What are investors starting to worry about during the half-year when AI is at the peak of its popularity?

Recently, the AI sector in the secondary market has been highly volatile, with chips, memory, and optical interconnection experiencing successive fluctuations. Meta has raised its 2026 capital expenditure forecast once again, bringing the question of how long and how fast AI infrastructure investment will continue back to the center of discussion.

Capital is still pouring into computing power, storage, and data centers, and the market has begun to ask: Has AI infrastructure reached a phased peak? How long can this boom last? At the same time, embodied intelligence, world models, Agents, Tokens, and hardware changes from the cloud to the edge to the end are all taking place simultaneously.

More than half of 2026 has passed. Liu Pengqi and Yan Qianhang, technology investors at FreeS Fund, sat down to review the past six months: Why do project valuations change every week? What remains after the hype around OpenClaw fades away? Why have world models seen a concentrated explosion? What kind of computing power competition does "power in, Token out" imply? If the bubble is bound to burst sooner or later, which companies will stay in the game?

We have compiled part of the conversation. For the full discussion, you are welcome to search for "High Energy" on the Xiaoyuzhou App and Apple Podcast to listen to this episode.

Valuations Shift Every Week: What Investors Fear Is Not Just Missing Out

Liu Pengqi: Looking back at the middle of 2026, the evolution of AI has not slowed down at all. From OpenClaw and world models to storage and optical interconnection, and then to Zhipu's market value exceeding one trillion Hong Kong dollars, hotspots have emerged one after another. All of these are packed into the same six months, so today we can go through them and see what's really happening behind the scenes.

First question for Dr. Yan: How many projects have you roughly talked to in the first half of the year?

Yan Qianhang: About 200 to 300, of which nearly 150 to 200 are AI-related.

Liu Pengqi: My feeling is about the same. If you use a few keywords to describe your experience of looking at projects in the first half of the year, what would you choose?

Yan Qianhang: The first one is FOMO (fear of missing out). In hot fields, a large number of unfamiliar people and projects will suddenly emerge, and valuations are rising rapidly. Many people will force themselves to review all projects: Did we invest today? How much did we invest? Should we invest a little more?

The second is structural differentiation. A small number of popular tracks are very active in financing, while most tracks remain calm. The same is true in the secondary market: technology stocks are performing well, but stocks outside the hotspots may remain flat or even fall.

The third is advanced valuation. In the past, for early-stage investment, everyone had a relatively clear judgment on what valuation corresponded to what Milestone. After the company reached a consensus and completed phased delivery, the valuation would rise. Now the hype and FOMO have broken many past experiences, and a $1 billion valuation can appear at rocket speed in just a few months.

Liu Pengqi: Some projects see their valuations change within a week, and it's even so exaggerated that three rounds of financing are underway at the same time: one is closing, one is negotiating the agreement, and the third is finalizing the plan. Facing such a market, would you as an investor feel anxious?

Yan Qianhang: It's unrealistic to say I'm not anxious. But what makes me more anxious than missing a certain project is that AI is evolving too fast. Every day there are new models, new buzzwords, and new ideas. When you see a news that's a few days late, you can't help but wonder: Have I already been left behind?

Liu Pengqi: Some investments are value-oriented, while others are opportunity-driven. We need to analyze them separately and not let market sentiment make decisions for us. Putting valuations and FOMO aside for now, what new things have truly emerged in the first half of the year?

Yan Qianhang: Let me start with research-driven entrepreneurship. There's a much-discussed concept in the US called NeoLab, which uses commercial companies to solve long-term research problems. Quite a few such companies have emerged in the first half of the year. People no longer just ask "is this a professor-founded startup," but rather see whether the research can progress faster than the original organizational form after the professor starts the company.

Looking at AI applications and AI hardware, people have also begun to return to the product itself. After the emergence of Manus, the market once experienced FOMO around Agents; later, Insta360 brought a new wave of attention to AI hardware. After chasing a round of hotspots, everyone still has to find the people who can truly define products.

AI Coding has also allowed entrepreneurs to go beyond algorithm researchers and engineers. Lawyers, designers, product managers, and even housewives can use coding tools to create products.

Liu Pengqi: In the past, we often said that technical entrepreneurs were "looking for nails with a hammer." Now it's a bit like the nails themselves are starting a business, and the hammer has become a universal tool.

Blow Away the Foam of Agents, What Lies Beneath Is Rich, Flavorful Beer

Liu Pengqi: OpenClaw should be one of the most talked-about events in the first half of the year. From the public "raising lobsters" and scrambling for Mac minis, to model vendors and IM platforms launching their own Agents one after another, the hype has gradually cooled down. Open source brought it to the public at once, but problems with security and user experience have also been exposed. After the excitement, what exactly has it left behind?

