CA @semarb

메모리 슈퍼사이클이 곧 끝난다고 보는 뷰가 늘었다. 논리는 대충 이렇다. 하이퍼스케일러가 비싼 가격에 무한정 사주진 않을 것고, 중국 메모리 공급이 풀리고, 테라팹 같은 수직통합이 들어오면 2028 컨센서스가 무너진다는 것. 그래서 지금이 메모리 주식의 피크아웃이라고 본다. 일부는 맞는 말이고, 메모리가 사이클 산업이라는 본질은 변하지 않는다. 가격이 꺾이면 이익이 사라진다는 것도 변하지 않는다. 다만 베어 관점이 착각하는 한 가지 가정이 있다. "효율화가 일어나면 수요가 줄어든다." 쉽게 말하면.. 메모리가 비싸지니까 빅테크가 안 살 것이고, 성능이 좋아지면 적게 사도 되니까 결국 수요가 감소한다는 시각. 그런데 역사적으로 보면 정반대였다. 효율화는 수요를 줄인 적이 없다. 항상 파이를 키웠다. 90년대 인터넷이 처음 깔릴 때.. 회선 속도가 빨라지자 무슨 일이 벌어졌나. 사람들이 인터넷을 적게 쓴 게 아니라 영상까지 보기 시작했다. 데이터 사용량이 곱셈으로 늘었다. 2000년대 모바일도 같았다. 통신 단가가 내려가자 사람들이 통화를 줄인 게 아니라 모바일 인터넷, 스트리밍, 영상통화를 같이 쓰기 시작했다. 클라우드도 마찬가지다. 컴퓨팅 단가가 내려가자 기업들이 서버를 안 산 게 아니라, 클라우드 위에 SaaS, 빅데이터, 머신러닝까지 올렸다. 경제학에는 이걸 부르는 이름이 있다. 제번스 역설. 자원의 효율이 좋아지면.. 자원 사용량이 줄어드는 게 아니라 오히려 늘어난다는 법칙이다. 19세기 영국에서 증기기관 효율이 좋아지자 석탄 소비가 줄지 않고 오히려 폭증했다는 데서 나온 개념이다. 이 법칙이 지금 AI에서 그대로 작동한다. HBM3 - HBM3E - HBM4로 가면 대역폭이 1.5~2배씩 뛴다. 같은 GPU 한 대가 처리할 수 있는 토큰 수가 그만큼 늘어난다. 토큰당 비용이 내려간다는 뜻이다. 그렇다면 베어 시나리오대로 빅테크가 메모리를 덜 살까. 반대다. 토큰당 비용이 내려가면, 지금까지 AI 쓰기 비싸서 망설였던 곳들이 더 많이 들어오기 시작한다. 더 긴 컨텍스트, 더 복잡한 모델, 더 많은 사용자가 가능해지는 것이다. 시장 파이가 통째로 커진다. 추상적으로 들리니까 현장 얘기를 해보면, 처남이 램리서치 7년차다. 얼마 전 전화가 왔는데 사장이 직원들한테 2030년까지 자사주를 팔지 말라고 강조했다고 한다. 램리서치가 어떤 회사냐. 메모리 반도체 장비, 그중에서도 식각 증착 핵심 공급사다. 회사 사장이 직원한테 자사주를 2030년까지 들고 있으라고 따로 강조하는 건 그냥 던지는 말이 아니다. 회사 내부에서 보는 메모리 자본지출 사이클의 끝이 2030년보다 한참 뒤라는 신호다. 베어가 그리는 "2027~2028 피크아웃" 시점이랑 얘기가 다르다. 내 경험도 같다. 음악 관련 AI 플랫폼이랑 MOU 맺고 기업용 버전을 쓰고 있는데, 내년 연 사용료 예산을 또 올려야 한다. 