Yesterday, I chatted with a friend who works in investment. He asked me what signals in the AI industry are worth paying attention to recently. I thought for a moment and told him about two scenes—one scene is DeepSeek's Li Wenfeng raising 70 billion yuan in funding, quietly hiring more people to develop Code Harness, directly targeting Claude Code; the other scene is Anthropic's quarterly revenue soaring 80 times, achieving profitability two years ahead of schedule, and then turning around to raise another 30 billion US dollars. He said that putting these two scenes together, it feels like the entire AI model layer is being torn in half—one side focuses on being expensive and deep, while the other focuses on volume and affordability. I said that's right, and this trend has become especially evident this week. Today's content features a major focus on AI programming tools, and there are also new developments in embodied intelligence. Keep reading.
Core content：DeepSeek's latest round of RMB 70 billion funding has officially closed, with its valuation exceeding RMB 100 billion. After the funding completion, DeepSeek immediately released a large-scale recruitment announcement, focusing on the core direction of "Code Harness" — creating the next-generation AI programming assistance platform comparable to GitHub Copilot and Claude Code. The recruitment scale is expected to exceed 100 people, with core technical positions offering salaries comparable to those of first-tier tech giants. Meanwhile, The Information revealed that industrial capital firms such as Contemporary Amperex Technology Co. (CATL), JD.com, and NetEase are in talks to participate in the investment. Liang Wenfeng stated that DeepSeek's ultimate goal is to achieve AGI, while also committing to maintaining an unwavering open-source path.
comments : RMB 700 billion secured + Code Harness team launch signals DeepSeek's expansion from "model player" to "tool ecosystem" very clearly. Previously, people were still speculating whether DeepSeek would enter the application layer, but the answer is here — first focus on programming tool chains, as this is the closest entry point to developers.
Core Content：DeepSeek announces the official adjustment of the V4-Pro API price to the current promotional price starting June 1st, with a permanent reduction and no plans to revert to the original price. This is another bold move by DeepSeek in its API pricing strategy, aimed at further expanding its market share. Combined with the completion of financing and the recruitment of the Code Harness team, DeepSeek's dual-track progress in the model layer and tool layer has明显 accelerated.
Comment：Price cut + financing + hiring, three things happening in the same week, DeepSeek's strategy is very clear: use low prices to increase the number of API calls, use open-source to lock in developers, and use Code Harness to deepen usage scenarios. The most direct pressure on this path comes from OpenAI.
Core Content：Anthropic released its Q2 2026 financial report: quarterly revenue reached $10.9 billion, achieving operating profit for the first time at $559 million, annual recurring revenue (ARR) exceeded $44 billion, surging 80 times year-on-year. The company originally planned to achieve profitability by 2028, but actually completed the goal two years ahead of schedule. The large-scale deployment of Claude Code Enterprise is the main driver of growth. Meanwhile, Anthropic is negotiating a new round of $30 billion funding at an estimated valuation of $900 billion, which would bring its valuation close to $1 trillion.
Comment：Anthropic's profitability came very quickly, with the core driver being the deep integration of Claude Code in the enterprise sector. When companies start using Claude Code to manage their core code repositories, the replacement cost becomes very high, which is exactly what Anthropic wants—selling not just APIs, but the workflow of "using AI to write code."
Main content：OpenAI has confidentially submitted its IPO prospectus to the U.S. SEC, with Goldman Sachs and Morgan Stanley as advisors, aiming for a valuation of $100 billion, with an expected official listing as early as September 2026. Currently, OpenAI's annualized revenue is approximately $25 billion, with 900 million weekly active users, but it is still operating at a loss. Analysts believe that Anthropic's profitability and DeepSeek's low-price strategy are significant external pressures pushing OpenAI to accelerate its listing.
Comment：OpenAI is being forced to go public, and this is not an exaggeration. Anthropic has already become profitable by adopting a high-end corporate approach, while DeepSeek is aggressively capturing user volume with a consumer-focused low-price strategy. OpenAI, caught in the middle, finds it difficult to maintain its current level of computational power investment without going public for funding.
Key content：Canadian AI company Cohere releases Command A+, a 218 billion parameter MoE architecture large model, fully open-sourced under the Apache 2.0 license, capable of running on a single NVIDIA B200, supporting 48 languages and native citation capabilities. This marks another landmark event in the field of open-source large models following Mistral, with the Apache 2.0 license meaning it can be used freely from research to commercial, making it highly attractive for enterprise users.
comments ：Cohere has fully open-sourced the 218 billion parameter model this time and chose the most friendly Apache 2.0 license, clearly targeting the enterprise market. For tech companies that don't want to rely on OpenAI/Anthropic APIs, Command A+ is a very attractive option.
core content：Anthropic has launched the officially managed Claude Code plugin directory (claude-plugins-official) on GitHub, a curated library of high-quality plugins. This marks a significant step by Anthropic in building an AI programming tool ecosystem, similar to the plugin market logic of VS Code but tailored for AI programming scenarios.
