Druckenmiller's Doctrine ⚾️🏟️
==================
Analyze macro from the bottom up
Posses an insatiable desire to learn
Have intuition to ensivion the winners in the next 2-4 years
Build a strong unshakeable belief in your own capabilities.
Seek Asymmetric Returns
Be patient to wait for a pat pitch.
When you really see the ball, swing really big for home run.
Position sizing is 70-80% of success.
Have no emotion to clean the slate and confidently restart from scratch, if the reason of buying a stock is no longer a case,
It's not whether you're right or wrong; it's how much you make when you're right and how much you lose when you're wrong.
Be open-minded to change your mind very quickly.
The way to build long-term returns is through preservation of capital and home runs.
. . .
Bottom Up -> Desire to Learn -> Intuition -> Self-belief -> Asymmetric Returns -> Be patient -> Swing Big -> Position Sizing -> No Emotion to Restart -> Right > Wrong -> Be open-minded -> Capital Preservation and Home Runs ⚾️🏟️
@demian_ai Short term SK Hynix wins.
Longer term Samsung wins from integrated Logic+Memory+Packaging in house. Foundry business just turns into profit.
$MU is only US HBM, so always has a premium in the price.
@JesusFerna7026 You can look at data, but if you don't deeply understand Thai culture, you will judge by Western standard, which is vastly wrong point of view.
You don't need to like or invest in $NBIS, but you need to follow this company's business and all neocloud's moves very closely.
If demand for the NeoCloud business drops, the rental price per GPU falls, the time to close the next deal expands, no new client names appear, or the margin from the next deal falls below the previous one, this is an early indicator of a cycle peak and slowdown.
$CRWV $NBIS $IREN
The AI infrastructure race is ON. CapEx spend has never been greater.
At the centre of it: Nebius.
$66BN market cap. Going head-to-head with the largest hyperscalers on the planet.
Leopold Aschenbrenner just made them one of his largest positions.
I sat down with @nebiusai Co-Founder, @romanchernin and I’ve condensed my notes below:
1. How Reducing the Cost of Intelligence Increases Consumption
Reducing the unit cost of intelligence triggers Jevons Paradox: total compute consumption rises as previously uneconomic tasks become viable. At scale, builders move toward tunable open-source architectures and specialized post-training, while frontier labs expand into larger, harder reasoning markets.
2. If Nebius Doubled Pricing, How Would That Impact Demand?
Nebius’s pricing power is capped by customer economics. If inference costs rise too high, customer margins break and demand stalls. The real edge is not nominal GPU pricing, but Total Cost of Ownership: caching, runtime optimization, and distillation can shift token economics by an order of magnitude.
3. If Nebius Had 10x the Capacity, Could They Sell It?
The real question is not whether raw demand exists, but whether Nebius can diversify it. Bare metal concentrates revenue around a few global giants. Moving up the stack into managed infrastructure and specialized inference expands the market to thousands of application developers.
4. What Is the Single Biggest Threat to Nebius?
The biggest threat is extreme consolidation into three to five closed tech empires. If a few conglomerates control the frontier model landscape, independent clouds risk becoming low-margin physical-layer vendors. Survival depends on a broad, democratized ecosystem of independent builders.
5. Who Actually Holds Power Against Nvidia?
Power against Nvidia comes from engineering credibility, not political posturing. Nvidia is deeply engineering-driven, so influence comes from proving technical capability across the stack. Differentiation requires a world-class team that reliably executes and earns operational respect.
6. Surviving the Hyper-CapEx War
Competing with hyperscaler CapEx requires respecting operational timelines. Capital cannot compress a six-month infrastructure bottleneck, but over 18 to 24 months it matters. It lets providers parallelize execution, secure power, lock data centers, and prepare capacity ahead of GPU deployment.
7. The Shark Rule: Move or Die
Cloud infrastructure is a post-sales business: every funding round or contract is only a credit and an opportunity to deliver. Survival requires relentless forward motion, disciplined execution, and focus on daily operations rather than emotional market spikes or consolidation noise.
(links below)
Thank you. You’ve highlighted a critical systemic risk.
If US AI Labs and Big Techs continue to lose ground to Chinese open-weights, the ultimate funders who subsidize the massive AI CAPEX bills may heavily scale back their spending. Could this one day trigger a severe margin compression and collapse across the entire hardware supply chain?
I had a fun run trading Japanese stocks this year. 🇯🇵
My individual stock approach was straightforward: use fundamentals to identify AI chokepoint monopoly or global leaders, then use valuation (RRR, P/E) and technicals to manage entries and exits. The average return per position was ~25%.
I even switched $SPY and $QQQ on my retirement account into Nikkei225 Index. This showed how much I like Japanese stocks.