업체가 직접 이유를 말했다. "서버 운영비, 그러니까 토큰 사용량이 계속 늘어나서 데이터 사용료를 올릴 수밖에 없다"는 것이다. 다른 산업의 AI 플랫폼 쪽 사람들 얘기도 같다. 고객 기업이 새로운 AI 모델을 계속 요구하는데, 결국 가장 큰 문제가 서버 운영비라고. 이게 뭘 뜻하느냐... AI를 이미 잘 쓰는 집단이 더 깊고 더 넓게 쓰고 있다는 것이다. 도입 단계의 수요 증가가 아니라, 정착 단계에서의 사용량 폭증이다. 한 번 AI를 업무에 쓰기 시작한 회사는 멈추지 않는다. 더 긴 컨텍스트, 더 많은 사용자, 더 복잡한 모델로 계속 확장한다. 제번스 역설이 현장에서 그대로 작동하는 모습이다. AI 단위 비용이 내려갈수록 사용량은 줄어드는 게 아니라 곱셈으로 늘어난다. 그래서 메모리-GPU-전력 수요가 같이 곱셈으로 늘어난다. 그러니까 베어가 그리는 "비싸지면 빅테크가 안 산다"는 시나리오는 한 가지 핵심을 놓친 그림이다. 빅테크는 비싸도 산다. 안 사는 게 비합리적이라서가 아니라, 사는 게 더 합리적이라서다. 효율화가 토큰당 비용을 낮추고, 토큰당 비용이 낮아지면 AI 활용 범위가 더 커지고, 활용 범위가 커지면 또 더 많은 메모리가 필요해진다. 이것이 끊이지 않는 한 사이클이 끝나기 어렵다. 중국 공급 얘기도 한 단계 들어가야 한다. CXMT가 DRAM 1위로 올라왔고 YMTC가 NAND 증설 중이라는 건 맞다. 그런데 중국이 들어오는 영역과 빅테크가 사는 영역이 다르다. CXMT는 DDR4와 일부 DDR5 같은 범용 메모리에서 양산 중이다. YMTC는 NAND다. 빅테크가 AI에 쓰는 건 HBM, 첨단 DDR5, LPDDR5X다. 이 두 영역 사이에 기술 갭이 상당히 크다. CXMT가 HBM 양산까지 가려면 EUV 장비 제약, 1c 노드 수율, TSV 패키징... 다 미국 제재 라인 안에 묶여 있다. 최소 2~3년, 아마 5년 이상은 더 걸릴 것이다. 그래서 베어가 그린 "중국 공급 풀어 가격 붕괴" 시나리오는 범용 메모리에선 가능해도, HBM과 첨단 DRAM에선 2028까지 거의 영향 없다고 봐야 한다. 그리고 메모리가 그렇게 쉽게 만들어지는 기술이 아니다. TSMC가 파운드리 생태계 만드는 데 30년 걸렸다. SK하이닉스가 HBM에서 지금 자리 잡는 데 10년 넘게 걸렸다. 베어 관점에 묻고 싶은 게 하나 있다. 인터넷이 빨라졌을 때 데이터 사용량이 줄었나? 클라우드가 싸졌을 때 서버 수요가 줄었나? 스마트폰이 보편화됐을 때 통신 인프라 투자가 줄었나? 답은 다 똑같다. 효율화는 한 번도 시장 파이를 줄인 적이 없다. 항상 파이를 키웠다. AI에서도 같다. 토큰당 비용이 내려갈수록 AI를 쓸 수 있는 산업이 늘고, 한 회사가 쓸 수 있는 범위가 늘고, 결국 더 많은 메모리가 필요해진다. 베어가 그린 "효율화 = 수요 감소" 등식은 역사적으로 한 번도 성립한 적 없다. 이번에도 다르지 않을 것이다.

