Xiaoliu BOT
中文EN

X Platform July 27 AI Brief | Kimi K3 Open Source, CXMT IPO, NVIDIA Invests in SSI

Kimi K3 Officially Open-Sourced, Moonshot Releases Multiple Underlying Technologies

Moonshot AI (Kimi) open-sourced the Kimi K3 model weights and technical report as scheduled on the evening of July 27. K3 is a 2.8T parameter MoE model supporting native visual understanding and a 1 million token context window. The company claims a 2.5x improvement in intelligence per unit of compute compared to its predecessor. Multiple bloggers’ tests suggest its performance is on par with leading closed-source models like Fable 5.

Simultaneously, Moonshot open-sourced FlashKDA (a high-performance Delta Attention kernel, achieving up to 2.22× speedup on H20), MoonEP (a MoE distributed communication library), and AgentENV (a large-scale Agent environment runtime system). Platforms including Nebius, Fireworks AI, Baseten, and Modal launched K3 inference services as Day 0 partners.

Notably, OpenAI and Google were reported to have quietly withdrawn from the open AI alliance advocated by Jensen Huang on the same day, forming a stark contrast with Moonshot’s open-source move.

Sources:

ChangXin Memory Tech’s STAR Market IPO Surpasses Tencent in Market Cap; Domestic Lithography Machine Enters Small-Batch Production

ChangXin Memory Technologies (CXMT) listed on the STAR Market on July 27 with an issue price of 8.66 yuan. Its stock surged over 530% at opening, pushing its market capitalization past 3.65 trillion yuan, surpassing Tencent to become China’s most valuable company. As the world’s fourth-largest DRAM manufacturer with a 7.67% market share, its IPO raised 57.9 billion yuan, setting a record for the STAR Market.

On the same day, The Information reported that a Chinese state-backed enterprise has begun small-batch production of domestic immersion DUV lithography machines. Approximately 5 units are expected this year, increasing to 20 by 2027, with plans to deliver them to SMIC, Hua Hong, and CXMT. Multiple bloggers commented that while the machines’ performance still lags behind ASML’s, the domestic substitution path has moved from the lab to the fab for the first time, with market repricing of the semiconductor sector reflecting this expectation. South Korean media also reported that Samsung is considering adopting Chinese-made DRAM.

Sources:

Ilya’s SSI Secures Strategic Investment from NVIDIA, Compute to Expand 10x

After two years of silence, Ilya Sutskever announced on July 27 that his company, SSI (Safe Superintelligence), has entered a long-term strategic partnership with NVIDIA. After reviewing SSI’s highly confidential research, NVIDIA decided to make a significant investment and provide next-generation Vera Rubin systems, aiming to expand SSI’s compute capacity tenfold within the next 12 months. Ilya stated the research now warrants scaling.

Multiple bloggers noted that SSI has not released any models or products in its two years of existence, previously relying mainly on Google TPUs. This shift to the NVIDIA ecosystem signifies that NVIDIA, through investment, is locking in the real chip architecture demands of next-generation models in advance. SSI has raised approximately $2 billion to date, with a valuation around $30 billion.

Sources:

NVIDIA in Talks to Provide $250B Financing Guarantee for OpenAI Data Center

According to The Wall Street Journal, NVIDIA is negotiating to provide approximately $250 billion in financing guarantees to help OpenAI lease a 10GW data center in Ohio being built by SoftBank’s SB Energy. The total project cost, including GPUs, energy, and construction, could ultimately exceed $500 billion, making it the largest data center project announced to date.

OpenAI plans to sign a 20-year lease, with the first phase expected to be operational in 2028. The U.S. government is providing federal land and controlling power resources, while Japan is providing $33.3 billion for 9.2GW of natural gas power generation through a U.S.-Japan investment pact. NVIDIA provides the credit guarantee, banks lend to the project, SoftBank handles construction, OpenAI commits to paying rent for 20 years, and the project purchases NVIDIA GPUs—ultimately, all funds flow back to NVIDIA. Multiple bloggers commented that AI capital expenditure has grown so large that tech companies’ cash flows cannot support it, making credit ratings and financial engineering key competitive capabilities for the next phase.

Sources:

Independent Developers Face AI Ecosystem Squeeze; Agents Emerge as New Delivery Paradigm

Pieter Levels posted about an overall decline in revenue and traffic for independent developers, arguing that “Big AI” is encroaching on the former indie app ecosystem. Wesley and Xiaohu both referenced and discussed this phenomenon. Xiaohu believes that as AI coding capabilities improve and enterprises deeply adopt AI Agents, small SaaS and independent apps are being rapidly eroded, while scarce content and精细化服务 (refined services) are finding more market.

The same day, multiple bloggers shared new thoughts on Agent workflows. Max For AI used a barber as a metaphor, pointing out that Agents need to learn to ask the most valuable questions at the right moment and build long-term user models (Memory). Practices from the LobeHub team show that a $200/month Claude Code subscription配合 (combined with) Agent orchestration can produce value worth about $10,000/month, effectively tripling the team’s output. Meta was reported to be planning to open-source Harness—a model external system (task decomposition, tool calling, Context management). After Alexandr Wang confirmed this, multiple bloggers suggested models are becoming commoditized, and the competitive delivery unit in AI will shift from the model alone to model + Harness + tool ecosystem.

Sources:

miHoYo Founder’s AI Companion Product AnuNeko to Shut Down

Anuttacon, under miHoYo founder Cai Haoyu, announced that its AI chat product AnuNeko will permanently shut down on July 29, with the app and all chat records becoming inaccessible. The product, launched just 8 months ago, was positioned as a digital life with personality, emotions, and opinions, not catering to users. Max For AI commented that the product has completed its阶段性实验 (phase experiment), and the team will focus resources on the AI game “Whispers from the Star.” However, this also exposes the残酷的一面 (cruel side) of AI companion products—for users, an emotional connection has been established, but for the company, it remains just a server that can be shut down at any time.

Sources:

ChatGPT Health Launches in US Market; Midjourney V8.2 and Seedance 2.0 Continue Iterating

Blogger Guizang (op7418) shared hands-on experience with the ChatGPT Health feature launching in the US. After connecting Apple Watch health data, it can analyze the impact of factors like cardio fitness and sleep on health, retrieve information from authoritative health websites, and output weekly health reports.

Additionally, Midjourney V8.2 was praised by multiple bloggers for its unique style and excellent quality. Chengzilhy shared deep video Skills based on Seedance 2.0, which use action reference videos to排除干扰 (eliminate interference), allowing AI to more accurately understand camera movement and action rhythm. Dash Huang (Boss Huang) also open-sourced an AI Agent client named Cindy,纯 Vibe 编码产出 (purely vibe-coded output).

Sources:

Stats: Timeline Scans=360/returnedTweets Matching Bloggers=36/matchedHandles Matching Tweets Total=212/matchingTweets Weighted Tweet Score=179.75/weightedMatchingTweets Original Tweets=115/originalTweets RT Tweets=28/retweetTweets Fetch Attempts=2/fetchAttempts Boundary Coverage Status=tail_confidently_crossed_target_boundary/coverageReason