What Are Investors Worried About in AI's Hottest Six Months?

Lately, the AI sector in secondary markets has been on a rollercoaster — chips, memory, and optical interconnects taking turns in the spotlight. Meta has revised its 2026 capex forecast upward again, pulling the question of how long and how fast AI infrastructure spending will continue back into center stage.

Money keeps pouring into compute, storage, and data centers, and the market has already started asking: Has AI infrastructure hit a temporary peak? How much longer can this boom run? At the same time, embodied intelligence, world models, Agents, Tokens, and hardware shifts from cloud to edge to end device are all unfolding simultaneously.

With 2026 more than halfway through, FreeS Fund tech investors Pengqi Liu and Qianhang Yan sat down to review the past six months: Why do project valuations change week to week? What remains after the OpenClaw hype died down? Why did world models explode all at once? What kind of compute competition does "power in, Token out" signal? And if the bubble does burst eventually, which companies will still have seats at the table?

We've edited excerpts from their conversation below. For the full discussion, search for "Gao Neng Liang" on Xiaoyuzhou App or Apple Podcasts.

We continue to track technological developments in the agent space. If you're a founder or practitioner in this field, feel free to reach out to Pengqi Liu (pengqi@freesvc.com) or Qianhang Yan (qianhang@freesvc.com).

Pengqi Liu: Looking back from mid-2026, AI's evolution hasn't slowed down one bit. From OpenClaw and world models to storage and optical interconnects, to Zhipu AI's market cap breaking a trillion Hong Kong dollars — the hot topics have come in waves, all compressed into the same six-month window. Today we'll walk through them in order and see what's actually been happening behind the scenes.

First question for you, Dr. Yan — roughly how many projects did you look at in the first half of the year?

Qianhang Yan: About 200 to 300, with AI-related projects making up roughly 150 to 200 of those.

Pengqi Liu: My sense is about the same. If you had to describe the first half in a few keywords, what would you pick?

Qianhang Yan: The first is FOMO (fear of missing out). Hot domains suddenly produce a flood of people and projects you've never seen before, with valuations rising fast. A lot of people force themselves to look at every single project: Did we invest today? How much? Should we be putting in more?

The second is structural divergence. A handful of hot tracks are seeing very active fundraising, while most sectors remain cold. The secondary market is the same — tech stocks are doing well, but stocks outside the spotlight may be lukewarm or even falling.

The third is valuation front-running. In early-stage investing, there used to be a relatively clear sense of what milestone matched what valuation. The company would build consensus, deliver on a phase, and then the valuation would move up. Now heat and FOMO have broken a lot of that experience — a $1 billion valuation can appear rocket-fast within months.

Pengqi Liu: Some projects see their valuations shift in a single week, even to the point where three funding rounds are happening simultaneously — one closing, one in term sheet discussion, another already nailing down terms. Facing a market like this, do you get anxious as an investor?

Qianhang Yan: It'd be unrealistic to say no. But what makes me more anxious isn't missing out on a particular project — it's how fast AI itself is changing. Every day there's a new model, a new buzzword, a new idea. You see a news item a few days late and can't help thinking: Have I already been left behind?

Pengqi Liu: Some investing is value-oriented, some is opportunity-oriented — you still have to break it down and not let market sentiment make decisions for you. Setting aside valuations and FOMO for a moment, what actually grew into something new in the first half?

Qianhang Yan: Let me start with research-driven entrepreneurship. There's a concept that's been discussed a lot recently in the United States called NeoLab — using a commercial company structure to tackle long-term research problems. A number of such companies emerged in the first half. People are no longer just asking "Is this a professor founding a company?" but rather, once the professor starts a company, can the research run faster than under the old organizational model?

Looking at AI applications and AI hardware, people are also returning to the product itself. After Manus appeared, the market went through a round of FOMO over Agents; then Insta360 sparked renewed attention on AI hardware. After chasing a round of hot topics, everyone still has to come back and find people who can actually define products.

AI Coding has also expanded entrepreneurship beyond algorithm researchers and engineers. Lawyers, designers, product managers, even stay-at-home parents — anyone can use coding tools to build products.