The Information: NVIDIA is testing Intel EMIB packaging for Feynman. The Information: NVIDIA is conducting early tests of Intel 18A. $INTC

Jukan @ICML @jukan05

3 weeks ago

The information: Google has placed an order with Intel to package more than 3 million TPUs. Across 2027 and 2028, Google is expected to produce more than 6 million TPUs using Intel’s packaging technology.

27 86 727 249K 119

TSMC fumbles Copackaged optics for the Nth time like some fucking donkeys and now the whole industry is limping towards NPO, and the pod bros who price the entire AI TAM off Nvidia’s BOM line items still can’t actually explain what the problem is. So let me do the engineering for you, since clearly nobody on here will. The bottleneck was never can you make light go through a waveguide. It’s all fucking thermals which is downstream of packaging. Specifically, how do you get a photonic engine onto the same substrate as a switch ASIC or XPU without your yield falling off a cliff and your reliability failing. TSMC’s answer is CoWoS where they bolt everything onto one big monolithic silicon interposer. Cute, until you hit the reticle limit and start duct-taping interposers together (CoWoS-S, then -R, then -L, soon -PleaseStop). Every chiplet and HBM stack you add to that single interposer compounds your defect probability and one bad die leads to a five-figure package going into the dumpster. CoWoS is thermally retarded and the whole industry knows this and it’s why capacity “can’t expand” and Jensen is acting like a bouncer in the front of a club choosing who gets pass the velvet rope. There is ONLY one company that will make copackaged optics work and expand in the rack… it’s not Lumentum, it’s not Coherent, it’s fucking INTELLLL. Intel’s EMIB gets rid of a giant reticle limited interposer and replaces it with a tiny silicon bridge that does the high-density coupling locally, exactly where you need it. You localize the hard part and the thermals in one area and your yield is ridiculously high. Comparing EMIB & CoWoS is so funny cause EMIB is north of 95% yield with like 12 reticle size equivalent package while CoWoS falls off a cliff after 5.5 reticles it’s that bad…now imagine adding thermally sensitive photonics. People don’t know this but Intel has been doing silicon photonics in-house for ~25 years... In 2024 they showed an Optical I/O chiplet doing 2 Tbps bidirectionally at ~5 pJ/bit, with the PIC and EIC co-packaged right against the ASIC and it’s all because of EMIB. And even more critically than that, they’ve actually run the fiber-attach and reliability/test flow to JEDEC-grade standards already, which everyone hand-waves until their links flap in production. My prediction is clear: Intel will capture over 90% of the copackaged silicon photonics market in the next five years because there is NO ALTERNATIVE.

keyboard_arrow_left Previous keyboard_arrow_right Next

SK하이닉스 일본 주재원 계신 분이랑 커피챗 내용 간단하게 공유 1. 메모리 수요가 너무 강력해서 일본 고객들에게 줄 메모리가 없다. 미국이랑 중국 고객 주고 나면 일본에 줄 물량이 없다고 함. 그래서 일본업체들 찾아가서 매번 쓰미마셍 고개 숙인다고 함. 2. 메모리 부족으로 닌텐도 같은 전자제품 가격 계속 올리는 중. 메모리가 없어서 난야 같은 업체 뿐만아니라 윈본드 같은 업체도 찾는중. CXMT는 기가바이트 통해서 이미 소비자용 제품으로 들어가고 있다고 함. 3. 일본에 있는 CXMT 한국엔지니어를 통해 듣기로는 CXMT 설계능력이 생각보다 엄청 뛰어나다고 함. 2-3년 격차가 아니라 그보다 더 좁혀질것 같다고 함. 진짜 중국의 똑똑한 인재들이 밤낮 안가리고 일한다고 함. 한국도 경계해야 할듯. 4. 메모리 가격은 계속 인상 중.. 몇프로인지는 밝힐 수 없지만 계속 인상중. 오늘 엔비디아 소캠에 메모리 줄었다는 이슈로 주가가 내리는데 메모리 가격은 오히려 올라서 영업이익은 계속 올라갈 전망. *개인적으로 걱정되는건 이런 메모리 가격 상승이 Capex 투자 부담으로 이어져서 AI 싸이클이 빨리 끝날까봐 걱정임. 구글도 채권찍고 있고 빅테크들도 현금이 떨어져 가는게 걱정됨. *투자권유, 매수,매도 추천 아닙니다.

Belated $.02 on $MU: It's fair to wonder how long GMs will remain around 80% (esp. with non-HBM DRAM GM higher than HBM GM). But all the same, Micron just guided for $75+ in annualized EPS, all while token consumption and KV cache needs keep soaring (see recent OpenRouter and GPU pricing data), agentic workloads also drive higher DDR5 demand, HBM4 takes the HBM trade ratio from 3:1 to 4:1, and $NVDA preps Rubin Ultra racks for 2H27 that consume 4x as much HBM as Rubin and GB300 racks. All cycles eventually end and China will in time build a ton of capacity for NAND and non-cutting-edge DRAM (lack of access to EUV will be a major handicap for them with high-end HBM and server DDR5). But at ~4x potential 2027 EPS (some buyside estimates have gone above $100), Micron seems to be pricing in a fairly short cycle when it's far from clear that supply will catch up to demand so quickly. (And of course, Hynix, which has a higher HBM share and just said it thinks memory shortages are likely until 2030, is even cheaper.)

