Midway through 2026, Fengrui Capital investors review the AI sector in the first half of the year. Project valuations fluctuate multiple times per week, leaving investors grappling with FOMO anxiety amid rising hype and growing divergence. The Agent space has transitioned from OpenClaw to Harness Engineering, with models shifting toward task execution and data flywheels becoming startup moats. World models have seen a concentrated surge, with domestic Chinese voices gaining louder traction in video generation. The gap between U.S. and Chinese models has narrowed to just a few months, and compute infrastructure is evolving into a “power in, tokens out” system competition. Regarding bubble bursts, investors believe research-driven companies must consistently deliver key milestones, while application-focused firms must validate the stability of their business models. The key variable for the second half of the year lies in real data闭环.

Author and source: Fengrui Capital

If the AI industry begins to "pop the bubble," which types of companies will be revalued first?

In recent times, the AI sector in the secondary market has experienced significant volatility, with chips, storage, and optical interconnects alternating through fluctuations. Meta has once again raised its capital expenditure forecast for 2026, bringing the debate back to the forefront: how long and how rapidly will AI infrastructure investment continue?

Money continues to flow into computing power, storage, and data centers, and the market is beginning to ask: Has AI infrastructure reached a阶段性 high point? How much longer can this wave last? Meanwhile, embodied intelligence, world models, agents, tokens, and hardware shifts from cloud to edge to endpoint are also unfolding simultaneously.

Halfway through 2026, Fengrui Capital’s technology investors Liu Pengqi and Yan Qianhang sat down to review the past six months: Why do project valuations change weekly? What remains after the hype around OpenClaw has faded? Why have world models seen a concentrated surge? What does “electricity in, token out” reveal about the competition for computing power? And if bubbles are bound to burst, which companies will stay on the table?

We’ve compiled a selection of this conversation—listen to the full episode by searching for “High Energy” on Xiaoyuzhou or Apple Podcasts.

Valuations change once a week—investors fear more than just missing out.

Liu Pengqi: Looking back from mid-2026, the evolution of AI has shown no signs of slowing down—from OpenClaw and world models, to storage and optical interconnects, all the way to Zhipu’s market capitalization surpassing one trillion Hong Kong dollars. One hotspot followed another within just this half-year period. Today, let’s walk through them one by one to understand what truly happened behind the scenes.

First, let’s ask Yan Bo: How many projects did you roughly discuss in the first half of the year?

Yan Qianhang: Around 200 to 300, with AI-related projects accounting for nearly 150 to 200.

Liu Pengqi: My experience is pretty similar. If you had to use a few keywords to describe your feelings about evaluating projects in the first half of the year, what would you choose?

Yan Qianhang: The first is FOMO (fear of missing out). Hot sectors suddenly attract a surge of new people and projects, with valuations rising rapidly. Many feel compelled to review every single project: Have we invested today? How much have we invested? Should we invest even more?

The second is structural differentiation. A few popular sectors are experiencing active fundraising, while most sectors remain subdued. The same applies to the secondary market: technology stocks are performing well, but stocks outside the spotlight may be flat or even declining.

The third is front-loaded valuation. In the past, early-stage investing followed a relatively clear framework: specific milestones corresponded to defined valuations, and valuation would increase only after the company reached consensus and delivered on阶段性 deliverables. Now, hype and FOMO have disrupted many of these established norms—valuations of $1 billion can surge within just a few months.

Liu Pengqi: Some projects see their valuations change within a week—so much so that three funding rounds can be happening simultaneously: one closing, one negotiating terms, and another finalizing its structure. In such a market, do investors feel anxious?

Yan Qianhang: It’s unrealistic to say I’m not anxious. But what worries me more than missing out on a particular project is how fast AI is evolving. Every day brings new models, trending terms, and fresh ideas. If you see a piece of news just a few days late, you can’t help but wonder: Have I already been left behind?