Pengqi Liu: People used to say technical founders were "walking around with a hammer looking for nails." Now it's a bit like the nails themselves are starting companies, and the hammer has become a universal tool.

Pengqi Liu: OpenClaw was probably one of the most watched events of the first half. From everyone "raising lobsters" and scrambling for Mac minis, to model vendors and IM platforms launching their own Agents, the heat has gradually cooled since. Open source brought it to the masses overnight, but security and experience issues surfaced right alongside. After the noise, what did it actually leave behind?

Qianhang Yan: OpenClaw let many ordinary people deploy an Agent themselves for the first time, download Skills, and gradually tune it through natural conversation into what they wanted. It's more like an Agent OS — not just a tool for completing a single task.

I have a friend in beauty content creation who wanted to pivot into tech content. She used OpenClaw to build several workflows, gathering tech information directionally and forming a content production pipeline. Coding, PPTs, e-commerce, and internal enterprise workflows — plenty of people are already using it stably. After the wave of hype, there was foam, volatility, and noise, but blow away the beer foam and what's underneath is the real brew.

Pengqi Liu: Recently people have started emphasizing Harness Engineering again. As large models shift from chat tools to task executors, what becomes more important?

Qianhang Yan: An Agent has to make decisions in an environment, execute tasks, and adjust based on results — the whole loop has to actually turn. The model is just the "brain" inside. Harness Engineering builds out the surrounding environment and feedback pathways so the model runs through workflows on its own, rather than having every step pre-scripted by humans.

Pengqi Liu: This also connects to data flywheels. Traditional internet personalization came from accumulated user data — could Agents become the vehicle for forming data flywheels and personalized experiences?

Qianhang Yan: What Chatbots leave behind is mostly linguistic context. What Agents leave behind is task trajectories: how it judged, how it executed, what tools it called, whether it actually got the thing done. These trajectories can be used for Agent RL (reinforcement learning). Vertical-scenario data may not flow back to foundation model companies, and that's where Agent startups could build real moats.

Pengqi Liu: Seen this way, whoever captures users' task trajectories has more initiative. Right now there are roughly three types of players on the field: model vendors, IM and workplace platforms, and local hardware. Who do you favor?

Qianhang Yan: They're doing different things. Model vendors are suited for general tasks like search, deep research, and coding; workplace platforms have existing files, documents, and business processes; local hardware connects more to privacy. But whether the "lobster machine" is a viable product path is still a question mark. The Mac mini was what people scrambled for most aggressively at the start, and after the heat died down, it was also what sold most on Xianyu.

Local hardware also has a problem: who owns the data? Say a company's core architect uses an Agent — of course they want the best model, but if their architectural designs and task trajectories get taken by the model company for training, that's a major risk. A lot of users face this conflict: I want to use the best model, but I won't accept giving it my data. How do you see the tradeoff between privacy and model capability?

Liu Pengqi: Based on past experience, most consumer users prioritize efficiency and experience first. But Pro C and B2B are different — technical documentation, product designs, and business data are core corporate assets. The opportunity might lie right here: letting enterprises use the best models while keeping local data, permissions, and security under control.

Qianhang Yan: Pengqi and I both looked at SaaS and fintech years ago, and we have a love-hate relationship with large enterprise clients: big budgets, but complex demands and heavy service overhead. Now there's also the FDE (Frontier Deployment Engineer) — you need people who understand both models and business to go on-site and help clients deploy. How will AI serving large enterprises differ from SaaS and fintech of the past?

Liu Pengqi: Large enterprise clients have big budgets and high expectations. The old problems still persist: more demands without more pay, high customization costs. Clients will also ask: if open-source models exist, why pay extra for AI services? Once AI truly penetrates enterprise infrastructure, data security, privacy, and system integration all need to be addressed. Then there's hallucination — if AI is used in core businesses like risk control, and errors cause losses, who bears responsibility: the model vendor, the enterprise, or the employee? That boundary hasn't been clearly drawn yet. Whoever solves these problems gets the opportunity.

