The next Micron could be hidden in the middle layer of physical AI.

How much is the market consensus on AI actually worth?

Just take a look at Micron Technology to find the answer.

Over the past year and a half, Micron's share price has surged from $61.419 to $1132.33, with a cumulative increase of 1743.62%, and the maximum drawdown during this period was only about 30.31%.

Apple took a full 10 years to achieve the same level of growth.

The story of Micron essentially reflects a new feature of asset revaluation in the AI era: the speed at which industrial value is realized is accelerating dramatically.

This also presents new challenges for investors. By the time the AI industry trend is confirmed by the market, the biggest opportunities have often already passed. The phase of truly high returns typically occurs before the market consensus is fully formed. As a result, the time window for AI investment is continuously shifting earlier.

What investors need to find is the next direction that has not yet been fully priced in.

We are at such a critical juncture right now. On one hand, the speed of model iteration is getting faster and faster. On the other hand, the AI Coding narrative has come to a temporary pause, and the market is looking for new growth drivers.

Where will the next AI consensus emerge? Who will be the next Micron? Today, we will dive into this topic.

The revaluation of Physical AI has begun

Over the past two years, every major revaluation of AI assets has stemmed from the formation of a new industrial consensus.

The first phase was marked by breakthroughs in technical capabilities.

ChatGPT opened the door to generative AI, making the market believe for the first time that large language models could become the next-generation computing platform.

Two years later, DeepSeek-R1 delivered another breakthrough in a different sense: it proved that high-performance models do not necessarily require infinitely stacked computing power, and a low-cost, high-efficiency model development path is equally viable.

This breakthrough means that China's AI industry chain has regained global competitiveness, and has directly changed the market's valuation of Chinese AI assets.

On January 20, 2025, when DeepSeek-R1 was officially released, the Hang Seng Tech Index stood at around 4595 points. In the following month, the index rose by about 30%, entering a technical bull market. By October 2025, the Hang Seng Tech Index peaked at 6715 points, up 46% from the time of DeepSeek-R1's release. Among the top beneficiaries was Alibaba, whose share price rose from HK$81 to HK$185, representing an increase of over 100%.

The second revaluation came from Anthropic's commercialization breakthrough.

In April 2026, Anthropic announced that its annualized revenue exceeded $30 billion, more than tripling from the roughly $9 billion at the end of 2025. By the end of May, this figure further surpassed $47 billion.

The Coding scenario was the first time AI truly found a productivity entry point for large-scale commercial implementation, which further drove the revaluation of AI asset values.

Following Anthropic's explosive revenue growth, May saw the most intense surge in AI hardware assets. The SOXX semiconductor ETF rose by about 23% in a single month. By the end of May, SOXX had posted a cumulative increase of around 89% for the year.

Looking back at the two previous rounds of AI asset revaluation, a distinct characteristic is that the speed of industrial value realization has clearly accelerated.

Micron is a typical case. Last year, Micron's share price started at $61.419, and in just a year and a half, it soared to $1132.330, with a cumulative increase of 1743.62%, while the maximum drawdown during this period was only 30.31%. Apple took a full 10 years to achieve the same level of growth.

This means that the time window for AI investment is continuously shifting earlier, and the core is to find the next direction where a consensus has not yet formed. From the current perspective, the next round of AI revaluation may come from two directions.

The first direction is the continuous expansion of AI commercialization into more professional service scenarios.

Coding is just the first entry point for AI to replace human labor. Last year, HSG conducted an internal calculation to estimate the scale of various scenarios in service fields including nurses, lawyers, and programmers. The calculation logic was: "number of practitioners in the occupation × median annual salary" (data from the U.S. Census Bureau), and the final result showed that the AI service market will exceed $10 trillion.

Among this total, AI can create 20% of the value of programmers, corresponding to an increment of about $300 billion in the IT service market. If it further covers 20% of white-collar jobs, the corresponding market space will reach $3.5 trillion.

This opportunity is mainly concentrated in the primary market. Based on this judgment, HSG has invested in many AI service companies, including Open Evidence and Freed in the healthcare sector, and Harvey, Crosby, and Finch in the legal sector.

The second direction is AI entering the physical world. In the past, Silicon Valley generally followed the AI-first principle, but recently, more and more companies and researchers have realized that the implementation of AI must rely on hardware.

From the current perspective, the market has already started placing bets in advance. As of June 22, 2026, global financing related to embodied intelligence has reached $18.8 billion, surpassing the total amount for the whole of 2025 in less than half a year.

The Chinese market is equally hot. According to Crunchbase data, by mid-May, Chinese robotics companies had completed 176 financing deals this year, totaling $5.6 billion, exceeding the full-year figure for 2025.

In the secondary market, although companies focused on end-user applications for robotics and autonomous driving have not seen a significant price increase, the "shovel sellers" in the Physical AI industry chain have begun to attract capital attention.

Since the beginning of this year, industrial machine vision company Cognex has risen 76.5%, Teradyne (which has a collaborative robotics business) has risen 72.4%, and lidar company Ouster has risen 60.7%.

This means that although a full consensus has not yet been formed, the revaluation of the Physical AI industry chain has already started.

The underappreciated middle layer

Over the past two years, the boom in large language models has proven one thing: in the early stage of an industry explosion, the most profitable players are often not the end application developers, but the "shovel sellers" who control core resources.