Liu Pengqi: Investments can be value-oriented or opportunity-driven—they need to be analyzed separately; don’t let market sentiment make decisions for you. Set aside valuation and FOMO for a moment: What new developments actually emerged in the first half of the year?

Yan Qianhang: Let me start with research-driven entrepreneurship. Recently in the U.S., there has been a lot of discussion around a concept called NeoLab—using commercial company models to solve long-term research problems. Many such companies emerged in the first half of this year. People are no longer just asking, “Is this a professor starting a company?” but instead focusing on whether the research can progress faster after the professor launches a company than it could under the original organizational structure.

Now, looking at AI applications and AI hardware, people are beginning to refocus on the products themselves. After Manus emerged, the market experienced FOMO around AI agents; later, Insta360 reignited interest in AI hardware. After chasing these trends, everyone still needs to return to those who truly define products.

AI coding also enables entrepreneurs to move beyond just algorithm researchers and engineers—lawyers, designers, product managers, and even stay-at-home parents can now use coding tools to build products.

Liu Pengqi: Previously, it was often said that tech entrepreneurs were “holding a hammer and looking for nails.” Now, it’s more like the nails themselves are starting businesses, and the hammer has become a universal tool.

Bust the bubble of the Agent—beneath it lies rich, flavorful beer.

Liu Pengqi: OpenClaw was likely one of the most talked-about events of the first half of the year. From everyone “raising lobsters” and scrambling for Mac Minis, to model providers and IM platforms launching their own agents, the hype has now gradually subsided. Open-source brought it into the mainstream, but it also exposed issues around security and user experience. After all the excitement, what has it truly left behind?

Yan Qianhang: OpenClaw enables many ordinary people to deploy agents, download skills, and gradually refine them into their desired form through natural conversation. It functions more like an Agent OS (Agent Operating System) rather than just a tool for completing a single task.

For example, I have a friend who runs a beauty-focused social media account and wants to venture into tech content. She built several workflows using OpenClaw to systematically gather tech-related information and established a content production pipeline. Many people are already reliably using it for coding, PPT creation, e-commerce, and internal corporate processes. After the initial surge of hype—with its bubbles, fluctuations, and noise—when you blow away the foam from a beer, what remains is rich, flavorful brew.

Liu Pengqi: Recently, people have begun emphasizing Harness Engineering (agent execution environment engineering). When large models shift from chat tools to task executors, what becomes more important?

Yan Qianhang: An agent must make decisions, execute tasks, and then adjust based on outcomes—this entire process needs to keep running. The model is merely the “brain” within it. Harness Engineering focuses on setting up the surrounding environment and feedback loops, allowing the model to operate autonomously within the workflow rather than having every step pre-defined by humans.

Liu Pengqi: This is also related to the data flywheel. In traditional internet platforms, personalized experiences come from accumulated user data—could agents become the vehicle for creating data flywheels and personalized experiences?

Yan Qianhang: Chatbots primarily retain linguistic context, while agents retain task trajectories—how they made decisions, how they executed actions, which tools they called, and whether they ultimately completed the task. These trajectories can be used for agent reinforcement learning (RL). Data from vertical scenarios may not return to foundational model companies, which could become a competitive barrier for agent-focused startups.

Liu Pengqi: From this perspective, whoever can access users’ task trajectories holds more initiative. Currently, there are roughly three types of players in the field: model providers, IM and office platforms, and local hardware. Who do you favor?

Yan Qianhang: They serve different purposes. Model providers are suited for general tasks like search, deep research, and coding; office platforms come with existing files, documents, and business workflows; local hardware is more related to privacy. However, whether the “lobster machine” represents a viable product path remains questionable. Initially, the Mac mini was the most aggressively pursued, and even after the hype faded, it remained the most commonly sold device on Xianyu.

There’s another issue with local hardware: who owns the data? For example, if a company’s lead architect uses an agent and wants to leverage the best available model, but the model provider uses the architectural designs and task trajectories for training, that poses significant risks. Many users face this conflict: they want to use the best models, but aren’t willing to hand over their data. How do you view the trade-off between privacy and model capability?