That said, I’ve now exited all of my Japanese positions, except for my allocation in the AsiaSemi ETF, due to the macro backdrop shown below.
At this stage of cycle, I don't want to find macro headwind head-on.
There will always be another opportunity. For now, I’m content to read manga and wait patiently.
My picks:
$5802 Sumitomo Electric - Safe choice
$4203 Sumitomo Bakelite - Hidden value play
$8035 TEL - The leader
$4368 Fuso - Hidden gems
$5332 TOTO - Low risk, mid return
$6834 Seikoh Geiken - Smallcap high growth play
$4963 Shin-Etsu - Half growth (Wafer), half flat (PVC)
$2802 Ajinomoto - The global chokepoint in ABF
$4062 Ibiden - The substrate leader, but expensive
"Chipflation: Navigating A Memory Crisis"
메모리 사이클 본질의 변화를 읽어라...
-
1. 메모리 사이클이 짧아지고 있다.
역사적으로 메모리 다운사이클은 4~8분기, 업사이클은 4~9분기였다.
그런데 2018년 조정은 4분기만, 2024년 다운사이클은 8개월만에 끝났다.
지금은 2025년 4월에 시작된 업사이클의 13개월차.
왜 짧아지는가...
모건스탠리는 LTA, 즉.. 장기계약 때문이라고 본다.
빅테크가 메모리 회사와 몇 년치 물량과 가격을 미리 고정하는 계약이다.
메모리 가격이 단기 스팟 가격에 좌우되던 과거와 달리, 이제 LTA가 가격 안정성과 실적 예측 가능성을 가져온다.
4Q25~2Q26 큰 인상 후 QoQ 인상폭이 8~13%로 좁혀질 거라 본다.
이것을 '사이클 끝'의 신호로 읽으면 안 된다는 것이... 모건스탠리 시각이다.
'사이클 본질의 변화' 신호로 읽어야 한다.
-
2. 메모리 회사 변곡점 신호
모건스탠리가 모니터링하는 4가지 신호.
DRAM 가격: LTA로 안정화 중. 변동성 자체가 낮아지는 단계.
재고: 공급사 사상 최저. 서버 고객은 여전히 적극 주문 중..
Capex: 과거 다운턴엔 23% 삭감됐는데, AI 수요가 워낙 크니 다음 다운턴에서도 capex가 유지될 가능성.
실적 상향 & 멀티플 확장: 시장이 아직 LTA 프리미엄을 가격에 반영 못 함.
-
3. Agentic AI가 진짜 게임체인저
이 부분이 보고서의 핵심 인사이트.
지금까지 'AI 메모리 수요 = GPU 옆 HBM' 위주였다.
그런데 agentic AI, 스스로 추론하고 행동하는 AI가 본격화되면서 CPU 옆 DRAM 수요가 같이 폭증한다.
이유는..
agentic AI는 단순 추론이 아니다.
여러 단계의 추론 체인을 weeks-to-months 단위로 유지해야 한다.
즉.. KV 캐시, 추론 상태가 한 세션이 끝나면 사라지는 게 아니라 몇 주에서 몇 달 단위로 메모리에 상주해야 한다.
그 결과 standalone agentic orchestration 서버 한 대당 DRAM이 1.5~3TB.
모건스탠리 표현을 그대로 옮기면 "2년 전엔 상상 불가능했던 수준."
그리고 한 가지 더...
아직 컨센서스 DRAM TAM 모델에 반영되지 안 됐다.
즉.. 베어가 보는 "2027~2028 컨센서스가 위태롭다"는 시나리오의 정반대다.
컨센서스가 오히려 과소평가라는 얘기다.
-
4. End-market fungibility - 메모리 부족이 PC 스마트폰까지 번진다.
이 부분도 중요한 인사이트다.
DRAM 비트는 PC, 서버, GPU 서버 간 호환된다.
같은 칩을 어디든 꽂을 수 있다.
그래서 삼성, 하이닉스, 마이크론이 end market별로 capacity를 따로 잡지 않는다.
증분 capacity는 가장 빠르게 성장하는 시장으로 성장한다.
무슨 뜻이냐...
AI 서버 수요가 폭증할수록 PC, 스마트폰, 게임기에 갈 메모리 비트가 줄어든다.
때문에 소비자 메모리 가격도 강제로 따라 오른다.
보고서 제목이 'Chipflation'인 이유가 여기 있다.
AI 발 메모리 인플레가 IT 전반으로 번진다는 것이다.
소비자 입장에선 갤럭시, 아이폰, 노트북 가격이 올라가는 식으로 체감된다.
메모리 회사 입장에선 모든 영역에서 가격을 받을 수 있는 환경이 만들어진다.
