Written by: Xiaobing
1. WAIC should have been held earlier, so that after visiting, one could clear out their tech holdings sooner.
There are too many signs of a peak: various side events are becoming increasingly crypto-centric, attractive women are everywhere, and the products on-site are highly homogenized… a sector has transformed from a game for geeks into a feast for the masses.
Whether in the primary or secondary market, when everyone starts believing the same story, and consensus becomes too concentrated, risks begin to accumulate.
Whether in the primary or secondary market, when everyone starts believing the same story, and consensus becomes too concentrated, risks begin to accumulate.
Because those who were bullish alongside you during the bull market will become sellers in the stampede during the downturn. This is a classic case of "long squeeze!" 2. Large models lack a true moat. The foundation of AGI remains the Scaling Law. The relationship between model capability, parameter size, training computational power, and data quality still follows a power law. Who can acquire GPU computing power more cheaply, higher quality data, better researchers, and burn money more efficiently? Therefore, the moat of large model companies is more fragile than many people imagine. Model capabilities will continue to converge, leading advantages will be constantly caught up, and prices will continue to decline. Those who are truly making sustained profits in this wave are those who sell water to all model companies; after all, everyone is just profiting from the capital expenditures of a few big companies. GPUs, HBM, high-speed interconnects, data centers, power, data services… Standing upstream in the arms race to get a share, or even just acting as an intermediary connecting resources, is currently the most profitable business. 3. Data Bottleneck is a potential opportunity to see at this conference. After talking with some friends, I found that companies making money in this wave of AI are very low-key and don't even participate in the exhibition; one area is data. The success or failure of unified multimodal representation essentially depends on high-quality multimodal training data. High-quality, accurately labeled, cross-modal aligned, and continuously iterative data assets are severely lacking. Models will become increasingly cheaper, GPUs will become more and more common, and truly high-quality data will become increasingly scarce. The multimodal data industry chain is becoming a highly profitable sector. Data cleaning, data labeling, data synthesis, vertical data assets, robot data collection… and algorithm companies that specialize in solving cross-modal unified representation and multimodal pre-trained encoders. GPUs are a one-time capital expenditure, while data is a continuous capital expenditure. I currently categorize data into three types: The first is expert data; the second is RL environment/agent data. In the future, training agents will no longer simply involve collecting question-answers, but rather constructing an Environment → Task → Trajectory → Reward → Verifier, which may be one of the largest incremental markets in the future. The third type is embedded/robotics data, which is even scarcer than LLM data and extremely difficult to collect. 4. The Dilemma of AI Applications: A Microcosm of the Crypto World's Past AI today is very similar to Crypto in the past: a fat protocol where L1 captures the vast majority of value. That's why Crypto VCs poured money into public chains, but applications? Not even a dog would invest. Currently, AI is in a similar situation. The large model is essentially a POW Layer 1, or even the model itself is the application. Over the past few years, most AI applications have been doing something dangerous: packaging capabilities that models don't yet possess into products. However, the boundaries of model capabilities have been constantly expanding. Each model upgrade is like L1 directly writing application-layer functionalities into the protocol: search, deep research, coding, image generation, video generation, computer use, agent… If each model upgrade reduces the value of your product, then you are essentially creating an AI feature, not an AI application. Therefore, you must create something that the model can do, but is also difficult to take away. Thinking about it, it boils down to data, context, workflow, permission, and distribution. The moat for AI applications is to become a customer of the model, not a competitor. 5. Embodied Intelligence, a Huge Bubble On the second floor, in Hall H4, watching numerous robots from the same supply chain slowly performing similar actions, and then looking at the valuations of various companies, I felt a chill run down my spine and a pang of sympathy for investors' money. The current problem with embodied intelligence is that capital is using the "software scaling law" to price "complex hardware and software systems." The miracle of large-scale models lies in the fact that a company training GPT-5 can theoretically serve hundreds of millions of people globally at near-zero marginal cost, but robots cannot do this. For every additional user a robot serves, one more machine needs to be built, involving BOM, manufacturing, supply chain, delivery, maintenance, and depreciation. AI intelligence can improve exponentially; for example, with each model upgrade, all users worldwide simultaneously become smarter, but the costs in the physical world do not decrease exponentially. If intelligence follows Moore's Law, then embodied intelligence will still follow the laws of manufacturing. Embodied intelligence may ultimately be a huge industry, but it may not necessarily be an industry with high-profit margins. Furthermore, it's too early to say. Autonomous driving is already a highly constrained embodied intelligence problem: clear objectives, limited room for maneuver, standardized road rules, massive amounts of real-world data, and a mature automotive industry system. Even so, this industry has burned through hundreds of billions of dollars and still hasn't fully solved some of the long-tail problems on the road. The real difficulty faced by embodied intelligence is many times greater than that of autonomous driving. 6. The Best Era for Pimps A harsh truth is that most AI startups aren't making much money, their profits are even less than what the famous woman Ququ sells her skills for. So, something interesting has emerged: many companies appear to have one business, but upon closer inspection, it turns out they also sell tokens, or computing power. It turns out that even large model companies resell computing power for profit. The real money is made by the pimps. FA (Financial Advisor), facilitating existing shareholders' transactions, trading B300, selling tokens, selling datasets, selling computing power, even selling an opportunity to meet a founder… AI simultaneously fulfills all the conditions for a thriving intermediary market: rapid technological change, a large enough information gap, ample capital, and a compelling narrative. VCs earn money by predicting the endgame, while FAs earn money by predicting consensus; they don't need to prove a trend is ultimately correct. Conversely, the more vague and grand the narrative, the more room FAs have to operate. For example, embodied intelligence, world models... these fields share a common characteristic: the future narrative is grand enough, the technology is complex enough, and it's difficult to disprove in the short term. Large companies are burning money to train models, investors are betting on AGI ten years from now, and middlemen are making money.
Either exploit the capital expenditures of large companies
Either exploit the FOMO expenditures of limited partners
7. Money is everything
The AI race is still a long way off. People often ask, what is the scarcest resource in the AI era?
Many people would answer: talent, computing power, data... But in the end, they all have only one name: money.
The biggest competitive advantage in the AI industry isn't technology, but its ability to raise funds. Entrepreneurs now need to keep raising funds, even if they already have enough in their accounts. The real battle will take place during future industry downturns. As long as they have money to survive, they can wait until their competitors run out of money and then acquire their talent, technology, and customers at low prices. The secondary market is similar; a major bubble collapse is highly likely within the foreseeable 2-3 years. Those who have enough money to buy at the bottom will outperform 99% of the market. [The following text appears to be an unrelated advertisement:] The Computing Power Bureau is dedicated to connecting B300 computing power with token industry opportunities and exploring large-scale overseas expansion models. Industry professionals are welcome to exchange ideas (V: blocktheworld).