it to stop. You just need it to slow — it's a second derivative question, and semiconductors are the classic canary in the coal mine." The complex peaked 26 June 2026 on tremendous orders and good fundamentals, sold off, and the attempted rally "seems to be rolling over." Cleanest tell is South Korea's index (~50% Samsung + SK Hynix), which also peaked in June and is down 30–35%; a new low there says the second derivative has shifted and "the valuations of these stocks can't support a second derivative slowing." The backlog is inflated by double ordering — an X survey found ~50% of AI-infrastructure buyers admitting to it — plus warehoused shadow inventory awaiting unbuilt data centres. 2026-SEP-02 (CNBC Halftime — Lebenthal / Weiss): the sector argument reframed as one binding-constraint question — "what NVIDIA is saying is that the whole space is supply constrained. What Broadcom has been saying… is that there's competition out there. Now which one of those two is right?" Lebenthal sides with supply ("they could do more than 70% if supply constraints come out of the way, particularly on memory"), which makes competition moot: "if there is competition, say from AMD versus Broadcom, particularly with the Alphabet partnership, then it really doesn't matter… there's enough of this pie." Weiss's expression is the chokepoint: Taiwan Semi (Stifel initiates buy, $515) is "the must own in the sector… a better must own, frankly, than NVIDIA because Nvidia's getting their chips made there. So is Google, so are all the others" — and he inverts the usual capacity warning: "they'll fill it with demand… it's going to take years to add it." Lebenthal's qualification: NVIDIA's CUDA software ecosystem is a switching-cost moat independent of who fabricates. 2026-AUG-28 (McCracken, Value Hive): memory has another leg, argued from the architecture rather than from hyperscaler capex. Digital compute's generality forces bit-flipping, so scaling is a memory bottleneck — "that's why you end up needing all this memory and Micron stock's going up." The incremental demand he can evidence first-hand: his own on-premise LLM quote was "like 4 million U.S. dollars and most of the cost is memory," implying demand from independents and small businesses, not only clouds. Verdict: "memory and compute are evolving into commodities now… memory has got another leg to its cycle that will probably break people who aren't long memory when they watch it go off." He bought Micron on the dip, deliberately small. The AUG-03 SSR fund-letter compilation carries the same argument from both sides. Baron (Samsung) makes the supply case physically — HBM is more wafer-intensive than DRAM, NAND capacity can no longer be converted into DRAM, greenfield fabs take two to three years — and adds a de-commoditization claim (long-term agreements, co-design with accelerators) plus a 4× P/E where "we're getting the non-memory parts of Samsung for free." Alger makes the behavioural version at Western Digital: two scaled HDD makers "prioritizing higher areal density… rather than adding significant unit capacity," already visible in margins above guidance. Harding Loevner, in the same document, prices the other side: disciplined statements "suggest that prices and margins could stay higher for longer," but the same demand "is now incentivizing capacity additions… For the companies, such investments may be rational. For the stocks, it could prove more complicated" — with SK hynix already funding advanced lithography from ADR proceeds. (fund pitches compiled by SSR, AUG-03) (2026-SEP-03, Dan Niles, Excess Returns) "I'm actually starting to wonder if the semiconductor bubble breaks before" the AI trade does — semis are still cyclical and "a lot of investors are vastly underestimating what China is doing." CXMT (DRAM) is public and lifting wafer starts 300,000 to 500,000 by the end of next year at ~8% global share "ramping incredibly fast"; YMTC (NAND) lists soon and "plan to be bigger than Samsung or SK Hynix in the NAND business by the end of next year"; CXMT has already produced high-bandwidth memory "yielding very very well… only a generation behind." If China simply meets its published targets, shortages "through 2030" have "very minimal chance." Memory valuations only look cheap "if… this is no longer cyclical." He exempts Nvidia specifically — at 15× calendar-27 earnings against the S&P at 19×, "this isn't Cisco in 2000." 2026-SEP-07 (Jay Singh) — a supply-chain census off Dell's call: constraints “remain the same across DRAM, NAND, spotty CPU shortages, disk drives, power ICs, microcontrollers, drivers, ABF substrate, T-glass and optical” — nine categories, alongside a $47bn (+58%) beat and HPE at +34%. Pricing: Samsung has locked ~70% of memory capacity through 2031 into LTAs (after SK Hynix and Micron), HBM3E 36GB spot at $2,100 = 4–5× contract, HBM4 16-stack spot $3,500, DRAM export prices +36.6% May→July on 13% lower volumes; Micron adding 60k wafers/month; China's CXMT stalled HBM3E production, “positive for Micron and SK Hynix and Samsung.” 2026-SEP-08 (WSJ AI & Business): the memory squeeze is now reaching the consumer device — PC unit sales are sagging yet prices are up and projected to climb further, partly because computer-memory costs “have gone through the roof” as memory-intensive AI data centers eat the supply (the other part being an on-device-AI premium at HP and Dell). 2026-SEP-10 — a live index short (JF Tardif, Timelo). "I'm short the SOXX index, S-O-X-X," sized against the tape rather than a target — "if it moved up from here I probably short more. If it went down 10, 15% here I'll probably cover half of it" — plus "some specific semiconductors." He does not need a collapse; the structural case is that if spending reaches $1.7trn and then falls 20%, "these stocks are going to go down, and because they're so important also in the S&P 500 they might take the whole stock market down with that." App Economy Insights (2026-SEP-11): the memory crunch ('RAM-aggeddon') is now in consumer list prices. Apple raised iPhone 18 Pro and Pro Max by $100, raised several older models, and dropped the standard iPhone 18, so 'hardware growth now relies almost entirely on pricing power rather than unit volume.' This echoes Best Buy's mid-teens PC ASP rise on falling units. Davolos (Sep 12, not a semi analyst): Chancellor's capital cycle says 'this time is different' margins mean-revert. Google and Amazon building their own chips to escape the 'Jensen tax' is the warning, though the cycle could run longer than expected. Dan Niles (2026-SEP-04): "We don't have a multiple bubble to some degree but we do have an earnings bubble… having memory companies having operating margins of 80% is not normal," and once China reaches its memory ambitions (CXMT listed, YMTC listing, a captive ~20% of world PCs and phones) "those margins will go back to below the average."