Lookonchain APP

App Store

SemiAnalysis: Kimi K3 significantly reduces KV (Key-Value) transmission bandwidth, but AI network demand will not shrink as a result.

6 hours ago

Independent semiconductor and AI research firm SemiAnalysis has published a report stating that while Kimi K3 uses KDA for roughly three-quarters of its network layers, cutting KV cache transmission bandwidth by up to 10 times compared to a full global attention model, this does not mean the AI network switch market will see a significant contraction. The Kimi K3 boasts 2.8 trillion parameters; even with MXFP4 precision, each forward computation requires around 1.5TB of HBM bandwidth. To maintain reasonable interaction speed while enabling profitable deployment, the model still needs high-bandwidth network connections such as GB300 NVL72 to link a large number of chips, and relies on WideEP for scaling services. WideEP distributes 896 expert models across multiple GPUs, performing token distribution and result merging twice per layer and per forward pass, with over 120 such operations required for a single forward computation. In contrast, KV cache transmission between pre-filling and decoding only occurs once per conversation round, meaning the bandwidth saved by KDA may be far less than the network scaling demands imposed by large-scale expert models. SemiAnalysis adds that more efficient attention mechanisms could also push context lengths from 1 million tokens to over 5 million tokens. Per Jevons’ paradox, efficiency improvements may expand the overall scale of AI usage, further increasing network requirements.

Relevant content

SK Group Chairman: Demand for storage chips is projected to grow by at least 50% to 60% next year, with the supply-demand gap likely to further widen.

According to South Korea’s Maeil Business Newspaper, SK Group Chairman and Korea Chamber of Commerce and Industry Chairman Choi Tae-won stated that driven by the expansion of artificial intelligence (AI), demand for AI semiconductors is projected to rise by at least 60% to 100% next year compared to this year, while overall memory chip demand will also grow by at least 50% to 60%. Choi noted that the additional supply each company can increase next year is very limited, so the supply-demand gap may widen further, with global firms currently scrambling for memory chip supplies. He added that existing expansion plans are still insufficient to meet the rapidly growing demand; SK’s current strategy is “build wherever possible”, but equipment, personnel and construction timelines continue to restrict capacity release. Choi also pointed out that current memory chip prices have deviated from normal ranges, and PC and smartphone manufacturers cannot keep passing cost increases to consumers. Semiconductor enterprises should not limit supply to maintain high prices; even if their profit margins decline, they should expand output and grow the market. Otherwise, excessive prices may attract new competitors and trigger government interventions. He further stated that the AI industry is facing shortages of infrastructure such as GPUs, storage and power, and new bottlenecks may emerge in the future. Regarding the possibility of a stock split for SK Hynix, Choi said the plan has not been fully studied, and adjustments for Korean domestic stocks and American depositary receipts (ADRs) need to be evaluated together.

6 minutes ago

AI hot stocks like NVIDIA have seen increased volatility, with their relative volatility standing at 4 times that of the S&P 500 index.

The Kobeissi Letter stated in a post that the three-week volatility of U.S. momentum stocks relative to the S&P 500 index has surged to four times, hitting an all-time high. This ratio has more than quadrupled over the past several weeks. The momentum stock group includes high-growth tech stocks at the center of the AI boom, such as NVIDIA, AMD, Palantir, D-Wave Quantum, and CoreWeave. By comparison, the ratio peaked at around 2 times during the 2020 COVID-19 market crash and roughly 1.8 times during the dot-com bubble burst. The current level is significantly higher than those periods. Meanwhile, the U.S. momentum stock index has dropped 24% since July, marking its largest monthly decline since the 2008 financial crisis. The stocks that previously performed the strongest are quickly losing market favor.

6 minutes ago

Changxin Technology's IPO winning numbers have been released, totaling 7,702,207.