Review ：Claude Code creating a plugin directory is a crucial move. Once the developer ecosystem takes off, Claude Code won't just be a programming assistant but an "AI programming operating system"—a layout that OpenAI Codex currently lacks.
Key content：The official Chrome DevTools team has launched the chrome-devtools-mcp project, based on the Model Context Protocol (MCP), which allows AI programming agents to access browser debugging features through a standardized protocol. This marks the evolution of AI-assisted development from code writing to browser environment interaction and automated debugging, enabling AI Agents to directly manipulate browser developer tools.
Comment：The official Chrome DevTools team getting involved in MCP means that the capabilities of AI Agents to operate on browsers will be significantly enhanced. Previously, for web automation, one had to write scripts to interact with ChromeDriver, but now it will be possible to directly communicate with DevTools.
Core Content：Developer colbymchenry released the open-source project CodeGraph, a pre-indexed code knowledge graph designed for AI programming tools like Claude Code, Cursor, Codex, etc. Running 100% locally, it aims to reduce token consumption and tool call frequency, enhancing code privacy protection. The core idea is to pre-establish a knowledge graph index for the codebase, allowing AI tools to query the graph directly instead of performing full-scale code searches each time.
comments ：This tool addresses a very practical pain point—AI assistants in large codebases have to read many files each time, consuming a huge number of tokens. CodeGraph's approach uses a knowledge graph for "pre-search," similar to pre-creating an index for the codebase, so AI tools can directly query the index without flipping through the entire book.
Key content：Embodied intelligence startup AtomBite.AI (元节智能) has completed a million-level seed funding round, led by InnovCo Fund and co-led by Shuimu Tsinghua Alumni Seed Fund. Dr. Wang Dong, the founder, is the former technical lead of the Meituan Delivery Department. The company focuses on the restaurant kitchen scene rather than humanoid robots, developing "Embodied Intelligence Restaurant Robots," and has secured product cooperation and deployment intentions from several top domestic and international companies.
comments ：The embodied intelligence track is starting to move towards vertical scenarios, not everyone making humanoid robots. The restaurant kitchen is an interesting starting point——the environment is relatively controllable, the tasks are clear (cooking, delivering dishes, preparing ingredients), and the shortage of labor is a real issue, not a pseudo-demand.
Core content：A domestic GPU manufacturer has launched the country's first full-stack domestically developed embodied intelligence simulation platform. During the live demonstration, a robot dog named "Xiao Fei" walked onto the stage, and Xiao Fei in the simulated world simultaneously completed a side flip on the screen, while Xiao Fei in the physical world immediately replicated the exact same action, achieving real-time synchronization between the physical and digital worlds. The platform covers the entire chain from simulation training to physical deployment.
comments : The "Sim2Real" migration from simulation to reality has always been a core challenge in embodied intelligence. This platform's ability to achieve real-time synchronization between the physical and simulated worlds indicates excellent latency control, which will significantly improve training efficiency.
Core content: AI matriarch Fei-Fei Li announces new research on spatial intelligence, aiming to establish standards for 3D spatial understanding and embodied intelligence. Core argument: Language models solve "thinking," visual models solve "seeing," but AI entering the physical world needs to address "spatial reasoning" — understanding three-dimensional relationships, predicting physical changes, and planning multi-step actions. She believes spatial intelligence is on the verge of an explosion similar to ImageNet.
comments on : Li Fei-Fei calls this the ImageNet moment for spatial intelligence, and this analogy is worth taking seriously. ImageNet spurred the explosion of visual AI; if spatial intelligence also has a similar benchmark, the pace of embodied intelligence research is likely to accelerate significantly.
core content：Microsoft's .NET official team has launched the "skills" open-source project, specifically designed to enhance the capabilities of AI programming agents in the .NET and C# environments. By using predefined "skills" to address the accuracy issues of AI in complex development environments, it marks an important step for the .NET ecosystem towards AI-native development.
comments ：Microsoft's layout in the AI programming toolchain is systematic — GitHub Copilot for code completion, Semantic Kernel for Agent frameworks, and now Agent Skills for .NET. For .NET developers, the appeal of this combination is significant.
Core content：Kunlun Wanwei CEO Fang Han stated at the 2026 China AIGC Industry Summit that experience is no longer a moat in the AI era; companies should aim to be the second best in their industry to maintain flexibility and combat effectiveness. He pointed out that "how many tokens are burned" has become a new measure of strength in the AI era—ordinary people can burn several million tokens in a month, while some heavy users (like those within OpenAI) can reach the scale of hundreds of billions of tokens. He also predicted that by 2027, 50% of programming work will be completed by AI.
comments : "How many tokens are burned" as a measure of strength is quite interesting. Fang Han likely means that in the AI era, it's not about who has more experience, but who can leverage AI more effectively. To some extent, token consumption can indeed reflect the depth of a team's AI integration.
