AnyoneBut

🚨지금 주식 사지마. 경고. -JP Morgan CEO 재이미 다이먼- 니가 그 6월에 코스피 15,000 콜했던 새끼구나?

[비트마인 26.07.13 8-K 요약] ■ 총보유 자산 약 113억 달러 - ETH $1,820 기준 - 현금·시장성 유가증권·문샷 포함 - 기존 111억 달러에서 약 2억 달러 증가 - ETH 추가 매수와 ETH 가격 상승에 따른 평가액 증가가 주요 원인 ■ 5,770,038 ETH 보유 - 기존 5,742,237 ETH → 5,770,038 ETH - 총 ETH 공급량 1.207억 개 기준 약 4.78% - 5% 목표 약 95.6% 달성 - 회사 표현 기준 약 96% 달성 - 5% 목표인 6,035,000 ETH까지 264,962 ETH 남음 ■ 지난주 27,801 ETH 추가 매수 - ETH $1,800~1,820 적용 시 약 $50.0M~$50.6M 규모 - 기존 주간 매수량 42,197 ETH보다는 감소 - 다만 2026년 들어 꾸준한 매수 기조는 유지 중 ■ 현금·시장성 유가증권 약 $482M - 기존 $527M에서 약 $45M 감소 - ETH 매수 추정액은 약 $50M 수준이므로 단순 계산상 약 $5M~$6M의 차이가 발생 - 운영비·스테이킹 수익·유가증권 평가변동 등을 무시하면 소규모 ATM 조달 가능성이 있음 ■ 약 0.00950 ETH / 주 추정 - ATM 반영 추정 보통주 약 6.074억~6.075억 주 기준 - 기존 0.00946 ETH/주 대비 약 0.4% 증가 - 이번 주 ETH 보유량은 약 0.48% 증가한 반면 추정 주식 수는 약 0.05~0.06% 증가 - 소규모 ATM이 있었다고 가정해도 주당 ETH는 증가한 것으로 추정 ■ 4,917,189 ETH 스테이킹 - 기존 4,879,157 ETH에서 38,032 ETH 증가 - 이번 주 신규 매수량 27,801 ETH보다 더 많은 물량을 추가 스테이킹 - 전체 보유 ETH의 약 85.2% - 기존 약 85.0%에서 스테이킹 비율도 소폭 상승 ■ MAVAN / 스테이킹 수익 - 7일 기준 연환산 스테이킹 수익률 2.70% - 기존 2.68%에서 소폭 상승 - 현재 연환산 스테이킹 수익 약 2.42억 달러 - 전량 스테이킹 시 예상 연환산 보상 약 2.84억 달러 ■ 5% 목표 달성 여력 - 남은 264,962 ETH를 $1,820에 매수하려면 약 $482.2M 필요 - 현재 현금·시장성 유가증권은 약 $482M - 현재 ETH 가격이 유지된다면 보유 현금만으로 사실상 5% 목표를 달성할 수 있는 수준 ■ 기타 자산 - 206 BTC - Beast Industries $180M - Eightco/ORBS $69M ■ Russell 1000 / 거래 유동성 - 2026년 6월 26일 Russell 1000 Large-cap Index 편입 - 기관투자자와 패시브 자금의 접근성 확대 - 7월 10일 기준 5일 평균 일일 거래대금 약 $475M - 미국 상장주 약 5,704개 중 215위 수준 ■ BMNP 우선주 - 9.50% Series A Perpetual Preferred Stock - 총 350만 주, 공모가 주당 $80 - 순조달액 약 $273.8M - NYSE에서 BMNP 티커로 거래 중 - 조건에 따라 주간 배당 지급 예정 ■ BMNP 배당 커버 - BMNP 연간 명목 배당 부담은 약 $26.6M - 현재 연환산 스테이킹 수익 $242M은 배당 부담의 약 9.1배 - 운영비와 세금을 제외하기 전 기준이지만 우선주 배당은 충분히 커버 가능한 구조 ■ 결론 - 5% 목표까지 이제 약 26.5만 ETH만 남음 - 현재 필요한 매수금액과 보유 현금성 자산이 사실상 일치 - 이번 주에도 소규모 ATM 가능성은 있지만, 추정 희석률은 약 0.05~0.06% $BMNR

주변 분위기를 보면 추매를 꺼리고 본전 매도를 기다리는 개인들이 많군요. 좋은 센티먼트 입니다.

$LAZR ETF 6월 30일에 상장했음 $sive 4,448주 보유하고 있음 루멘텀, AXT, AAOI, Win Semi, Soitec, IQE 등등 보유 중

Bitmine executed a 5.5 million common stock buyback in the past week.

