From atoms to agents: Physical + Cognitive Compute Stacks in Era / Material Science Tech & Insights → Highlights For business: DM, [email protected]
Joined April 2016
- Tweets17.3K
- Following458
- Followers4K
- Likes62.3K
Pinned Tweet
Thanks for 4k followers!
3k 팔로워를 달성했던 시기가 4월 30일
4k까지 정확히 5개월이로군요 👀✨️
저는 펀더멘탈이나 가설을 말하는,
자극적이라기 보단 다소 슴슴한 계정이지만
꾸준히 찾아주셔서 감사할 따름입니다
매번 저와 소통 및 놀아주시는 분들
그리고 어제 저녁부터 4천 팔로워를 달성할 수 있게 도와주신 분들께도
정말 진심으로 감사드리고,
앞으로도 잘 부탁드립니다 🙇♂️
슈포 (Superposition) retweeted
I respect Jukan, but I feel a little frustrated whenever I see this kind of argument.
You can build any scenario around plausible assumptions, and yes, of course, the outcome will follow from those assumptions. If you assume demand collapses, customers accumulate inventory, and contract terms get renegotiated, then naturally, memory companies’ revenue and profits will decline.
But honestly, this feels like the perspective of analysts who treat memory as just a commodity you manufacture and sell, without properly understanding the technological changes taking place in custom memory.
Who exactly are these customers we are talking about when we say they could cancel LTAs or defer deliveries? They are the very customers working with memory companies to design the next generation of HBM and 3D stacked DRAM.
After jointly designing and validating these products, can customers really cancel those volumes so easily or push deliveries out by one or two years? How would they manufacture their own products, which were designed around that memory? What happens to the launch and system deployment schedules they have already coordinated? Switching suppliers also raises questions of design compatibility and requalification. If you are going to argue that contracts can be adjusted, you also need to explain what that decision would mean for the very customers canceling them.
This is one of the biggest reasons I keep saying that memory is no longer the same cyclical industry it used to be. The parties to these LTAs are already tied together from the design stage. But if you still view memory simply as a commodity where rising prices lead to more production, and inventory buildups lead to price cuts or canceled orders, you will inevitably miss this shift in the technical and business relationships.
One more thing. These arguments about a collapse in memory demand seem to assume that planned AI data center investments will be canceled on a massive scale, or that AI infrastructure expansion will slow so substantially that memory demand falls sharply. What interests me more than whether contracts can be adjusted is why you expect demand to weaken that dramatically in the first place.
I think it would be more realistic to discuss directly whether the AI era we expect will actually fail to materialize, and whether much of the AI infrastructure investment currently planned will disappear. If that is your outlook, I can at least understand why you would expect memory demand to collapse and customers to cut volumes even for products they have jointly developed.
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.
슈포 (Superposition) retweeted
TSMC has denied reports that it is exploring 3D DRAM production at Nvidia’s request. It says it remains focused on logic chip manufacturing.
$TSM
$NVDA
Tom’s Hardware reports the response. A TSMC spokesperson said the company focuses on its core business at all times, which is dedicated integrated circuit foundry work. Its mission is to be the trusted technology and capacity provider of the global logic integrated circuit industry for years to come. The statement effectively rejects the earlier report from The Information. That report, citing two people familiar with the matter, said TSMC had begun drafting plans for 3D DRAM manufacturing after a request from Nvidia, its largest customer. The effort was said to be led by Hsu Kuo-chin, TSMC’s vice president of advanced packaging and a former head of Micron’s Taiwan operations.
TSMC works with DRAM producers to package memory into system-in-packages for customers. It lacks the expertise and production capacity to make DRAM, let alone 3D DRAM. Even major memory makers are years away from 3D DRAM. Pursuing it would not address current shortages for at least five years. Micron, Samsung, and SK Hynix are all working on 3D DRAM. Micron’s strategy includes at least four more planar DRAM process technologies before 3D DRAM becomes a reality. Making DRAM would require TSMC to develop manufacturing processes distinct from its logic production. 3D DRAM also needs different techniques to build vertically stacked memory cell layers while controlling defects and maintaining acceptable yields.
A denial is not a finished supply plan.
If TSMC has already stated that it stays focused on logic foundry work and that 3D DRAM remains years from market viability even at the major memory makers, does the next constraint stay on how the existing DRAM and high-bandwidth memory suppliers scale output through 2027 and 2028, or on whether Nvidia and other AI chip designers still need additional long-term memory sources beyond the current three?
슈포 (Superposition) retweeted
얹혀가는 김에 저도 하나 올립니다 ㅋㅋ
김크립토님 말씀처럼 오케스트레이션의 핵심은 전이와 검증입니다. 그런데 이 두 가지는 멀티에이전트 없이 세션 하나에서도 똑같이 무너집니다.