Core content：Jason Liu（author of the open-source library Instructor, 13k stars）has recently joined the OpenAI Codex team and publicly shared detailed practical tips on how to fully utilize the Codex API, including advanced uses such as letting Codex automatically follow up on Amazon refunds, scheduling Slack requirement scans, and continuously running pending tasks. The article demonstrates the potential of Codex as a general automation engine, far beyond the single scenario of "writing code."
comments on ：Someone from the Codex team is teaching everyone how to "milk" Codex, and this signal is worth noting. OpenAI clearly wants developers to understand Codex as a general Agent platform, not just a programming assistant. This is the key difference compared to Claude Code.
core content：Developer Lum1104 launched Understand-Anything, transforming code repositories into interactive knowledge graphs, emphasizing "teaching-oriented graphs." It deeply integrates mainstream AI development tools such as Claude Code, Cursor, Copilot, and Gemini CLI, providing developers with a new way to understand and learn code. Unlike CodeGraph, this project focuses more on the "teaching" scenario—helping new members quickly grasp large codebases.
comments ：Understand-Anything and CodeGraph share similar ideas but differ in context: one leans toward "helping AI save tokens," while the other focuses on "helping people understand code." Both projects topped GitHub Trending on the same day, indicating that the direction of "code knowledge graphs" is indeed very hot this week.
Core content：Meituan Drones announced that its “Low-Altitude Air Network” has officially entered regular operation, while also launching a recruitment program for authorized service providers nationwide for low-altitude logistics operators. Leveraging its newly released self-developed hardware and software products, Meituan is opening up its low-altitude delivery capabilities to the industry. This means Meituan’s drone delivery has moved from the “self-operated pilot” stage to the “platform-based open” stage.
Comment ：From doing deliveries itself to opening it up to the entire industry, Meituan’s drone business is following a path similar to its food delivery business back then—first running it smoothly itself, then opening it up to third parties. The low-altitude logistics proposition is indeed accelerating from a demonstration phase to regular operation by 2026.
Core Content ：Jinri Toutiao announced staff reductions in its Toutiao Baike business team, reallocating resources to focus on core information distribution services. Internal sources said that the encyclopedia business has limited relevance to core competencies, and the company has decided to reduce investment in this direction, shifting resources toward core areas such as AI recommendation algorithms and content safety. Employees affected will receive an N+2 compensation package.
Comment: Toutiao Baike has not taken off after years, showing a clear gap compared to Wikipedia. Now they are cutting it back, focusing resources on AI recommendations and content safety. This decision is actually too late—it should have been done two years ago. However, for internet companies, doing cuts is always better than dragging things on.
Key Content: SpaceX has included a retail allocation round in its IPO design for the first time, giving ordinary investors a chance to participate in one of the highest-valued non-listed tech companies globally. SpaceX's valuation is expected to exceed $250 billion, and its IPO size may set a record for tech stocks. Analysts believe the inclusion of retail allocation is aimed at expanding the investor base and boosting market enthusiasm.
Comment：SpaceX IPO with retail allocation, this design is clearly intended to stir up market hype. With a $250 billion valuation in place, whether retail investors can actually make money is uncertain, but Musk needs the FOMO sentiment of retail investors to push up the offering price.
Key content：Microsoft is beginning to cancel Claude Code enterprise licenses for some companies, which is interpreted as a measure to prevent enterprise users from switching to competitors and protect the market share of GitHub Copilot. In the short term, it can help retain Copilot's market share, but in the long term, it may damage Microsoft's "open ecosystem" brand image. This move has triggered significant backlash in the developer community.
Comment：Microsoft's move is a bit short-sighted. Developers choose Claude Code because it's genuinely useful, and revoking the license will only make people more反感微软的做法. The competition between Copilot and Claude Code will ultimately depend on the product, not on bundling.
Key content : Constellation responded to investors' inquiries about expanding cooperation with SpaceX on an interactive platform, stating that the impact of a single user on the company's overall performance is limited, and the company's business does not rely on a single customer. Previously, there were reports that Constellation intended to expand its supply chain cooperation with SpaceX, and the market responded positively, with the company's stock price rising by about 12% this week.
Comment : Constellation's response is very standard—not outright denying anything but also not over-promising, avoiding abnormal fluctuations in the stock price. It's not easy for Chinese companies in the SpaceX supply chain to secure orders, and if Constellation can stabilize this customer, there is still room for growth.
This content is automatically aggregated by InertiaRSS (RSS Reader) for reading reference only. Original from — Copyright belongs to the original author.