Yan Qianhang: OpenClaw allowed many ordinary people to deploy Agents and download Skills for the first time, and then adjust them to their desired state through natural conversation. It's more like an Agent OS, not just a tool for completing a single task.

For example, a friend of mine who runs a beauty-focused We-Media account wanted to expand into tech content. She used OpenClaw to set up several Workflows to collect tech information in a targeted manner, forming a Pipeline for content production. Many people are also stably using it for Coding, PPT, e-commerce, and internal corporate process tasks. After a wave of boom, even though there are bubbles, fluctuations, and noise, if you blow away the beer foam, what's underneath is rich, flavorful beer.

Liu Pengqi: Recently, people have begun to emphasize Harness Engineering again. When large language models evolve from chat tools to task executors, what things become more important?

Yan Qianhang: An Agent must make decisions, execute tasks, and adjust based on results in an environment, with the entire process running continuously. The model is just the "brain" inside. Harness Engineering focuses on building the surrounding environment and feedback loops, allowing the model to run independently in the Workflow instead of having humans pre-program every single step.

Liu Pengqi: This is also related to the data flywheel. The personalized experience in traditional internet comes from accumulated user data. Could Agents become the carrier for forming data flywheels and personalized experiences?

Yan Qianhang: Chatbots mainly leave behind the linguistic context, while Agents leave behind task trajectories: how it makes judgments, how it executes, what tools it calls, and whether it finally completes the task. These trajectories can be used for Agent RL. Data from vertical scenarios may not necessarily flow back to the foundation model companies, which could become the moat for Agent startups.

Liu Pengqi: From this perspective, whoever can access users' task trajectories will have more initiative. There are roughly three types of players in the market now: model vendors, IM and office platforms, and local hardware. Which one are you more optimistic about?

Yan Qianhang: They do different things. Model vendors are suitable for general tasks such as search, in-depth research, and Coding; office platforms have ready-made files, documents, and business processes; local hardware is more related to privacy. But whether the "lobster machine" is a valid product route is still questionable. The most sought-after product at the beginning was the Mac mini, and after the hype cooled down, the Mac mini is also the most sold on Xianyu.

There is another problem with local hardware: who owns the data? For example, if a company's core architect uses an Agent, he certainly wants to use the best model, but if the architectural design and task trajectories are taken by the model company for training, it poses a great risk to him. Many users face this conflict: I want to use the best model, but I can't accept giving my data to it. What's your take on the trade-off between privacy and model capabilities?

Liu Pengqi: Based on past experience, most C-end users will prioritize efficiency and experience first. But Pro C and B-end users are different: technical documents, product designs, and business data are all core assets of the enterprise. The opportunity may lie here: allowing enterprises to use the best models while keeping local data, permissions, and security under control.

Yan Qianhang: Pengqi and I both looked at SaaS and Fintech in our early years, and we have a love-hate relationship with large B-end customers: they have big budgets, but their needs are complex, making service delivery very heavy. Now there's the addition of FDE (Field Deployment Engineer), who needs to bring model and business expertise on-site to help customers deploy. How will AI serving large B-end customers be different from the SaaS and Fintech of the past?

Liu Pengqi: Large B-end customers have large budgets and high requirements, and old problems like "more features without price increases" and high customization costs still exist. Customers will also ask: Since there are open-source models, why should I pay extra for AI services? When AI truly penetrates into the underlying systems of enterprises, issues like data security, privacy, and system integration need to be addressed. There's also the hallucination problem: if AI is used in core businesses such as risk control and causes losses due to errors, whose responsibility is it—the model vendor, the enterprise, or the employees? This boundary hasn't been clearly defined yet. Whoever can solve these problems will have a chance.

Power in, Token out: AI Infrastructure Enters System Competition

Liu Pengqi: Over the past six months, has the speed of capability improvement in large models been accelerating or slowing down?

Yan Qianhang: If we look at leapfrog innovations like the o1 reasoning architecture, we indeed haven't seen anything like that for a while. But the models' ability to handle real tasks is still improving. In the past, we looked at math tests and knowledge rankings, which was like making models take exams; now we look at SWE-bench, Computer Use, and Agent Benchmark, which test whether the model can fix bugs, adjust tools, and complete long tasks. Models have moved from the examination room to the workplace.

Liu Pengqi: It's a bit like a student graduating from school and stepping into society. The knowledge reserve may have been basically formed, and the real assessment is how much capability he can exert at work. Which company do you think has stood out the most in this round?

Yan Qianhang: I would pick Anthropic. It bet heavily on Coding very early, and now this decision looks very critical. It realized earlier than others that models should not just compete on exam scores, but also on task execution. After Claude 3.5 was released, the experience of AI Coding tools improved significantly, and products like Cursor became more and more user-friendly; better tools attract more users, whose feedback flows back to the model, gradually forming a positive flywheel.

