Summary of Key Points
In the first half of 2026, the AI industry continued to gain momentum, albeit with fluctuations: the investment community was plagued by "FOMO" (Fear Of Missing Out), leading to soaring and rapidly changing project valuations. Technologically, Agents such as OpenClaw, world models, and embodied intelligence became hot topics, with models transitioning from research environments to practical applications. The gap between Chinese and American AI models has narrowed, but infrastructure (computing power, storage, and transmission) remains a weakness in China. The capital market saw a wave of AI companies going public, raising concerns about potential bubbles. Investors are focusing on those companies that can survive the bubble burst—either by having ongoing research achievements or a stable business model.
I. The "Craziness" in the Investment Community: Valuation Chaos driven by FOMO
The number of AI projects surged in the first half of the year, and investors were generally anxious about missing out on opportunities. For instance, some project valuations changed within weeks, with multiple rounds of financing occurring simultaneously (one round for payment, one for signing agreements, and another for finalizing plans). This phenomenon reflects a shift towards "prevalent valuation" – whereas early-stage projects used to require significant progress before their valuations increased, current trends allow valuations of $1 billion to emerge in just a few months. However, investors' main concern is not about missing out on projects, but rather the rapid pace of change in AI: new models and buzzwords emerge daily, and missing out on news for a few days could mean falling behind. Some warn that market sentiment should not dictate investment decisions; it's important to distinguish between value investing and opportunistic investing.
II. Technological Hotspots: What Really Matters After the Hype?
1. Agents (Intelligent Agents): The OpenClaw craze has given way to the development of "agent operating systems" that allow ordinary users to create workflows (e.g., beauty bloggers using them to gather tech information and generate content). The focus now shifts to "agent runtime engineering," enabling models to make autonomous decisions and execute tasks without explicit human instructions.
2. World Models: These models have suddenly gained attention this year due to the advancement of video generation and 3D technology. They represent the ability to predict global changes (e.g., reducing bugs in video games). The current challenge is the lack of data that incorporates physical laws, but real sensors, simulations, and videos can complement these datasets.
3. Embodied Intelligence: This field is still working on overcoming the "impossible triangle" of long tasks, high success rates, and cross-scenario functionality. Tactile adaptation remains a challenge, but there are low-cost examples (e.g., two-finger grippers with sensors that can handle fragile objects like tofu).
4. AI for Science: AI is being applied to streamline scientific research, from formulating questions to conducting experiments, even extending to social sciences (using AI for predictive analysis).
III. The Gap Between Chinese and American AI Models: Nearing Closure, but Infrastructure Remains a Weakness
- Changing Gap: What seemed insurmountable in 2022 is now within just a few months of catch-up. China has made significant progress in areas like fast versions of models and video generation (e.g., with platforms like Seedance), thanks to more efficient architecture design.
- Infrastructure Weaknesses: The U.S. has larger computing clusters that can run multiple experiments simultaneously, while China is catching up through advanced algorithms and engineering efficiency. However, infrastructure (computing power, storage, and transmission) still lags behind.
- Chinese Advances: Companies like Zhipu are focusing on coding capabilities, DeepSeek has raised funds and collaborated with domestic computing power providers, and MiniMax has received support from OpenClaw and been included in NVIDIA's GTC agenda.
IV. AI Infrastructure: More Than Just Computing Power
AI infrastructure encompasses more than just chips; it also requires coordinating computing, storage, and transmission capabilities. For example, if data transfer is slow among ten thousand computing nodes, it's like workers waiting for materials to arrive. Therefore, technologies like high-bandwidth memory (HBM) and liquid cooling become increasingly important. The core principle is to use the least amount of energy and money to generate the maximum number of tokens (the units used by models). Additionally, there's a division of labor between cloud, edge, and client devices: simple tasks are handled on mobile phones, private data on enterprise equipment, and complex tasks on the cloud. This approach presents opportunities for startups with limited resources.
V. The Capital Market: A Wave of Public Offerings and the Bubble Question
- Public Offering Trend: Companies like Yushu have registered on the STAR Market, while Zhipu and MiniMax have seen varying stock prices after going public. Kimi and DeepSeek have secured substantial funding. Going public is about raising more money for continued research and development; valuations are no longer based solely on PE ratios but on the company's position within the ecosystem.
- Bubble Concerns: It's difficult to predict whether bubbles will burst, but if they do, funds will likely flow towards "real assets." Research-driven companies need to continuously deliver milestones (e.g., progress with world models), while practical applications require stable business models and genuine customer orders. The bursting of a bubble could also be beneficial, as it might lower computing costs and facilitate more projects.
- Second Half Expectations: The availability of real data will be crucial; companies that can create feedback loops for their technologies (including those selling related tools) will have a better chance of success.
In summary, the AI industry is evolving rapidly, and bubbles may exist, but technology will continue to advance. Only those companies that can implement their ideas and achieve tangible results will survive the fluctuations.