Liu Pengqi: Based on past experience, most C-end users prioritize efficiency and experience. However, Pro C and B-end users are different—technical documentation, product design, and business data are all core assets for enterprises. The opportunity may lie here: enabling enterprises to use the best models while maintaining control over local data, permissions, and security.

Yan Qianhang: Pengqi and I both looked into SaaS and Fintech early on, and we’ve always had a love-hate relationship with large B2B clients—they have substantial budgets but complex needs, and servicing them is heavily resource-intensive. Now we’ve added FDEs (Frontline Deployment Engineers), requiring people who understand both models and business to go on-site to help customers with deployment. How will AI serving large B2B clients differ from past SaaS and Fintech approaches?

Liu Pengqi: Large enterprise clients have big budgets and high demands—old issues like increasing volume without raising prices and high customization costs still persist. Clients also ask: If there are open-source models, why pay extra for AI services? When AI truly integrates into enterprise infrastructure, issues like data security, privacy, and system integration must be addressed. There’s also the problem of hallucinations—if AI is used in core functions like risk control and makes an error causing losses, who is responsible: the model provider, the enterprise, or the employee? This boundary has not yet been clearly defined. Whoever can solve these problems will have the opportunity.

Power in, tokens out; AI infrastructure enters the system competition.

Liu Pengqi: Over the past six months, has the pace of advancement in large model capabilities accelerated or slowed down?

Yan Qianhang: Truly groundbreaking innovations like the o1 reasoning architecture haven’t been seen in a while. However, models continue to improve in their ability to perform real-world tasks. Previously, we evaluated models using math tests and knowledge benchmarks—like having them take an exam. Now, we assess them using SWE-bench (software engineering benchmarks), Computer Use, and Agent Benchmark, which test whether they can fix bugs, use tools, and complete complex, multi-step tasks. Models have moved from the exam hall to the workplace.

Liu Pengqi: It’s like a student graduating from school and entering society. Their knowledge base may already be largely established, but the real test is how much ability they can demonstrate in their work. Among the companies in this round, who would you choose as the most outstanding?

Yan Qianhang: I would choose Anthropic. It placed a strong focus on coding very early on—a decision that now appears crucial. It recognized earlier than others that models shouldn’t be judged solely by exam scores but by their ability to execute tasks. Since the release of Claude 3.5, the user experience of AI coding tools has improved significantly, and products like Cursor have become increasingly intuitive. As these tools improve, more users adopt them, and the feedback loops further enhance the models, creating a self-reinforcing cycle.

Another reason is its rapid revenue growth. Anthropic has repeatedly disclosed annually recurring revenue figures this year that have been rising quickly, prompting growing discussions about what kind of company Anthropic would be if it went public. Improvements in inference costs and profitability have also led many to reassess its potential.

Liu Pengqi: It seems they’ve chosen a market that appears vertical but is actually large enough, focusing first on deeply penetrating this user base. It’s better than spreading too thin without a clear, leading direction.

Yan Qianhang: I don’t really consider coding a vertical domain. Coding inherently involves human logic, reasoning, and decision-making—turning an algorithm into a functional application requires all of these. When models delve deeply into coding, they naturally develop stronger reasoning, associative thinking, and chain-of-decision skills. That’s why users find Claude to be meticulous and intelligent.

Some models have great ideas but lack reliability. Anthropic entered through coding and agents, and real-world tasks are simultaneously training the model. It’s somewhat like “the unity of knowledge and action”: it integrated “knowledge” and “action” into the same feedback loop earlier than others.

This year, Anthropic launched Fable 5 and Mythos 5. Fable 5 is available to general users, while access to Mythos 5 is more restricted. Both models were temporarily suspended at one point but have since been restored. What are your thoughts on the discussions surrounding these restrictions?