Another reason is its rapid revenue growth. Anthropic's annualized revenue, which it has disclosed publicly several times this year, is rising rapidly. People have begun to discuss: if Anthropic goes public, what scale will the company reach? The improvement in reasoning cost and profitability has also led many people to reassess it.

Liu Pengqi: It sounds like it chose a market that seems vertical but is actually large enough, and deeply penetrated this group of users first. That's better than spreading out in all directions without being a leader in any one area.

Yan Qianhang: I wouldn't call Coding a vertical field. Coding inherently involves human logic, reasoning, and decision-making, to turn an algorithm into a runnable Application. When models deepen their capabilities in Coding, reasoning, associative thinking, and chained decision-making will all be practiced accordingly. That's why people find Claude rigorous and smart when they use it.

Some models are creative but not reliable enough. Anthropic entered through Coding and Agents, and real tasks are in turn refining its models. It's a bit like the "unity of knowledge and action": it put "knowledge" and "action" into the same loop earlier than others.

This year, Anthropic also launched Fable 5 and Mythos 5. Fable 5 is open to regular users, while Mythos 5 has a more limited access scope; both models were once suspended, and then access was restored. What's your view on the discussions surrounding these restrictions?

Liu Pengqi: There are already many reviews of these models online, but not many people have truly used them in depth. After a quick look, I found that their performance in terms of context length, adaptive loop thinking, and integration with Agent capabilities is indeed very strong. As for the so-called "ban," I think there may be some smoke bombs and hunger marketing elements in it. Whether they are really powerful enough to become tools for competition between nations remains to be seen.

Yan Qianhang: Turning back to domestic models, what do you think is the biggest change in the past six months?

Liu Pengqi: Zhipu is a very typical example. It is also focusing on Coding, making its own attempts in architecture and long-horizon task training. We've also heard that the team has introduced methods such as process reward to enhance model capabilities. Domestic teams are constantly pushing forward, with no pauses in architecture, algorithms, or model capabilities.

The data side is also being improved. If you only focus on distillation, you'll hit the ceiling very quickly. After model companies get more funding, in addition to purchasing computing power and recruiting talents, they are also willing to invest in high-quality data.

DeepSeek is also worth watching. After releasing its model, it stayed quiet for a period of time, but has been continuously accumulating its capabilities. This year, there's a noticeable change: it has adopted a more open attitude towards financing and domestic computing power partnerships than before.

I understand that it needs to retain talents through valuation and equity incentives, and hopes to build deeper partnerships with domestic computing power vendors. This not only improves its own models, but also drives the entire industrial ecosystem.

Yan Qianhang: When DeepSeek first emerged, the capital scale of several domestic model companies was mostly in the tens of billions. At this scale, a Research Lab supported by a quantitative fund could still compete with everyone. From last year to this year, large tech companies have begun to invest heavily: Alibaba and Xiaomi are pouring money into the field, and after MiniMax and Zhipu go public, their capital is no longer at the tens of billions level. If DeepSeek continues to operate in its old way, it will actively make things more difficult for itself.

But DeepSeek's style hasn't changed much. All its choices still revolve around one thing: building a better large model, rather than pursuing short-term commercialization first.

Liu Pengqi: After talking about several domestic model companies, let's look at the gap between China and the US. Do you think this gap is widening or narrowing?

Yan Qianhang: In the fast-version and general standard models, the gap between China and the US is indeed narrowing. In 2022, people thought it was hard to catch up; in 2023 and 2024, people said the gap was about one or two years; this year, the feeling is that the gap may only be a few months. Recently, Tang Jie, co-founder of Zhipu, had a conversation with Elon Musk on X. Musk said that China's large models might catch up with the cutting edge level in the first quarter of 2027; Tang Jie replied, "It won't take that long."

The gap that's hard to overcome in the short term still lies in infrastructure and computing power. The US has larger computing clusters that can run more experiments simultaneously; China is better at achieving big results on a small budget and catching up quickly through algorithms and engineering efficiency.

Liu Pengqi: Exactly. With more machines, you can run more experiments at the same time, leaving more room for exploring model architectures. China's ability to supplement data is stronger than before, and the contribution of data to model improvement is also increasing. Looking at the fast-version and lightweight models, we do have advantages in efficiency and cost.

Yan Qianhang: Chinese models are also being noticed by more people. MiniMax's model is supported by OpenClaw; in the agenda and related speeches of Nvidia's GTC in March, Kimi and MiniMax were both mentioned. We're catching up fast in the fast, standard, and general-purpose model layers.

Video generation is even more interesting, as Chinese players have a louder voice in this field. After Seedance was released, there were market rumors that its ARR had reached $2 billion, but Volcano Engine later stated that the revenue figures circulating publicly