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

Qianhang Yan: If you're looking at leapfrog innovations like the o1 reasoning architecture, there hasn't been one for a while. But models' ability to perform real tasks keeps improving. In the past we looked at math tests and knowledge leaderboards — like making models take exams. Now we look at SWE-bench (software engineering benchmark), Computer Use, and Agent Benchmark — testing whether it can fix bugs, orchestrate tools, and complete long tasks. Models have moved from the examination hall to the actual workplace.

Liu Pengqi: A bit like students graduating from school and entering society. Their knowledge base may be largely formed; the real test is how much capability they can demonstrate at work. Of the companies performing most prominently in this round, who would you pick?

Qianhang Yan: I'd pick Anthropic. It bet heavily on coding very early, and that decision looks crucial in hindsight. It realized sooner than others that models can't just compete on test scores — they need to be judged on task execution. After Claude 3.5 came out, the experience of AI coding tools improved noticeably; products like Cursor became increasingly usable. Better tools attract more users; feedback flows back into the models, and slowly it starts to snowball.

Another reason is its rapid revenue growth. Anthropic's publicly disclosed annualized revenue has been climbing quickly this year, and people have started discussing: if Anthropic goes public, what scale of company would it be? Improvements in inference costs and profitability have also led many to reassess it.

Liu Pengqi: Sounds like it chose a market that looks vertical but is actually large enough, and penetrated that user base deeply first. Better than spreading thin without leading in any single direction.

Qianhang Yan: I wouldn't quite call coding a vertical domain. Coding inherently involves human logic, reasoning, and decision-making — turning an algorithm into a running application. When models go deep on coding, reasoning, associative thinking, and chain-of-thought decision-making all get trained alongside. That's why people using Claude find it rigorous and intelligent.

Some models are creative but not reliable enough. Anthropic cut in through coding and agents, and real tasks are in turn training the model. A bit like "the unity of knowledge and action": it put "knowing" and "doing" into the same loop earlier than others.

This year Anthropic also launched Fable 5 and Mythos 5. Fable 5 was opened to general users, while Mythos 5 had narrower access; both models were temporarily suspended and later restored. What do you make of the discussions around these restrictions?

Liu Pengqi: There's already quite a bit of online commentary about these models, but the number of people who've genuinely used them in depth may still be limited. From my brief look, they are indeed very strong in context length, adaptive recursive reasoning, and integration with agent capabilities. As for the so-called "ban," I suspect there may also be an element of smoke and mirrors and scarcity marketing. Whether they're truly powerful enough to become tools of competition between nations remains to be seen.

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

Liu Pengqi: Zhipu AI is a very typical example. It's also pushing toward coding, making its own attempts in architecture and long-horizon task training. We've also heard the team has incorporated process rewards and other methods to improve model capability. Domestic teams are still constantly experimenting — architecture, algorithms, model capabilities haven't stopped.

The data piece is also being addressed. Distillation alone hits a ceiling quickly. After model companies raised more money, beyond buying compute and hiring, they've also become willing to spend on high-quality data.

DeepSeek is also worth watching. After releasing its model, it went quiet for a while but has been building its own accumulation. A notable change this year is that its attitude toward fundraising and domestic compute partnerships has become more open than in the past.

I understand it needs to retain talent through valuation and equity incentives, and also hopes to establish deeper cooperation with domestic compute vendors. Doing so isn't just about improving its own models — it's also driving the entire industrial ecosystem.

Qianhang Yan: When DeepSeek first emerged, most domestic model companies still had funding in the range of a few billion RMB. At that scale, a research lab backed by a quant fund could still compete with everyone. From last year to this year, the tech giants started investing heavily — Alibaba, Xiaomi are pouring in money; after going public, MiniMax and Zhipu AI no longer have just a few billion RMB on hand. If DeepSeek continues with its original operating model, it's effectively increasing the difficulty for itself.

But DeepSeek's style hasn't changed much — all its choices still revolve around one thing: making better large models, rather than pursuing short-term commercialization first.

Liu Pengqi: Having discussed several domestic model companies, let's look at the China-US gap. Do you think this gap is widening or narrowing?