Liu Pengqi: There are already many evaluations of these models online, but relatively few people have truly used them in depth. After a quick look, I found that their performance in context length, adaptive iterative reasoning, and integration with agent capabilities is indeed impressive. As for the so-called “bans,” I suspect there may be some smoke and mirrors and scarcity marketing involved. Whether they have truly become powerful enough to serve as tools in national competition remains to be seen.

Yan Qianhang: Looking back at domestic models, what do you think has been the biggest change over the past six months?

Liu Pengqi: Zhipu is a very typical example. It is also pushing forward in the direction of coding, making its own attempts in architecture and long-horizon task training. We’ve also heard that the team has incorporated methods such as process rewards to enhance model capabilities. Domestic teams continue to push forward, with ongoing advancements in architecture, algorithms, and model capabilities.

We’re also catching up on data. If you only rely on distillation, you’ll hit a ceiling quickly. Now that model companies have secured more funding, in addition to buying more computing power and hiring talent, they’re also willing to invest in high-quality data.

DeepSeek is also worth watching. After releasing its models, it remained quiet for a while, but has been steadily building its capabilities. A notable change this year is its more open attitude toward fundraising and partnerships with domestic computing power providers.

I understand that it requires valuation and equity incentives to retain talent, and also aims to build deeper partnerships with domestic computing power providers. Doing so not only enhances its own models but also helps drive the entire industry ecosystem.

Yan Qianhang: When DeepSeek first emerged, most domestic model companies had funding scales of just a few billion yuan. At that level, a research lab backed by a quantitative fund could still compete with others. But from last year to this year, major tech companies have begun pouring massive investments—Alibaba and Xiaomi are both spending heavily, and after their IPOs, companies like MiniMax and Zhipu no longer operate with just a few billion yuan in funds. If DeepSeek continues to operate in the same way, it would be deliberately making things harder for itself.

But DeepSeek’s approach has remained largely unchanged—all choices still revolve around one goal: building better large models, rather than prioritizing short-term commercialization.

Liu Pengqi: After discussing several domestic model companies, let’s look at the gap between China and the U.S. Do you think this gap is growing or shrinking?

Yan Qianhang: The gap between China and the U.S. is indeed narrowing on both the Ultra and General Standard models. In 2022, many felt it would be hard to catch up; in 2023 and 2024, people still said it was one or two years away, but this year, the perceived difference may be just a few months. Recently, Tang Jie, co-founder of Zhipu, had a conversation with Musk on X. Musk said China’s large models might not catch up to the cutting edge until Q1 2027; Tang replied, “It won’t take that long.”

The gap that remains difficult to overcome in the short term is infrastructure and computing power. The U.S. has larger computing clusters that can run more experiments simultaneously; China excels at achieving more with less, quickly catching up through algorithmic and engineering efficiency.

Liu Pengqi: Yes. With more machines, we can run more experiments simultaneously, allowing for greater exploration of model architectures. Domestically, our ability to augment data has improved significantly compared to before, and data's contribution to model advancement is also increasing. Looking at the Ultra and Lightweight models, we indeed have a clear advantage in efficiency and cost.

Yan Qianhang: Chinese models are also gaining more visibility. MiniMax’s models are supported by OpenClaw; both Kimi and MiniMax were mentioned in NVIDIA’s GTC agenda and related talks in March. We’ve been moving quickly to catch up across the Ultra, Standard, and General versions.

Video generation has become more interesting, with louder voices emerging domestically. After Seedance’s release, the market briefly circulated claims that its ARR had reached $2 billion, though Volcano Engine later clarified that publicly circulated revenue figures were generally overstated. Abroad, apart from Google Veo 3.0 still maintaining attention, Sora’s market buzz has significantly diminished since its launch. In terms of video models, domestic players appear to be stronger.

Liu Pengqi: Because video generation consumes a large number of tokens, the cost and efficiency per token are significantly amplified. Domestic manufacturers can reduce costs to an acceptable level for users through architectural design and more efficient utilization of computing power.

