JPMorgan increased its holdings in Strategy by approximately 36% in the third quarter, and holds around $540 million worth of BlackRock’s Bitcoin ETF.

，散户投资者在 11 月 10 日净买入约 7.3 亿美元英伟达股票，为 3 月以来单日最大买入规模。此前 11 月 9 日净买入约 5.9 亿美元，11 月 11 日再净买入约 3 亿美元。三个交易日合计净买入约 16 亿美元，为至少 12 个月来最大的三日买入。同期，2 倍做多英伟达 ETF NVDL 录得约 4.01 亿美元流入，有望创下 2025 年 11 月以来最大单周流入。另一只 2 倍做多英伟达的 ETF NVDU 录得约 6400 万美元流入，有望创下 2026 年 2 月以来最大单周流入。11 月 12 日开盘后的前 30 分钟，散户买入约 4.55 亿美元、卖出约 3.45 亿美元，净买入约 1.11 亿美元，为有记录以来开盘前 30 分钟的最大净买入，散户正在加码买入英伟达。

Dongcha Beating AI Flash News: Figure AI founder Brett Adcock’s personal AI assistant Hark has launched a new-generation web operation model Handoff2, which is now available to all Hark Pro users. Compared with the previous version, the new model doubles the average overall speed of completing web tasks, with flight booking tasks up to 3.8x faster, and long research tasks even over 5x faster. Hark tested the new model on 478 tasks across 238 websites, and its task completion rate rose from 73% to 82% (a 9-percentage-point increase). Its CAPTCHA handling capability also improved by 80%. According to Hark’s self-test, Handoff2 outperforms Claude Opus 5.5 High in both web operation performance and speed. Compared with the higher-tier Opus 5.5 Max, the accuracy gap is less than 5 points, but its task completion speed is about twice that of Opus 5.5 Max. Hark mainly improved speed by optimizing reinforcement learning training: during training, it penalizes excessive reasoning, repeated operations, and time-consuming steps, teaching the AI to execute multiple operations at once, such as filling multiple form fields consecutively or reusing repeated action combinations. After training, the single-task time of the new model was reduced by 46%, coupled with backend optimizations, leading to an average overall speed twice that of the previous generation. Hark also disclosed that since its official launch in October this year, Handoff has run more than 5.6 million browser sessions and processed over $20 million in shopping transactions.

Beating AI Express (Dongcha) reports: Anthropic’s official status page shows rising error rates for claude.ai, Claude Code, and Claude Console (platform.claude.com). The company is investigating login issues affecting users of Claude.ai and Claude Code, and will provide updates promptly. Currently, claude.ai is marked as a major outage, while Claude Console and Claude Code are partial outages. The outage also impacts Claude Cowork, a service that had previously been suffering from performance degradation. Claude API and Claude services on cloud platforms including AWS, Google Cloud, and Microsoft Foundry remain operational.

JPMorgan Chase’s latest 13F filing shows that as of September 30, it held 848,900 common shares of Strategy, with a market value of roughly $131.3 million calculated at end-of-quarter prices. Compared to 624,400 shares held at the end of Q2, the holding volume rose by 224,500 shares, a 36% increase. In the same period, the firm held 11,408,900 shares of BlackRock’s Bitcoin ETF (IBIT), valued at approximately $540 million, up 9.6% from 10,407,600 shares at the end of Q2. The same filing also lists 1,581,800 common shares of Coinbase, worth around $301 million, plus holdings in publicly traded bitcoin miners including Robinhood, Core Scientific, MARA Holdings, Riot Platforms, Hut 8, and CleanSpark. The filing further reveals JPMorgan holds a large position in Strategy put and call options: the put options correspond to roughly 1,148,500 shares with a notional value of $178 million, while the call options cover about 349,500 shares with a notional value of $54.06 million.

Beating AI News Brief: Tencent’s Hunyuan has launched a new real-time speech model named Hy Realtime, adding online search capabilities for singing and voice conversations. The model has been integrated into Yuanbao Voice Call and Tencent Customer Service. Users can use voice commands to search for the latest information, or request songs by singer, genre, language, and release year. For example, asking it to sing an English folk song released between 2010 and 2020 will prompt the model to select and perform the song based on context. The singing feature is co-developed by Hunyuan and Tencent Music (TME)’s Lingyin AI team. The new model also addresses a key pain point of speech AIs: the so-called "intelligence drop" when speaking. Knowledge and reasoning capabilities originally held by text large language models (LLMs) may be lost when adapted to speech dialogue models. Tencent redesigned its training data production process, stating that the effective retention rate of high-quality text data migrating to speech tasks has risen from 21% to 69% — roughly 3.3 times the previous level. Hy Realtime also enhances understanding of real colloquial speech. When users mumble, speak half-sentences, or change their request mid-sentence, the model can infer intent based on context. When generating responses, it simultaneously organizes text and synthesizes speech, cutting down wait times. Currently, users can experience the model by selecting the "Chuyu" voice tone in the voice call feature of the latest Yuanbao App. Tencent Cloud’s API is set to open soon, with duplex interaction and long-term personalized memory planned as future upgrade directions.

Industry sources and analysts note that Jensen Huang’s sudden resignation from NVIDIA could grant Korean memory chipmakers greater leverage in upcoming supply negotiations. However, the future of various AI projects he personally championed now hangs in the balance. Ahn Ki-hyun, vice chairman of the Korea Semiconductor Industry Association, told reporters in an interview: “This may not be a bad thing for us. Even if other firms fill the market gap left by NVIDIA, those companies will still require our HBM.” Ahn emphasized that NVIDIA’s reliance on a small number of memory suppliers is the core issue, rather than the personal relationship between Jensen Huang and Korean business leaders. “If NVIDIA wants to switch suppliers, it must find alternatives—but where are those alternatives? We don’t have to concentrate our production capacity on a single client; we can spread supply across multiple customers, which is actually more advantageous for us.”