Qianhang Yan: For turbo models and general standard models, the China-US gap is indeed narrowing. In 2022 people thought catching up would be very difficult; in 2023 and 2024 they were still saying the gap was one to two years; by this year the gut feeling might be just a few months. Recently Zhipu AI co-founder Jie Tang had an exchange on X with Elon Musk. Musk said Chinese large models might not catch up to frontier levels until Q1 2027; Tang replied, "It won't take that long."

The gap that's still hard to overcome in the short term remains infrastructure and compute. The United States has larger compute clusters and can run more experiments in parallel; China is better at doing more with less, catching up quickly through algorithmic and engineering efficiency.

Liu Pengqi: Right. More machines mean more experiments running simultaneously, and more room to explore model architectures. Domestic capabilities in data supplementation have improved compared to before, and data's contribution to model progress is also increasing. Looking at turbo and lightweight models, we do have advantages in efficiency and cost.

Qianhang Yan: Chinese models are also starting to get more recognition. MiniMax's models are supported by OpenClaw; at NVIDIA's March GTC, both Kimi and MiniMax were mentioned in the agenda and related talks. Across turbo, standard, and general tiers, we're catching up fast.

Video generation is more interesting — domestic players are actually louder here. After Seedance's release, the market briefly circulated rumors that its ARR had reached $2 billion, though Volcano Engine later stated that externally circulated revenue figures were generally inflated. Overseas, beyond Google Veo 3.0 still maintaining attention, Sora's market buzz has diminished considerably compared to its initial launch. On video models, domestic players actually look stronger.

Liu Pengqi: Because video generation consumes a lot of tokens, the cost and efficiency per token get magnified exponentially. Domestic vendors can bring costs down to user-acceptable ranges through architectural design and more efficient compute utilization.

Going further, we can't avoid AI infrastructure. Model vendors keep raising capex, token consumption is rising, and compute costs will eventually become a bottleneck. What changes do you see in infrastructure?

Qianhang Yan: What people are asking now isn't just "is there enough compute," but whether the entire system can run smoothly. Beyond "computing," you have to look at "storage" and "transmission." Suppose you have ten thousand compute nodes — if results can't be transmitted to the next node in time, other nodes just have to wait. It's like hiring ten thousand assembly line workers: if materials arrive too slowly, someone's always slacking off.

That's why optical interconnect, HBM (high-bandwidth memory), DRAM (dynamic random-access memory), and SSD (solid-state drive) are becoming increasingly important. Abstract the whole system, and it's really one thing: "power in, tokens out." How to convert electricity and operating expenses into more tokens requires joint optimization across chips, storage, networking, liquid cooling, and software.

Liu Pengqi: Looking at infrastructure, you can't just look at the cloud. The device side covers phones, watches, earphones — devices around you; the edge side can be NAS (network-attached storage), all-in-one machines, or enterprise internal equipment; the cloud then provides large-scale compute. Simple tasks go on-device, personal and enterprise private data stays at the edge, and complex programming and long-horizon reasoning go to the cloud.

Qianhang Yan: The cloud side has absorbed massive resources in the short term, while the device side has to deliver functionality with less compute and lower cost. For startups, this constraint also hides opportunities.

Liu Pengqi: In the first half of 2026, "world model" suddenly became a buzzword. Embodied intelligence, video generation, gaming, physics simulation — everyone was talking about it. My understanding is that a world model doesn't necessarily correspond to one specific model; it's more like: given an external intervention to the world, predict what happens next.

By this definition, couldn't Newton's laws also be understood as a kind of world model? Dr. Yan, how do you understand world models? What is their relationship with VLA (vision-language-action models), video generation models, and large language models respectively?

Qianhang Yan: World models aren't actually a new concept — they've always existed in reinforcement learning. Now we can see several approaches: video models generating the next frame, 3D world models building out interactive spaces, JEPA (Joint Embedding Predictive Architecture) abstracting the world into latent space representations, and WAM (World Action Model) connecting world models with action policy.

Liu Pengqi: Language is highly abstract information, but the real world also includes vision, touch, and spatial relationships. I think world models are more like innate human abilities, while language models are more like acquired learning. They iterate differently, but together they shape human intelligence. Do you agree with this assessment? What kind of relationship will world models and LLMs form in the future?

