WAIC Observation: Crowded Consensus, Huge Bubble

Author: Xiao Bing, Deep Tide TechFlow

1. WAIC Should Be Held Earlier, So That Tech Can Be Cleared Out Soon After Visiting.

There are too many peak signals: various side events becoming crypto-centric, a gathering of beautiful women, and highly homogeneous products on site... A track has transformed from a geek's game into a public feast.

Whether in the primary or secondary market, when everyone starts believing the same story and consensus becomes too concentrated, risks begin to accumulate.

Because in a bull market, the bulls standing beside you will all become sellers in a downturn.

More kills, more!

2. Large Models Have No Real Moat

The foundation of AGI still relies on Scaling Law.

The relationship between model capability, parameter scale, training computing power, and data quality still follows a power law.

Who can acquire GPU computing power more cheaply, obtain higher quality data, recruit better researchers, and burn money more efficiently?

Thus, the moat of large model companies is weaker than many imagine.

Model capabilities will continue to converge, leading advantages will be chased, and prices will keep falling.

The ones truly making sustained profits are those selling water to all model companies, as everyone is tapping into the capital expenditures of a few big firms.

GPU, HBM, high-speed interconnect, data centers, electricity, data services...

Standing at the upstream of the arms race to share the spoils, or even just acting as a middleman connecting resources, is currently the most profitable business.

3. Data Bottleneck is a Potential Opportunity Seen at This Conference

After discussing with some friends, I found that the companies making money in this wave of AI are very low-key and do not even come to exhibit, one direction being data.

The success or failure of unified multimodal representation essentially depends on high-quality multimodal training data. There is a severe shortage of data assets that have been thoroughly cleaned, accurately labeled, cross-modal aligned, and can be iterated continuously.

Models will become cheaper, GPUs will ultimately increase, but truly high-quality data will become increasingly scarce.

The multimodal data industry chain is becoming a lucrative track for profit.

Data cleaning, data labeling, data synthesis, vertical data assets, robotic data collection... as well as algorithm companies specifically solving cross-modal unified representation and multimodal pre-training encoders.

GPUs are one-time capital expenditures, while data is a continuous capital expenditure.

Currently, I categorize data into three types: the first type is Expert Data; the second type is RL Environment / Agent Data. In the future, training agents will not just involve collecting question-answer pairs but constructing Environment → Task → Trajectory → Reward → Verifier, which could be one of the largest incremental markets ahead.

The third type is Embodied / Robotics Data, which is even scarcer than LLM data and extremely difficult to collect.

4. AI Application Dilemma, Once a Reflection of the Crypto World

Today's AI is very similar to past Crypto, with fat protocols and L1 capturing the vast majority of value, so Crypto VCs are frantically pouring money into public chains, applications? No one is investing in them.

Currently, AI is the same.

Large models are like a POW Layer1, and even the model itself is the application.

In the past few years, most AI applications have been doing something dangerous: packaging capabilities that the model temporarily lacks into products.

But the boundaries of model capabilities have been expanding outward. Every model upgrade is like L1 directly writing functions that originally belonged to the application layer into the protocol, such as search, deep research, coding, image generation, video generation, computer use, agent...

If every time the model upgrades, your product loses a layer of value, then you are essentially a feature of AI, not an AI application.

Therefore, it is essential to create things that the model can do but are also difficult to take away, which boils down to data, context, workflow, permission, and distribution.

The moat of AI applications is to become a customer of the model rather than a competitor to the model.

5. Embodied Intelligence, a Huge Bubble

In Hall H4 on the second floor, watching a large number of robots produced by the same supply chain performing similar actions slowly, and then looking at the valuations of various companies, it makes one’s scalp tingle and feel sorry for the investors' money.

The current problem with embodied intelligence is that capital is pricing "software Scaling Law" for "soft and hard complex systems."

The miracle of large models is that if a company trains GPT-5, it can theoretically serve hundreds of millions of people globally at nearly zero marginal cost, but robots cannot.

For every additional user a robot serves, a new machine must be built, involving BOM, manufacturing, supply chain, delivery, maintenance, and depreciation.

AI intelligence can progress exponentially; for instance, when a model upgrades, all global users become smarter simultaneously, but the costs in the physical world do not decrease exponentially.

If intelligence follows Moore's Law, then embodiment still adheres to manufacturing industry rules. Ultimately, embodied intelligence may become a huge industry but may not possess the profit margins of large models.

Moreover, it is still too early.

Autonomous driving is already a highly constrained embodied intelligence problem with clear objectives, limited action space, standardized road rules, and a wealth of real data and a mature automotive industrial system. Even so, this industry has burned through hundreds of billions of dollars and still has not completely solved some long-tail problems on the road. The real difficulty faced by embodiment is N times that of autonomous driving.

6. The Best Era for Pimps

A harsh truth is that most AI startups are not making much money, and their profits are even less than those of some women selling their skills.

So, I discovered something interesting: many companies appear to be one business, but upon closer inspection, "Oh, you are also selling tokens," "Oh, you are also selling computing power," and it turns out that large model companies are also reselling computing power for profit.

The ones truly making money are still the pimps.

FA, matching old stocks, flipping B300, selling tokens, selling datasets, selling computing power, or even selling an opportunity to meet a certain founder...

AI meets all the conditions for a thriving intermediary market: rapid technological changes, significant information gaps, ample capital, and compelling narratives.

VCs earn money by judging the endgame, while FAs earn money by judging consensus, without needing to prove that a trend is ultimately correct.

On the contrary, the more vague and grand the narrative, the more space there is for FAs to operate.

For example, embodied intelligence, world models... these tracks share a common characteristic: the future narrative is sufficiently grand, the technology is complex enough, and it is sufficiently difficult to falsify in the short term.

Big companies are burning money to train models, investors are betting on AGI ten years down the line, and intermediaries are making money.

Either tap into the capital expenditures of big companies

Or tap into the FOMO expenditures of LPs.

7. Money is Everything

The AI race still has a long way to go. People often ask, what is the most scarce resource in the AI era?

Many would answer: talent, computing power, data... but ultimately, they all have one name: money.

The biggest competitive advantage in the AI industry is not technology, but the ability to raise funds.

What entrepreneurs need to do now is to keep raising funds, even if there is enough money in their accounts to spend, they still need to raise.

The real war will be in the future industry downturn cycle.

As long as there is money to keep living, one can wait until competitors run out of money and then acquire their talent, technology, and customers at a low price.

The secondary market is the same; in the foreseeable 2-3 years, a massive bubble collapse may occur, and as long as there is enough money to buy the dip at that time, one can defeat 99% of people.