The fact that this post became the most successful one of my entire journey in terms of likes should make one think. To me, it confirms that $NBIS is currently the most crowded trade on X within my feed. Everyone seems to agree that neoclouds are the winners from Kimi. And that is usually the moment when it becomes interesting to look at the other side of the trade. Demand is not the problem for the neoclouds. It was already huge before Kimi, and you can argue that open models push it even higher, because every new deployment is a new compute customer. The real question marks are financing and execution, and those get bigger, not smaller. Neoclouds borrow against long term offtake contracts. Here it helps to be concrete about who is actually on the other side of those contracts. Nebius does not primarily sell to small startups, its anchor customers are Microsoft and Meta, backed by multi year deals worth tens of billions. That is a strength, because it makes the financing bankable, but it is also the hidden dependency. The transmission does not run straight from the frontier labs, it runs through the hyperscalers. If the US frontier ecosystem cannot hold its high valuations, the labs cost of capital rises, and the hyperscalers that host and fund them start to rethink their own capex. And their offtake is exactly the collateral the neocloud financing rests on. So a repricing that begins at private labs nobody can short ends up landing on the balance sheet of a listed neocloud, one layer removed. On top of that comes the quality of the new demand. Many small, short term customers hosting a Chinese open weight model are no substitute for a few large, bankable contracts. More GPU hours, yes, but not the kind of collateral you can build a senior secured facility on. So let us be clear about it. Life is easier for US neoclouds when two or three American frontier labs lead the market, not Chinese open source models. In other words, the level of uncertainty has increased, and markets generally do not like uncertainty.

Wow, apparently there’s claims ~3.4% of the adult population in Korea faces liquidation risk? “As of July 13, over 1.2 million leveraged retail traders' accounts hit margin call lines, with 320,000 to 360,000 accounts force-liquidated by brokers.” (There were other sources claiming this data came from Goldman Sachs desks rather than SK financial supervisory service data) Regardless, just evidence recent crash was accelerated by liquidation cascades/deleveraging rather than fundamentals:

$sive 인도에서도 뭐 하고 있나봄 - RF and mmWave IC Layout Engineer Wireless 분야인 거 같은데 인도에서도 좋은 소식 좀 가져와보자!

However, all of these assumptions begin with the same premise: that a flood of new supply will arrive in 2028 and cause memory prices to collapse. The analysis I read made the following argument. Since the 1980s, memory prices have never fallen because demand declined. Demand continued to grow during the PC era, the smartphone era, and the cloud era. Every major collapse in memory prices has been caused by supply. I agree. If Big Tech eventually lacks the financial capacity to absorb all the available memory, a new wave of consumer AI innovation will emerge to absorb it instead. Perhaps memory prices will first need to fall before that consumer AI innovation can take place. But that is a chicken-and-egg problem, not the central issue we are discussing here. The analysis then went on to say this: everyone who has ever argued that “this time is different” in the memory industry has eventually been proven wrong. Yet there are several reasons why this time might actually be different. One is the growing use of LTAs, although many people remain skeptical of them, so I will set that aside. The more important point is that in today’s AI industry, as P falls, Q increases by much more than P declines. A paper by Zhang and Zhang titled The Economics of Digital Intelligence Capital was posted on arXiv in January 2026. It estimates the price elasticity of demand for AI tokens at 1.42. In other words, a 1% decline in price leads to a 1.42% increase in volume. The crucial point is that the elasticity is greater than one. When that is the case, falling prices can actually increase total industry revenue. The paper models a scenario in which API prices are cut in half. Token consumption does not rise linearly; it accelerates in a convex fashion as developers begin adopting more compute-intensive inference architectures. What matters here is that elasticity does not prevent a price collapse. It reduces the damage caused by that collapse. Suppose DRAM selling prices fall by 30%. In the old world, shipment volumes would barely increase. Revenue would then fall by 24%, and because costs would remain largely unchanged, margins would collapse from 60% to below 25%. That is what happened in 2019, when Samsung Electronics’ operating profit fell by 52.8%. In an AI-driven world, however, an elasticity of 1.42 means that volume rises by 42%. Assume that production costs also decline by roughly 15% through process migration. Revenue then remains almost unchanged, margins fall from 60% to 36%, and profit declines by only 15%. Negative 52.8% versus negative 15%. That difference lies at the heart of whether memory companies deserve to trade at more than five to six times earnings.

NVIDIA argues that continuous post-training will become a central compute workload in the agentic AI era and proposes “intelligence per dollar” as the broader metric for measuring its economic efficiency. Reinforcement-learning-based post-training repeatedly generates model rollouts, evaluates rewards and updates model weights. NVIDIA is positioning NeMo Gym and NeMo RL as infrastructure for coordinating environments, reward verification and distributed training at scale. NVIDIA Nemotron 3 Ultra, a 550-billion-parameter MoE model, scored 71.7% on SWE-bench Verified. NVIDIA also claims Vera Rubin can train the largest models using one-quarter as many GPUs as the Blackwell generation, although this remains a company-provided platform performance claim. Prime Intellect reported that Vera CPUs delivered 30% higher average throughput than alternative x86 architectures in selected RL sandbox workloads. Perplexity’s stack uses hundreds of GPUs and an RDMA-based engine to synchronize trillion-parameter model weights between training and inference nodes in under two seconds. The broader message is that lowering inference cost per token is no longer sufficient. AI platforms must also reduce the cost of continuous post-training so that models can keep improving as tools, production environments and user requirements change. --> The key point is that GPU demand does not disappear once pretraining is complete. AI agents must be continuously post-trained as tools, environments and failure cases change, requiring large-scale inference rollouts and repeated model-weight updates. Even if platforms such as Vera Rubin can complete the same workload with fewer GPUs, lower costs may encourage companies to post-train more models, across more environments and at a higher frequency. Greater hardware efficiency may therefore expand the overall training and inference cycle rather than reduce aggregate compute demand. AI infrastructure demand is shifting from one-time pretraining toward a recurring model in which inference and continuous post-training run together. That supports sustained demand not only for GPUs, but also for HBM, networking, CPUs, storage and broader data-center infrastructure.