Qianhang Yan: I also lean toward them being complementary. Language and the five senses — including vision, hearing, touch, and smell — are all ways humans interact with the world. Large language models have already solved intelligence at the language level, so naturally we'll 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?

Qianhang Yan: First, video models matured — the generated results look increasingly real. 3D representation technology also advanced, from Mesh, to point clouds, to NeRF (Neural Radiance Fields), to 3D Gaussian Splatting. Startup teams already have the technical foundation to generate and manipulate 3D worlds. Plus, with open-source video models, many ideas can be validated quickly.

Liu Pengqi: But both world models and embodied intelligence lack data containing physical information. The industry often talks about a data pyramid: internet videos and synthetic data at the base, first-person data with multimodal information in the middle, and real robot teleoperation data at the top. How do we solve this bottleneck?

Qianhang Yan: We don't lack ordinary video data — what we lack is data that encodes physical states and laws. There are three types: first, real sensor collection; second, physics simulation and digital twins; third, inferring contact and force states from video. Real data is best but most expensive. Ultimately, we need to combine all three into a gradient system.

Let's switch to something lighter. If world models start showing clear commercial progress, in which scenarios do you think consumers will feel it first — gaming, content consumption, or embodied intelligence, which everyone's been talking about lately?

Liu Pengqi: My guess is gaming and content consumption. These scenarios don't demand absolute accuracy — they just need to fool the human eye. Embodied intelligence operates in the real world, with higher requirements for data, success rates, and safety.

Qianhang Yan: If world models can really be used in games, I'd quite look forward to no longer seeing all kinds of bizarre clipping and bugs.

Liu Pengqi: Moving from content consumption to embodied intelligence, what changes have world models brought to this industry? Some people now think VLA will be replaced by world models. Is that really the case?

Qianhang Yan: I don't think so. When people play tennis, facing an unfamiliar shot they might first imagine how to return it; but with a fast incoming ball, professional athletes rely on bodily instinct to counter directly. World models are more like "think first, then act," while VLA is more like directly outputting actions from visual input. The two can complement each other.

Liu Pengqi: There's also an "impossible triangle" for embodied intelligence deployment: long-horizon complex tasks, high success rates, and cross-scene generalization — currently hard to satisfy simultaneously. Touch is also critical. If a robot can reconstruct an object's approximate 3D shape using only tactile sensors on its hand, that indicates a fairly high level of capability.

Qianhang Yan: We saw a very intuitive case. A simple two-finger gripper costing just one or two hundred yuan, plus tactile sensors forming a force control loop, could pick up tofu without crushing it. But how to adapt touch to models is still in very early stages.

Liu Pengqi: Besides the hottest areas like Agent, world models, and embodied intelligence, what other directions do you think deserve entrepreneurs' special attention?

Qianhang Yan: One direction I've been most interested in recently is the broader AI for Science. Break down the scientific research process — from asking questions, reviewing literature, to experiments, analysis, and validation — and see which steps can be handed to AI, and whether research speed can be accelerated. It goes a layer beyond AlphaFold, which people were familiar with a few years ago.

Liu Pengqi: A bit like turning scientific research itself into a reorganizable Pipeline.

Qianhang Yan: Exactly. This year we've already seen quite a bit of Auto Research work, though mostly still limited to certain fields. What's more worth watching next is whether it can enter more disciplines and greatly accelerate scientific innovation. Another example relates to world models we discussed earlier. Traditional physics simulation often relies on solving partial differential equations at the底层. There used to be a painful tradeoff in simulation: accuracy meant slowness, speed meant inaccuracy. Now people are starting to use AI and neural operators to replace part of the PDE solving. If it can truly be both fast and accurate, that's solid progress.

Something I've found particularly interesting recently is social science. Previously, social science research lacked efficient experiment and推演 mechanisms: either treat real society as an experimental environment for large-scale surveys, or do small-scale sampling. Can AI now provide new推演 methods for social science,切入 some of these steps and accelerate research? This immediately opens up the AI for Science思路 — it's no longer just AI for materials, AI for biotech.