According to an announcement by Changxin Technology, the company’s initial public offering (IPO) and listing on the STAR Market has released its offline preliminary placement results and online lottery results. The offering price is RMB 8.66 per share, with an initial share offering size of approximately 6.688 billion shares. A total of 7,702,207 winning lottery numbers have been issued, and each winning number entitles holders to subscribe for 500 Changxin Technology A-shares.

6 minutes ago

Renowned trader closes all crypto short positions, resumes buying Bitcoin spot

Renowned crypto trader Doctor Profit announced in a post that he has closed all his cryptocurrency short positions, including Bitcoin shorts established in the $115,000–$125,000 range, another Bitcoin short in the $79,000–$82,000 range, and over 100 altcoin shorts opened in recent months, noting that all these positions have generated significant profits. He also said he has repurchased Bitcoin spot at $64,000, marking his first long-term allocation since September 2025. His plan is to invest 5% of his planned capital daily in spot Bitcoin purchases when the price is in the $54,000–$64,000 range, for up to 20 days total; if the price approaches $54,000, he will increase his buying activity. Doctor Profit pointed out that the current market is showing clear "herd behavior": investors who were previously bullish on Bitcoin up to $150,000 at high levels are now widely waiting for the cryptocurrency to drop to $40,000–$50,000, and are targeting September or October as the bottom of the four-year cycle. When a large number of investors are waiting for the same price level and time point, the market may not move as expected, so he chose to build positions in advance and judges that this cycle’s bottom may arrive earlier than the market’s general expectation. He also cited regulatory clarity, asset tokenization infrastructure, and progress in institutional adoption as the structural reasons for his shift to buying, and retracted his previous prediction that Bitcoin would fall to $40,000–$50,000. However, he still holds all his S&P 500 short positions, arguing that the crypto market has completed a large repricing, while U.S. stock valuations remain elevated.

6 minutes ago

Binance Wallet’s Meme Rush adds new launchpad filter options for Robinhood Chain-based projects including Virtuals Protocol, Flap, and Bankr.

According to official announcements, Binance Wallet’s Meme Rush has added new launchpad filter options for Robinhood chain projects including Virtuals Protocol, Flap, and Bankr, helping users discover more on-chain token opportunities. Users can now track tokens across BSC, Solana, ETH, Base, and Robinhood chains via Meme Rush, with a single feed to stay updated on multi-chain hotspots.

6 minutes ago

Zcash launches Zakura full node, aiming to boost its privacy transaction throughput to 50,000 per second.

Zcash has launched its new full-node client Zakura 1.0.0, the first implementation component aimed at scaling its private transaction throughput from the current ~1 transaction per second (TPS) to payment-network-level capacity. Maintained by Sean Bowe, an early Zcash zero-knowledge cryptography contributor, and Dev Ojha, head of Valar Group, the client operates independently of the Zcash Foundation and is funded via private ZEC donations. Built on the Zcash Foundation’s Zebra client, the team provides an ~11GB blockchain snapshot, allowing new nodes to sync up in two minutes—an approximately 680x speed improvement. Its compatibility mode also replicates the legacy zcashd interface, enabling wallets and exchanges to continue operating after the original client’s maintenance ends on July 18. The team notes that Visa and Mastercard process over 50,000 transactions per second, a benchmark it has set as its minimum target. To reduce the verification burden of large-scale private transactions, Bowe’s Tachyon project is developing recursive proofs, which let nodes validate thousands of proofs with a single proof. Valar Group is building private information retrieval (PIR) technology, allowing wallets to access relevant transaction data without exposing their query content. Zakura is also testing a fast block propagation system that delivers new blocks to all nodes in half a second. Its upcoming testnet is the Ironwood upgrade, set to activate on the mainnet on July 28. Ironwood will use a "rotating door" mechanism to restrict inflows and outflows of the Orchard privacy pool, mitigating risks of fake ZEC entering circulation due to prior zero-knowledge proof vulnerabilities.

6 minutes ago