Puppet @2xpuppet

Kioxia director and Bain rep, Yuji Sugimoto, on whether the stock is fairly valued: "As a current director, it's difficult to comment, but all I can say is that the stock price has risen on the back of improved performance, and the projected P/E ratio is still at a level where it might or might not reach double digits." "[Cyclicity] will likely continue. However, the demand for AI has raised the baseline by a notch, so it will be a cycle on a plateau rather than on flat ground. I don't think it will ever fall back to flat ground."

이 과정에서 가장 중요한 하드웨어 요소는 모델의 기억에 해당하는 컨텍스트 창인 것으로 확인됐다. 컨텍스트 용량이 100만 토큰인 모델은 20만개의 모델을 상대로 테스트 전 구간에서 우위를 점했다. 최종 12시간 시점에서의 성능 상한선도 100만 모델은 97.8점, 20만 모델은 81.5점으로 갈렸다. 이는 장기 업무를 수행할 때 과거의 실패 로그와 빌드 오류를 잊지 않고 기억하는 능력이 학습의 핵심이라는 의미다.

대략적으로 1GW 급 AIDC 하나가 아이폰17 pro 3천만대 정도의 dram을 소모함 매년 미국에서만 10GW 이상의 AIDC가 새로 들어서니(매년 2배 이상 성장) 아이폰 3억대 분량의 dram 수요가 생기는 것 참고로 아이폰 라인 전체 판매량은 연간 2.3억대 도대체 수요 계산 못 하는 사람이 누구냐?

This last week JEDEC announced a new standard for SPHBM4: Standard Package High Bandwidth Memory (JESD330-4). It utilizes the same DRAM stacks as HBM4, but swaps in a different buffer die. The goal? Enable HBM assembly in standard packaging and break the AI Advanced Packaging bottleneck. (1/6)🧵

$META Chief AI Officer Alexandr Wang says Meta’s current model in training codenamed Watermelon is matching GPT-5.5. He also said Watermelon uses an “order of magnitude” more compute than Avocado.

keyboard_arrow_left Previous keyboard_arrow_right Next

Bernstein estimates that SK hynix’s DRAM gross margin in Q2 will reach 90.9%…

UBS raises its DRAM and NAND price forecasts DRAM prices are now expected to rise 32% QoQ in Q3 and 18% QoQ in Q4. NAND prices are expected to rise 30% QoQ in Q3 and 12% QoQ in Q4.

Knockouttrader @knockouttrader

2 days ago

UBS raises its memory pricing forecasts again, arguing DRAM and NAND shortages are likely to persist well into 2028 despite ongoing LTA negotiations. The latest supply-chain checks continue pointing out that demand is still outpacing supply. UBS increased its DRAM contract

3 14 91 216K 28

>> TrendForce: AI server demand underpins memory prices… Q3 DRAM +13 to 18%, NAND +10 to 15% QoQ increase expected • According to TrendForce's latest memory price survey, the DRAM market in Q3 2026 is expected to continue experiencing severe supply shortages across the board. That said, softening demand in consumer applications and a high base effect will moderate the pace of the increase somewhat, with DRAM contract prices projected to rise +13 to 18% QoQ. • For NAND Flash, AI inference demand and large scale data center buildouts continue to drive the bulk of demand. However, with contract prices already at record highs and consumer market demand slowing, customers' capacity to absorb further price increases has reached its limit. As a result, TrendForce forecasts NAND Flash contract prices to rise +10 to 15% QoQ, with the magnitude of the increase narrowing noticeably versus prior quarters.

@paurooteri アメリカ半導体産業協会：メモリ供給難は政府の規制ではなくLTAでのみ解決可能

VCSEL CPO is gaining momentum.

@TBU12345678 The whole photonics trade will blow up if VCSELs are adopted for scale up 🤣🤣🤣

My 2 cents: First of all, excess capacity is a “if”. Nevertheless, it is not a new news, as Meta is likely responding to investor pressure by leasing part of its compute capacity, if any, as a way to help ROE amid high rental px. Please note that the transformation into Neocloud usually leads to higher capex, and we mentioned in our earlier report of “SpaceX’s Move to Neocloud on June 22” Recent developments suggest that Meta remains compute-constrained. News reported that Google restricted Meta’s access to Gemini compute due to high demand. Meanwhile, Meta recently signed a new $21 billion AI compute agreement with CoreWeave for AI capacity. #Meta #NVDA

@

57 years ago

0 0 0 0 0

$META is as likely to cut capex as $AAPL is to get CXMT approval

$META is reportedly developing a cloud business to sell access to excess AI compute, per Bloomberg. The internal initiative is called Meta Compute. The plans being considered: AI model access hosted on Meta infrastructure, similar to AWS Bedrock Raw AI compute capacity, closer to CoreWeave Developer access to Meta’s data centers, chips, and models

Bad news for neos & other hyperscalers

Zephyr @zephyr_z9

4 days ago

LMAO Zuck finally takes the L

65 40 2K 339K 106

*META IS BUILDING A CLOUD BUSINESS TO SELL EXCESS AI COMPUTE

BYTEDANCE OPENS A NEW $39 BILLION DATA CENTER IN BRAZIL, EXPANDING BATTLE FOR AI CONTROL BETWEEN THE US AND CHINA.

Vertically integrating the full inference hardware stack is extremely hard. Hear @UbertiGavin and @robertwachen share three years of battle stories leading into today's launch

Etched @Etched

5 days ago

We're coming out of stealth. We've built our first racks after a successful A0 tapeout, $1B+ in customer contracts, and $800m raised. Early customer tests show us achieving SOTA throughput, latency, and power efficiency on inference workloads. Our first racks ship this summer.

616 913 9K 5.9M 4K

With Micron earnings last week absolutely smoking all estimates and putting up truly insane numbers, the timeline is once again asking if this is the top of the memory cycle. I sat down to do the work, and it seems we are in for a much longer shortage than the market believes. A few key numbers: 1. Token demand is projected to 24x by 2030. 2. HBM supply is projected to double by 2030. 3. HBM demand is projected to 5x by 2030. There’s a massive structural supply/demand imbalance in the market, and it’s actually growing not shrinking. Link to read the full free report below.