Another direction is AI hardware. China has mature hardware supply chains, a batch of very smart young founders and product managers, plus increasingly good AI infrastructure. They can quickly turn ideas into products. What I'm more curious about is whether this soil will randomly grow some things we didn't anticipate.

Liu Pengqi: AI hardware isn't just providing users with a function — it's also closely connected to world models. The internet has already沉淀大量数字化数据, but how humans interact with the physical world, including behavioral data, vital signs data, and brainwave data, is still very lacking. In the future, these hardware devices may carry computing power, AI software, and various sensors, gradually understanding us through use, then反过来提供服务.

Startups still need to think clearly: are you building a data collector, or a Data Hub? If it can also bring in users' information from other scenarios, there's much more it can do.

Liu Pengqi: Finally, let's talk about capital markets. This year Unitree has already received IPO registration approval for the STAR Market and entered the issuance stage. After Zhipu AI and MiniMax listed earlier this year, their stock prices experienced fluctuations and even began diverging. Moonshot AI, DeepSeek, and StepFun have completed large funding rounds in succession. How do you view this wave of AI company listings? How will it affect the primary market?

Qianhang Yan: From a primary market investor's perspective, everyone making money is of course a good thing. For these foundation model companies, going public is also a capitalization兑现: their ecosystem position already has阶段性证明, and they still need to fight longer battles ahead, requiring more弹药.

These companies have long-term value, but in the short term they need to keep burning money on research. Capital markets give them the conditions to continue investing. OpenAI and Anthropic are similar. You can't fully value them using traditional PE frameworks — people care more about what ecosystem position they ultimately occupy. When internet companies went public early on, people didn't just look at whether they were making money right then either.

Liu Pengqi: But in this round of AI, everyone acknowledges there's a bubble. After more giants go public, will the bubble burst?

Qianhang Yan: It's hard to predict which day the bubble will burst. Anyone can say bubbles eventually burst, but that's a correct yet useless statement. In the short term, listings may add another wave of FOMO to the primary market; but secondary markets反馈快 — if the business isn't doing well, or research progress can't support expectations, stock prices will reflect that quickly. If头部 companies perform well, there's still room in the primary market; if they can't sustain it, that will反过来压低 primary market valuations. Then people will recalculate: if secondary market valuations can't hold, why is the primary market still so expensive, even with一二级倒挂? Should we calm down? Valuations that rose quickly in the first half of this year may also face a cooling-off period.

Liu Pengqi: Can't guess which day exactly. Then let me rephrase — assuming the bubble starts bursting, what happens? Which companies are more likely to survive, or even become truly good companies?

Qianhang Yan: The mentality in a bubble is a lot like stock trading — everyone thinks their pick will hit the涨停 tomorrow. When markets cool, money flows toward better assets. For research-driven directions like world models and embodied intelligence models, companies need to keep delivering key milestones and maintain their position in the industry. Relying only on demos and storytelling without outputting real成果 makes it hard to raise more money later.

For companies focused on落地, it ultimately comes down to whether the business model is stable and orders are real. Whether they can withstand shocks determines whether they survive.

Liu Pengqi: I actually think a bubble bursting isn't necessarily all bad. Whether for software, large models, or embodied intelligence, upstream computing and infrastructure costs may come down. Many projects can't落地 now because they're stuck on capital expenditures and costs. If users can afford it,落地 will actually accelerate.

Qianhang Yan: Right, and for the primary market, listing isn't the endpoint either. When it's time to actually exit, stock prices still need to return to reasonable value — secondary markets react quickly.

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

Qianhang Yan: I'd watch for real data flywheels. Whether embodied intelligence or Agent, whoever can keep data and usage feedback continuously cycling, or help others build that mechanism, has a chance — even "selling shovels" works.

Liu Pengqi: My assessment is similar. The second half should see more projects that can落地 and generate revenue. When the bubble bursts is hard to guess, but technology can't skip cycles and jump straight to AGI — there will definitely be fluctuations in between. Whether you can落地, get revenue, funding, or data during those fluctuations determines whether a company can build阶段性优势 and have a chance to compete in the next round.