$NVDA buying up all the memory. This makes more sense now that we realize they have secured most, if not all, they need for 2026 and some of 2027.

🚨 NEW: Anthropic announces 10 new Claude cowork plugins to eliminate 10 new industries. Man plans, Claude laughs.

Added 65 new write-ups to the site (pt final): @SuEricGP - $AMD, $MNSO, $CRM, $PYPL, $ADBE, $EVVTY, $VRTX, $KNSL, $V, $TMDX, $NVO @__MattLindsay (watchlist) - $215A.T, $DUOL, $ICLR, $LFCR @ArmsGarrett - $3445.T @accrued_int - $PINS (update) @BanyanLaneCap - $VOD (update)

keyboard_arrow_left Previous keyboard_arrow_right Next

The next $1T AI bottleneck isn't power, compute, or memory. It's connectivity. And the only solution is light. So I launched @joinautopilot's first ever photonics themed stock portfolio. AI clusters are hitting a "Copper Wall." Signal decay, thermal loss, and power constraints make electronic interconnects unviable past 800G. Networking already eats up to 50% of data center power. As Gilad Shainer, Senior VP of Networking at NVIDIA, recently noted, the exponential growth of AI requires a “new class of high-speed, scalable networking". The physics are simple: electrons move through copper and generate heat. Photons travel through fiber with near-zero loss. As we scale to 1.6T and 3.2T architectures, every watt saved on interconnects is a watt returned to compute! Without this transition to light, the industry faces a physical ceiling that threatens to stall the progress of LLMS. A full supply chain repricing over optical tech is just beginnnig. The "Photonics Is Next" portfolio maps every layer: $LITE (21.07%) Single most critical chokepoint. ONLY supplier of 200Gbps EMLs. No EMLs = no transceivers = no AI cluster. NVIDIA GB200 and Google TPUv7 critical supplier. $COHR (20.33%) Most vertically integrated player in photonics. 6-inch InP wafers mean 4x the chips at half the cost. Book-to-bill over 4x. $5.81B record backlog booked through 2027. No competitor matches this integration depth. $TSEM (18.96%) ONLY pure-play silicon photonics foundry. NVIDIA partnership is the strongest endorsement in the space. 50+ customers. Tripling capacity with a $650M expansion. CPO foundry tech. $CIEN (15.59%) WaveLogic 6 is the only 1.6T coherent system deployed at scale. $7.8B backlog. Just re-entered the S&P 500. 50% North American optical networking share. In-house 3nm silicon. $AAOI (14.81%) $4B Amazon deal transformed this company. In-house InP laser fab gives them a supply edge during the industry-wide EML shortage. 3 of the top 5 hyperscalers as customers. Ramping to 200K units/mo. $INTC (9.24%) Pioneered silicon photonics. Shipped over 8M photonic integrated circuits with 32M on-chip lasers. Their OCI chiplet is the first fully integrated optical I/O co-packaged with a CPU. Smaller position, optionality on a turnaround. Link below 👇

Morgan Stanley PT for Kioxia ==> +50% current SP But (correct me if my math is wrong), that means a P/E 6.5 by Q1-2027 🥴 So imagine what that can means for $SNDK SanDisk (more potential, higher multiple) Also bullish for the other SSD manufacturers: $MU SH hynix

Jukan @ICML @jukan05

4 months ago

Morgan Stanley raises Kioxia target price to ¥33,000 The firm significantly raised its CY2026 ASP growth forecast for Kioxia from +75% YoY to over +100% YoY, with Q1 ASP expected to surge approximately 90% QoQ, followed by an additional 10% QoQ increase in Q2. This suggests

4 19 167 23K 54

This made me laugh. 25 year-old professional podcaster @dwarkesh_sp : I don't get it. If I were you and had country-of-geniuses-level AI, I'd be happy to buy $5 trillion of compute! Dario, founder of a $380B AI model company growing at 10X per year: Bro! $300 billion of compute per year in two years is a lot, OK? It's a lot! We don't want to go bankrupt if our forecasts are off by a year.