-
5. TeraFab 스케일 - 한 시설이 DRAM WFE 14조
가장 충격적인 숫자..
연 12GW 컴퓨트 capacity를 위한 시설 한 곳 = 1년차에 반도체 장비 33조 달러 필요.
이 중 DRAM 장비만 14조 달러.
보수적인 3GW 시나리오에서도 한 시설에서 DRAM 장비 수요만 3.5조 달러/년이다.
지금 미국, 유럽, 중동, 한국, 일본, 인도에서 AI 데이터센터가 동시에 진행 중이다.
오늘 네이버-엔비디아 뉴스가 증명.
모건스탠리 표현 그대로 옮기면..
"AI DC 배포는 단순 칩 수요 스토리가 아니라 풀스택 메모리 capacity 스토리"
-
그래서 밸류에이션은?
마지막에 모건스탠리가 던진 숫자가 있다.
만약 70% LTA 커버리지(메모리 매출의 70%가 장기계약 기반)에 10x P/E를 적용하면, 2027년 삼성, 하이닉스 P/E가 현재 5x에서 8.5x로 확장 가능하다고 본다.
70% 업사이드다.
지금 시장은 메모리 회사에 사이클 디스카운트를 그대로 반영하려 한다.
모건스탠리는 디스카운트가 LTA 시대엔 잘못된 가격이라고 본다.
This excellent analysis sparks a lot of second-order thinking.
@elonmusk team have shown that their executions are the best in business.
But my view is the ultimate beneficiaries may be the neoclouds, which offer superior value once their scale comes online. $IREN $NBIS $WYFI
$SPCX
Two deals with $70 billion in headline value signed within 30 days of each other.
MY TAKE
> SpaceX built data centers faster than anyone in history.
> Grok failed and the talent left. Musk admitted it needed a rebuild. Renting the GPUs out is a rational pivot.
> The pricing on both deals is well above market. I think scarcity explains part of it. IPO dynamics explain the rest.
> 90-day termination clauses make these contracts cosmetically large but structurally fragile.
> Anthropic’s deal probably sticks for 1-2 years. Google’s might not survive 2027 once their TPU buildout catches up.
> Orbital compute is a real long-term idea wrapped in a short-term marketing need.
THE DEAL MATH
> Anthropic deal: $1.5B/month for 220,000 GPUs.
That’s ~$9.32 per GPU-hour. Market rate for H100s on neo-cloud: $2-3/hr. On AWS: $6-8/hr.
Even adjusting for the H200s and GB200s mixed in, and the full-stack nature of the deal (networking, cooling, power, ops), Anthropic is paying 3-4x the open market rate.
> Google deal: $920M/month for 110,000 GPUs.
That’s ~$11.47 per GPU-hour. Even more expensive than Anthropic’s deal on a per-GPU basis.
I think the pricing tells you everything about scarcity. Nobody pays 3-4x market unless there’s no alternative.
But maybe the pricing tells you something about motivation. Google has a $12B-to-$1.75T unrealized gain on its SpaceX investment. Overpaying for compute by a few billion is a rounding error next to the IPO upside?
THE HARDWARE
150,000 H100s + 50,000 H200s + 30,000 GB200s.
~300MW. A former Electrolux factory in South Memphis.
Built in 122 days and doubled in 92. The build speed is like nothing else in the industry. Anthropic is getting ALL of it.
> Colossus 2 (Google’s deal is here likely):
110,000 GB200 GPUs being brought online.
Targeting 1GW+. That would make it the first gigawatt single datacenter in the world.
Power comes from a fleet of gas turbines via Solaris Energy. ~460MW installed and growing. 1,140MW in the order book.
I think the build speed is what people keep underestimating. Everyone else is waiting 15-18 months. SpaceX did it in four.
THE TERMINATION CLAUSES
Both deals: 90 days’ notice to terminate after the initial ramp. No multi-year lock-in and no penalty.
I think this is the most important detail nobody is talking about. These are $70B in combined headline contract value, but they’re functionally quarter-to-quarter commitments.
If Anthropic finds cheaper capacity (and they’re building 5GW with Amazon, 5GW with Google/Broadcom, $30B on Azure), they walk or renegotiate. If Google’s own TPU capacity catches up, they walk or renegotiate.
$SPCX $NBIS $CRWV $GOOGL $ASTS $RKLB
I mostly agree with what you wrote. We can estimate ARR, then simply multiple by proper P/S (say 5 is mid point). The deeper valuation would be EV/EBITDA which requires more assumption on debt and EBITDA margin, and apply 15x multiple. What we don't know is numbers of share counts, but we can assume they are double as base/conservative case.