Below this, you can't avoid AI infrastructure. Model providers are continuously increasing capital expenditures, and token consumption is rising—computing power costs will inevitably become a bottleneck. What changes have you observed in the infrastructure?

Yan Qianhang: What people are asking now isn’t just whether the computing power is sufficient—it’s whether the entire system can run smoothly. Beyond “computation,” we also need to consider “storage” and “transmission.” Imagine ten thousand computing nodes, but if results can’t be delivered to the next node in time, all the others have to wait. It’s like hiring ten thousand workers on an assembly line, but materials arrive too slowly—so many are just standing around idle.

Therefore, optical interconnects, HBM (High Bandwidth Memory), DRAM (Dynamic Random-Access Memory), and SSDs (Solid-State Drives) are becoming increasingly important. Abstract the entire system, and it really comes down to one thing: “Power in, Tokens out.” To maximize the number of Tokens gained per unit of power and operational cost, chips, memory, communication, liquid cooling, and software must all be optimized together.

Liu Pengqi: When evaluating infrastructure, don’t focus solely on the cloud. Edge devices include everyday gadgets like smartphones, smartwatches, and headphones; the edge layer can also encompass NAS (Network-Attached Storage), all-in-one systems, or enterprise internal devices; while the cloud provides large-scale computing power. Simple tasks are handled at the edge, personal and enterprise private data remain on the edge layer, and complex programming and long-range inference are delegated to the cloud.

Yan Qianhang: Cloud-side has temporarily absorbed a large amount of resources, forcing edge-side development to deliver functionality with less computing power and lower costs. For startups, these constraints also harbor opportunities.

World model, not another large language model

Liu Pengqi: In the first half of 2026, the term "world model" suddenly became a hot topic. Embodied intelligence, video generation, gaming, and physics simulation are all discussing it. My understanding is that a world model doesn't necessarily refer to a specific type of model; rather, it's more like: applying an external intervention to the world and predicting what will happen in the next moment.

By this definition, could Newton's laws also be understood as a world model? Yan Bo, how do you understand world models? What is their relationship with VLA (Vision-Language-Action) models, video generation models, and large language models?

Yan Qianhang: World models are not a new concept—they have long existed in reinforcement learning. Today, we can see several approaches: video models generating the next frame, 3D world models constructing interactive environments, JEPA (Joint Embedding Predictive Architecture) abstracting the world into a latent space representation, and WAM (World Action Model) integrating the world model directly with the action policy.

Liu Pengqi: Language is highly abstract information, but the real world also includes vision, touch, and spatial relationships. I believe world models are more like innate human abilities, while language models are more like learned skills. Their modes of iteration differ, yet together they shape human intelligence. Do you agree with this assessment? What relationship will world models form with LLMs (large language models) in the future?

Yan Qianhang: I also tend to favor a complementary approach. Language, along with the five senses—including visual, auditory, tactile, and olfactory perception—are all ways humans interact with the world. Large language models have first addressed intelligence at the linguistic level; naturally, we will continue exploring AI’s understanding and prediction of the physical world.

Liu Pengqi: Why did world models explode specifically in the first half of this year?

Yan Qianhang: First, video models have matured, producing results that increasingly resemble reality. 3D representation technologies have also advanced—from meshes and point clouds to NeRF (Neural Radiance Fields) and 3D Gaussian Splatting—giving startup teams a solid technical foundation for generating and manipulating 3D worlds. Coupled with open-source video models, many ideas can now be rapidly validated.

Liu Pengqi: But both world models and embodied intelligence lack data containing physical information. The industry often refers to the data pyramid, with internet videos and synthetic data at the base, first-person data with multimodal information in the middle, and real-world teleoperation data at the top. How can this bottleneck be addressed?

Yan Qianhang: We don’t lack ordinary video data; what we lack is data that captures physical states and underlying principles. The first type is data collected from real sensors; the second is data from physical simulations and digital twins; the third is inferred contact and force states derived from video analysis. Real data is ideal but most expensive—ultimately, we need to integrate all three types into a graded system.