$MU Micron Technology is a $600 stock on CFO's comment on guidance and margins: "We guided 68% last quarter, and we indicated today that our financial outlook has improved since the guidance... We do believe margins will expand in Q2 and Q3. We expect the market to remain tight beyond 2026. That really supports the pricing environment. We're ramping out gamma node, so our cost performance has been good. Our spend control has been very good. On product mix, it has also been a positive story."

$MU $SNDK $SMIC The below is from the SMIC Q4 earnings call (available on @Quartr_App) . The CEO, Zhao Haijun provided commentary on the memory shortage but it is a bit more nuanced than just "shortage eases in Q3." He laid out how the memory cycle plays out in two entirely separate tracks, consumer and AI, driven by different supply constraints and different timelines. Starting with demand, Zhao noted that "everyone's compute capacity is insufficient right now because everyone has grand visions for AI, and they want to build the next ten years' worth of data centers in one or two years." He compared it to building highways or high-speed rail ahead of actual traffic. On that basis, he stated the AI-driven memory shortage "should persist for several years" because "you can't just massively reinvest in memory, front-end, back-end, HBM and have it all come online at once." New fabs can ramp memory wafer production relatively quickly, but "those wafers from new equipment and new fabs will take a very long time to be validated for use in AI data centers." New capacity therefore cannot serve AI, but it can serve consumer electronics. The current tightness in phone and PC memory exists not because there isn't enough total memory capacity in the world, but because "phone and computer production capacity has been redirected to AI, because these are existing, established fabs." Established fabs with qualified processes got pulled toward higher-margin AI work, leaving consumer short. New fabs coming online can't fill the AI gap because they lack the qualification, but they can fill the consumer gap that was created when the established fabs shifted over. AI memory follows a completely different timeline because the constraint is not wafer production. Zhao stated that even a year from now, the binding constraint on HBM "won't be the front-end wafer fabs anymore" because "the bottleneck isn't at the front-end, it's at the back-end." He specifically pointed to stacking, final testing, yield improvement, and quality validation: "Final testing with so many wafers stacked together is extremely slow, yield improvement is very slow, and quality validation is even slower." Even if you redirected wafer fabs toward AI, "it won't help" because the gating factor is downstream. His conclusion was that "the few largest players who started doing HBM early will always be in the lead. Newly built fabs won't capture the HBM upside that quickly." They can, however, "immediately address the shortages in mid-to-low-end phones, computers, edge devices, watches, and consumer memory." So the distinction is: consumer memory is short because qualified capacity was reallocated to AI, and that gap gets filled relatively quickly once new unqualified capacity comes online. AI memory is short because of back-end bottlenecks that no amount of new front-end wafer capacity can solve, and that persists for years. Zhao emphasized that these were his personal observations and "not necessarily correct." *** The above information was translated from Chinese. The original earnings call is available on the Quartr App linked below.

keyboard_arrow_left Previous keyboard_arrow_right Next

Ryan Leung 梁興盛, CFA @RYANHINGSHING

5 months ago

SMIC’s CEO expects memory shortage in consumer electronics to ease in Q3 2026 1. New memory capacity will come online, but qualification for data center takes time, so near-term supply will flow into consumer electronics. 2. Memory hoarding will unwind

11 18 150 103K 77

@byul_finance Doyred

LFG!!!!!!!!!!