But when ARR is 10x, and share counts are 2x, it give us 5x potential. Multiple depends on market condition and relative valuation, but I think 5x is a quick napkin math for $IREN for 5GW.
But again, the stock won't go up in straight line and market can crash for several macro reasons, so the strick risk management should be applied along the journey, and we may not need to chase the last dollar.
SK HYNIX, SAMSUNG, MICRON
$000660 $005930 $MU $DRAM $EWY
I started accumulating memory stocks since October 2025.
I added more during March and April when market crash amid war escalation.
I bought a bit more yesterday.
I’ll probably buy again today, and in the days ahead.
So far, I haven’t sold a single share.
That said, I already have an "exit plan" in place. 🚪
And I’m fully willing to absorb losses on my latest purchases if the market corrects.
In fact, if the market pulls back in June, I’ll happily buy more again.
I believe the rerating has only just begun, from being viewed as purely cyclical to becoming increasingly structural, or extended cyclical, if you will.
And frankly, a 5x P/E multiple on just memory stocks makes little sense to me when other parts of the AI supply chain trade at 50–100x earnings while still relying on the same AI cycle, even if to different degrees.
Thanks to the 𝕏 community - especially the highly insightful and intellectually generous people from Korea, Japan, Taiwan, and even China 🇰🇷🇯🇵🇹🇼🇨🇳 - my understanding of the industry and these companies has deepened as my allocation increased.
That’s where my conviction comes from, and I’m truly grateful for. 🙏
However, I think things could change dramatically by 2027, when everything peaks, if not earlier.
But as we should know, the stock market is always 6–12 months forward-looking, so I need to stay prepared, not complacent.
I believe this quarter may be the last great opportunity to accumulate.
Next quarter may be the time to sit, watch, and do nothing.
The trimming process may begin once the rerating becomes consensus and everyone starts believing the cycle will last forever.
That’s when risk starts to far outweigh reward, holding” turns into “trading,” from "buy the dip" turns into "trim the dip".
You can leave the last dollar on the table because you’ve already taken your principal off the table, and let the profits ride for free.
Memory was, is, and will always be cyclical, just like every other manufacturing industry.
If you don’t have a predetermined "exit plan" in place, the market will eventually choose your "exit price" for you.
"Sam had invested 500 million dollars in Anthropic that would be worth $77b today"
Only problem is.. that $500m was not his money. It's FTX clients' deposit money.
FTX was literally bankrupted with LUNA, so he had no money left. Or, he had to liquidate his Anthropic's stakes to pay margin that he lost during Luna crash.
There is no what if.
$DELL
I was late to the party and just recently accumulated the stock after President Trump said in mid-May, “Buy $DELL.”
I hate to chase the hot stock, but this wasn’t a blind nor FOMO move.
$DELL is President Trump's “buy American" story. Before his speech, $SMCI, one of $DELL’s key competitors, was hit by U.S. regulators' investigation.
This $DELL story also directly related to $IREN.
Days after, $NVDA "indirectly" secured access to $IREN’s 5GW power portfolio to build flagship AI DGX factories.
$IREN had already partner with $DELL as its core GPU infrastructure: $5B in 300MW $MSFT deal, plus another ~$1.6B 60MW $NVDA deal.
With 4–5GW of $IREN "open" capacity, my "connect-the-dots" view is that the market still hasn’t priced in the possibility of $50B–$75B of infrastructure business flowing from $IREN to $DELL over the next 3-4 years.
Fundamentally, $DELL’s ISG (Infrastructure Solutions Group) is growing over 100% YoY and generating 3x the margin of CSG (Consumer Solutions Group), which remains relatively flat and stable. But with AI, PC upgrade cycle will be coming, sooner or later.
When I looked at 2027 EPS consensus estimates of ~$12, it felt far too low to me. My own "conservative" estimate was $14–$15.
At a 30x P/E multiple, justified by AI and U.S. premium, the stock "could" reach $450 vs the then price of $240, or PE 16x, criminally cheap by almost any standard.
And here we are now: after-hours, $DELL jumped to $437 - an astonishing +80% move in just a matter of days.
And after earning release, my 2027 EPS at $14-$15 looks low now. I think it's highly likely that $DELL could achieve 2027 EPS = $20.
At PE 30, you do the math.
Our exceptional first quarter performance reflects strong, broad-based execution across our business and continued momentum in AI demand.
✅ Record revenue $43.8B, up 88% YoY
✅ Record diluted EPS $5.24, up 282% YoY
✅ Record non-GAAP diluted EPS $4.86, up 214% YoY
✅ Record Q1 cash flow from ops $4.1B
Learn more about our Q1FY27 #earnings results: https://t.co/HeRz0Gw36T
w/ @JClarkeatDell