Let’s keep it light. If world models start showing clear commercial progress, where do you think consumers will first notice them—in gaming, content consumption, or the embodied intelligence we’ve been talking about lately?

Liu Pengqi: I guess it’s gaming and content consumption. These scenarios don’t require absolute accuracy—just enough to fool the human eye. Embodied intelligence, however, must operate in the real world, demanding higher standards for data, success rates, and safety.

Yan Qianhang: If world models could really be used in games, I’m looking forward to no longer seeing all sorts of strange clipping and bugs.

Liu Pengqi: As we shift from content consumption to embodied intelligence, what changes has the world model brought to this industry? Some now believe that VLAs will be replaced by world models—is this really the case?

Yan Qianhang: I don’t think so. When a person plays tennis, they might first imagine how to return a shot when facing an unfamiliar movement; but when faced with a fast-moving ball, professional athletes rely on instinct to respond immediately. World models are more like “think first, then act,” while VLA is more like directly outputting actions based on visual input—these two can complement each other.

Liu Pengqi: Another "impossible triangle" in the deployment of embodied intelligence is achieving long-term complex tasks, high success rates, and cross-scenario generalization—all of which are currently difficult to satisfy simultaneously. Tactile feedback is also crucial. If a robot can reconstruct the approximate 3D shape of an object using only tactile sensors on its hands, it indicates that its capability has reached a relatively high level.

Yan Qianhang: We’ve seen a very intuitive example—a two-finger gripper costing only a few hundred yuan, when combined with a tactile sensor to form a force-control loop, can pick up tofu without crushing it. However, how to align tactile feedback with models is still in its early stages.

After the bubble bursts, where will the money flow?

Liu Pengqi: Besides the hottest areas such as agents, world models, and embodied intelligence, what other directions do you think entrepreneurs should pay special attention to?

Yan Qianhang: One of the areas I’m most interested in right now is the broader concept of AI for Science—using artificial intelligence to drive scientific research. By breaking down the research process—formulating questions, conducting literature reviews, performing experiments, analyzing data, and validating results—I’m exploring which steps can be delegated to AI and whether this can accelerate the pace of discovery. This goes a step beyond AlphaFold, which was widely discussed a few years ago.

Liu Pengqi: It’s somewhat like turning the research process itself into a pipeline that can be reorganized.

Yan Qianhang: Yes. This year, we’ve already seen a significant amount of work in Auto Research, though most of it remains limited to certain domains. What’s more worth watching is whether it can expand into more disciplines and dramatically accelerate scientific innovation. Another example relates to the world models we just discussed. Traditional physical simulations often rely on solving partial differential equations (PDEs). In the past, a major challenge was that achieving high accuracy meant slow computation, while speeding up the process often led to inaccuracies. Now, researchers are beginning to use AI and neural operators to replace parts of PDE solvers. If this can truly be done both quickly and accurately, it would represent a very tangible step forward.

Recently, I’ve found social sciences particularly fascinating. Previously, social science research lacked efficient experimental and deductive mechanisms: either treating real societies as experimental environments for large-scale surveys, or conducting small-sample sampling. Now, could AI provide new deductive methods to intervene in certain aspects of this field and accelerate research? This immediately opens up the concept of AI for Science, extending beyond just AI for materials or AI for biotechnology.

Another direction is AI hardware. Domestically, there is a mature hardware supply chain, a group of highly talented young founders and product managers, and increasingly robust AI infrastructure—enabling them to quickly turn ideas into products. What I’m most excited about is whether this environment will unexpectedly give rise to something we haven’t anticipated.

Liu Pengqi: AI hardware isn't just about providing users with a single function—it's closely connected to world models. While vast amounts of digital data have already been accumulated online, there is still a significant lack of data on how humans interact with the physical world, including behavioral data, vital signs, and brainwave data. In the future, these devices may carry computing power, AI software, and various sensors, gradually learning to understand us through use and then proactively delivering services in return.