작업 분배를 잘 했느냐? (전이)
부패: 문서와 메모리가 실제 코드와 어긋남 → /doc-drift
오염: 외부 스킬, 플러그인, MCP를 검토 없이 들여옴 → /integration-intake
분산: 같은 사실이 파일 여러 곳에 흩어지고 서로 모순됨 → /ssot-audit
혼란: 어디까지가 이번 작업인지 모름 → /scope
충동: 하다 보니 범위 밖 파일까지 고침 → /freeze
분배된 작업을 진짜 잘 한 게 맞냐? (검증)
목표를 잠그고 PLAN→DO→VERIFY 루프를 강제함 → /goal-lock
검증이 스스로를 확인하는 순환 구조인지 31개 패턴으로 점검함 → /eval-leakage-audit
푸시 전에 시크릿 스캔과 리뷰를 함 → /pre-push
세션 사이 전이도 있습니다.
세션이 끝날 때 /session-checkpoint로 남기고, 다음 세션은 /session-start로 받습니다. 사람이 매번 "어디까지 했더라"를 설명하지 않아도 됩니다.
스킬 22개는 각각 따로 써도 되고, 순서대로 쓰면 프로젝트 셋업부터 리뷰, 거버넌스까지 덮습니다. Claude Code, Codex, Cursor에서 씁니다.
github.com/AlexZio00/soverei…
자자 오케스트레이션에서 중요한건 컨텍스트 전이 그리고 검증입니다.
결국 오케스트레이터도 에이전트거든요.
1. 작업 분배을 잘 했느냐? (전이)
- 부패, 오염, 분산, 혼란, 충동
2. 분배된 작업을 진짜 잘 한게 맞냐? (검증)
- 재현 가능성
이 두가지를 잘 해야합니다. 그리고 이 두가지를 잘 하는 방법은 무엇이냐? 제 오픈소스를 보는것입니다. (갑분 광고 ㅈㅅ..)
대부분의 작업에 멀티에이전트를 쓰지마십시오. 모델들의 성능이 좋아지면서 멀티에이전트의 가치는 점점 떨어져가고 있습니다. 하지만 밀도있고 규모있는 소수의 작업을 위해서라면 멀티에이전트는 고려해봄직합니다.
밀도있는 workflow를 만들어주고 전이, 검증을 도와주는 오픈소스!!! github.com/AgwaB/pi-workflow
pi 생태계 한정이지만 어떤 provider든 다 쓸 수 있게 개선 중입니다. (github.com/AgwaB/agent-workf…)
율무님한테 얹혀갑니다 ㅋㅋㅋ
창발을 일으키는 부분을 오히려 역이용하여 아이디어로 활용하기도 합니다
'AI가 창발로 만드는 내용이, 가능성이 아예 없진 않을 수도 있다'라고 생각하기도 하고
특정 분야의 전문가로서 보는 view와는 다른
편견이 덜 한 시각을 가지고 있다... 라고도 생각하기 때문입니다.
![[AI를 보는 렌즈들 #23] 귀납과 연역](https://pbs.twimg.com/media/HWe4nLBaEAA-pvz.jpg)
슈포 (Superposition) retweeted
No matter how revolutionary AI turns out to be, there’s no reason every dollar poured into it should survive. If it really transforms productivity, some companies will make a fortune while others lose billions, drown in debt, or get bought out for a fraction of what they were once worth. That’s how markets work.
Except now we’re told AI is the new normal, that markets have learned from past crashes, and that investors and policymakers know enough to prevent another collapse. Maybe they do. But it’s quite a leap from being better prepared for a crisis to believing the market no longer needs to go through a serious correction. And then there’s China, which seems to come up whenever someone questions how much money is being poured into AI. America can’t afford to fall behind, we’re told, as though that settles the question of whether the investments themselves make financial sense.
I’m optimistic about what AI can accomplish, but a company that can’t even earn its cost of capital and needs fresh debt just to stay alive doesn’t get a pass on its financial losses simply because the technology might change the world.
And if markets have learned so much from past crashes, why is restructuring still treated as something we should avoid? We’ve spent generations operating under a system where companies fail, assets change hands, and investors take losses, even when the technology really does change everything.
If the new normal means we can skip all that, maybe we’re talking about a different economic system altogether.
슈포 (Superposition) retweeted
GOOGLE $GOOGL WILL SPEND $17B ON A NEW DATA CENTER IN WYOMING
The site in Laramie County will be Google's first in the state, per Bloomberg. Google says it will pay for all the power the data center uses.
The data center will be air-cooled to limit water use.
슈포 (Superposition) retweeted
中 국유기업도 부채 위기…4130억위안 빚에 채권시장 출렁 [차이나 워치]
중국의 부동산 침체가 지방 국유기업의 신용위기로 번지면서 채권시장이 크게 흔들리고 있다. 광둥성 주하이시 최대 국유기업인 주하이화파그룹의 대규모 부채와 유동성 문제가 불거지자 투자자들이 채권 매도에 나서면서다.
헝다그룹 등 민간 부동산 개발업체의 채무불이행에 이어 지방정부가 소유한 대형 국유기업까지 자금난에 직면했다는 점에서 시장의 경계심이 커지고 있다. 중국 부동산 시장이 내년부터 회복할 수 있다는 전망이 나오고 있지만 장기간 누적된 부채 부담이 금융시장 불안 요인으로 부각됐다는 분석이다.
슈포 (Superposition) retweeted
Korean media report that JNTC, a Korean glass substrate maker, has been registered as a TGV glass substrate supplier to NVIDIA.
thelec.kr/news/articleView.h…