Illiquid @illyquid

5 months ago

FSC has approved IBKR access to Korean stocks so Fintwit can mop up the Korea Discount.

keyboard_arrow_left Previous keyboard_arrow_right Next

18 40 469 138K 164

What has become apparent If you don’t invest huge your company goes away. Simple. Every large company ceo know this. Not hard to know .. Nvda 221 comes. Shorts had their chance. Now if you miss it. Look out

My favorite @perplexity_ai prompt I’ve made for traders. Feel free to copy and use. ———————————————— Please analyze [TICKER] for me and provide the following, concise and clearly organized: 1. **Explain what the company does in like I'm 12 years old** - three short bullet points about what it does and any helpful relatable examples and analogies. 2. **Professional summary (max 10 sentences)** - industry, main products/services, primary competitors (list tickers), notable metrics or achievements, competitive advantage/moat, why they are unique and if they are a biotech provide if they have a commercial product or in clinical stages. 3. In a table, provide the follwoing: * Any hot theme, narrative or story of the stock * Any catalysts (earnings, news, macro) * Any significant fundamentals (huge growth in earnings or revenues, moat, unique product or service, superior management, patents etc) 4. **Show all the main news/events for the last 3 months:** - Use a bullet-point table for: - Date (YYYY-MM-DD) - Event type (Earnings, Product Launch, Analyst Upgrade/Downgrade, etc.) - Short summary (max 1-2 sentences) - Direct source link - Mark any major price-moving events (surprise earnings, large guidance shift, top-tier analyst actions). 5. **Mention any recent insider buys/sells or institutional filings if visible.** 6. **Summarize how the stock is moving vs. main competitors and overall sector trend in past month (up/down).** 7. **Flag upcoming catalysts (earnings, product launches, regulatory events) in the next 30 days.** 8. **Note any changes in analyst price targets for this ticker during the period above.** - Format for easy review. If possible, use tables for events and peer moves. - Respond in clear, concise, easily readable style for use in trading decisions. Overall, Focus on the reasons why the stock can make a big move in the future - earnings, sales, guidance, product launches, analyst upgrades/downgrades, insider buying especially from CEO/Founder and executive team, partnerships, and sector/news catalysts. I want to focus on stocks with catalysts and themes as catalysts are the cause of big moves in the stock market. Finally, discuss and bring up any relevant previous perplexity queries and conversations.

The Kobeissi Letter @KobeissiLetter

5 months ago

BREAKING: Perplexity announces an advanced version of Deep Research, which beats both OpenAI and Anthropic in benchmarking tests. The AI race just keeps on accelerating.

256 305 4K 559K 441

Wasn’t it already widely known news that Micron wouldn’t be able to supply HBM4? I don’t understand why everyone is making such a big fuss about it. And selling Micron because of this is foolish, too. Micron can earn even higher margins from server DDR5 instead of HBM4.

We wouldn’t characterize this as Micron having zero Rubin HBM4 share, nor do we think it’s particularly negative for Micron. Based on our supply chain checks, they still appear to retain some allocation. More broadly, Micron likely benefits under almost any scenario — that’s simply the nature of a tight memory market where supply remains the binding constraint. More importantly, from an economic perspective, DRAM is currently a better capital allocation choice for Micron than HBM4. On a wafer basis, conventional DRAM delivers roughly ~3× the bit output versus HBM once you factor in stacking complexity and yield drag, and with DRAM pricing still elevated, the effective revenue per unit capacity can significantly exceed HBM in this part of the cycle (2x or above). Given the ongoing tightness in memory supply — which we expect to persist through at least 2026–2027 — prioritizing DRAM production is rational rather than a sign of competitive weakness. And even if Micron pushes further into HBM4, Nvidia isn’t the only buyer: hyperscaler ASIC programs, TPUs, and other AI accelerators will likely absorb essentially all incremental supply. This remains a supply-constrained market, not a demand shortfall story. Therefore, Micron probably benefits under almost any scenario — that’s the upside of being in a structurally tight memory market where supply, not demand, is the binding constraint. $MU

Halal Investor @_HalalTrader_

5 months ago

$MU If true, this is actually bullish for Micron! As @fundaai put it - “1GB of HBM sells for $18, while switching to DDR5 can fetch a price of $60.” Why would Micron chase after HBM when the DRAM shortage will last into 2030 and they can make 3x more on DDR5?

2 15 128 64K 25

Elon Musk on Cheeky Pint: "Nvidia’s output is FTPing files to Taiwan. It’s digital. Now, those are very, very difficult. They’re the only ones that can make files that good, but that is literally their output. They FTP files to Taiwan."

keyboard_arrow_left Previous keyboard_arrow_right Next