Startups must also consider: Are you building a data collector or a data hub? If it can integrate information from users’ other contexts, the possibilities expand significantly.

Liu Pengqi: Finally, let’s talk about the capital markets. This year, Unitree has received approval for its IPO on the STAR Market and has entered the issuance phase; after going public earlier this year, Zhipu and MiniMax have experienced stock price volatility and even begun to diverge. Kimi, DeepSeek, and Jiepao Xingchen have each completed large-scale funding rounds. How do you view this wave of AI company IPOs? How will it impact the private equity market?

Yan Qianhang: From the perspective of a primary investor, it’s certainly positive if everyone can make money. For these foundational model companies, going public is also a way to realize capital value: their position in the ecosystem has already been partially validated, but they still have a longer battle ahead and will need more resources.

These companies have long-term value but require continuous cash burn on research in the short term, and capital markets have provided them with the conditions to keep investing. OpenAI and Anthropic are much the same. Evaluating them cannot rely entirely on traditional corporate P/E valuation frameworks; what matters more is the ecological position they ultimately secure. Similarly, when internet companies went public early on, investors didn’t focus solely on immediate profitability.

Liu Pengqi: But everyone acknowledges that there is a bubble in this round of AI. Will the bubble burst after more giants go public?

Yan Qianhang: It’s hard to predict exactly when a bubble will burst. Anyone can say the bubble will eventually burst, but that’s just a truism. In the short term, an IPO might fuel another wave of FOMO in the primary market; however, the secondary market reacts quickly—if a company’s business performance falls short or its research progress fails to justify expectations, the stock price will quickly reflect that. If leading companies perform well, there’s still room in the primary market; but if they struggle, they’ll drag down primary market valuations in turn. At that point, everyone will reassess: if secondary market valuations are unsustainable, why are primary market valuations still so high—or even inverted? Shouldn’t we take a step back? The valuations that surged rapidly in the first half of this year may also enter a period of recalibration.

Liu Pengqi: It’s impossible to predict exactly when the bubble will burst. Let’s rephrase the question—suppose the bubble does begin to burst, what would happen? Which companies are more likely to survive and even become truly great companies?

Yan Qianhang: The mindset in a bubble is much like stock trading—everyone believes the asset they bought will hit the daily price limit tomorrow. When the market cools, money flows toward better assets. For research-driven fields like world models and embodied intelligence models, companies must continuously deliver key milestones and maintain their position within the industry. Relying solely on demos or storytelling without tangible output makes it very difficult to secure funding in the future.

Ultimately, a company’s success depends on whether its business model is stable and its orders are genuine. Its ability to withstand shocks determines whether it can survive.

Liu Pengqi: I actually think that the bursting of a bubble isn’t necessarily all bad. Whether it’s software, large models, or embodied AI, the upstream costs of computing power and infrastructure may decrease. Many projects are currently stuck due to capital expenditure and cost constraints. Once users can afford them, adoption will accelerate.

Yan Qianhang: Yes, and for the primary market, going public isn’t the end. When it’s finally time to exit, the stock price still needs to return to its fair value, and the secondary market reacts quickly.

Liu Pengqi: Finally, if you had to bet on one variable for the second half of the year, what would it be?

Yan Qianhang: I focus on real data闭环. Whether it’s embodied intelligence or agents, as long as they can keep data and user feedback circulating, or help others build this mechanism, there’s opportunity—even in selling shovels.

Liu Pengqi: My assessment is similar. We should see more projects in the second half of the year that can be realistically implemented and generate revenue. It’s hard to predict when the bubble will burst, but technology cannot skip cycles and jump directly to AGI—there will inevitably be fluctuations along the way. Whether a company can successfully implement its technology, generate revenue, secure funding, or acquire data during these fluctuations determines whether it can build阶段性 advantages and have a chance to compete in the next round.