Currently, the combined market value of global large language model companies is around $2.5 trillion, while NVIDIA's market value is as high as $4.96 trillion. The total market value of the world's five major storage companies also stands at $3.45 trillion, not even including ChangXin Memory Technologies which is about to go public.

It is almost certain that Physical AI will follow a similar industrial pattern.

At the WAIC (World Artificial Intelligence Conference), when the Zhiyuan G2 announced mass production at the 10,000-unit scale, Unitree demonstrated its industrial assembly scenarios, and Galaxy Universal launched a wheeled dual-arm solution, all eyes were focused on embodied intelligence companies. Few people noticed that the Physical AI industry chain is also undergoing stratification.

According to the latest report released by IDC, the Physical AI industry chain can be roughly divided into three layers:

The top layer consists of applications and business scenarios, simply put, embodied intelligence and autonomous driving. The middle layer is the software infrastructure layer, including models, data, and simulation training platforms. The bottom layer is the infrastructure layer, which provides computing, connectivity, energy supply, perception, and physical execution capabilities.

Among these, IDC believes that the middle software infrastructure is the key to the large-scale implementation of the entire Physical AI industry.

The real difficulty of Physical AI lies in the fact that physical testing fields are too expensive, and most enterprises do not have the ability to build high-precision simulation environments. This makes it difficult for AI to operate stably in complex real-world environments. The software layer can convert operational feedback into reusable capabilities, improving the evolutionary speed and large-scale replication capacity of Physical AI.

Specifically, software solves this problem through three dimensions:

First, models and strategies. Solving how machines understand their environment, perform reasoning, and make decisions;

Second, simulation and verification. Using simulation platforms to build controllable, reproducible, and traceable virtual environments to enable large-scale testing of machines before they enter the real world;

Third, scenario data and synthetic data. Using synthetic data to supplement extreme weather conditions, dangerous interactions, long-tail events, and complex scenarios that are difficult to collect, solving the problem of covering long-tail situations in the real world.

More importantly, these three elements do not exist in isolation. Instead, they need to form a continuous iterative closed loop around real scenarios: after scenario data is collected and processed, it is used for model training, and the model then undergoes continuous iteration through simulation verification, synthetic data expansion, and real deployment feedback.

This also explains why NVIDIA has invested heavily in developing Omniverse and Cosmos, and why Tesla insists on building its own simulation system. Whoever can successfully run the closed loop of "real data - simulation verification - synthetic expansion" will control the essential infrastructure (like water, electricity, and gas) in the Physical AI era.

However, on a global scale, the software infrastructure for Physical AI is still in its early stages.

Models, simulation, and data originally belonged to different industrial segments: model companies were responsible for training the "brain", simulation vendors provided testing environments, and data service providers handled data collection and annotation. As Physical AI begins to move into real-world scenarios, these originally independent capabilities are gradually being connected.

The business evolution of 51World can be seen as a sample to observe this trend.

It initially started with autonomous driving simulation, but with the development of end-to-end models and Physical AI, simply providing a virtual testing environment is no longer sufficient. Simulation platforms need to integrate real data, continuously create new scenarios through world models and generative AI, and then precipitate the verification results back into training data.

Following this path, 51World has gradually extended its business to world models, simulation training, and synthetic data, attempting to connect the previously scattered segments into a unified Physical AI software infrastructure.

At present, 51World's strategic layout has begun to take shape.

In the model layer, 51World uses world models to model spatial relationships, motion trajectories, and physical laws in real environments, generating more complex scenarios for autonomous driving and embodied intelligence models, especially for low-frequency, dangerous, and hard-to-collect long-tail scenarios in the real world.

In the simulation and verification segment, 51World's SimOne simulation training platform plays a key role in connecting the virtual world with real roads.

The SimOne platform can restore real roads into interactive, editable digital scenarios, and on this basis, modify variables such as weather, traffic participants, and obstacles to repeatedly test the models.

In the data segment, the DataOne synthetic data engine is responsible for converting real data and simulation processes into trainable scenario assets, filling gaps in training data through automatic annotation, scenario generalization, and Corner Case generation.

Once the data closed loop is fully operational, this software capability built around the real world will help 51World build an extremely deep moat in the Physical AI industry chain. Such changes have already begun to take place first in the autonomous driving sector.

The simulation software business is transforming

Among all Physical AI directions, autonomous driving may be one of the fields closest to large-scale commercial implementation.

The reason is that it meets three conditions simultaneously: clear real-world demand, a R&D process that is highly dependent on data, and strong rigid demand for verification processes.

For L3/L4 level autonomous driving, the real difficulty is not to make the vehicle complete a single driving task, but to maintain safety in countless complex, extreme, and unpredictable scenarios.

This means that the development of autonomous driving is increasingly dependent on a complete software infrastructure. This year, the new access regulations of the Ministry of Industry and Information Technology have established virtual simulation as a statutory pre-link, which further validates the rigid demand for this infrastructure from a policy perspective.

To some extent, autonomous driving is the most practical test field for 51World's software closed-loop system.

From a market performance perspective, this development path has been partially validated. According to data from Frost & Sullivan, 51Sim ranks first in China's end-to-end advanced intelligent driving simulation and data platform market with a 53.5% market share.

Benefiting from its high market share, the market has begun to re-evaluate 51World's growth potential.

According to a research report from Guotai Junan Securities, the company's revenue is expected to reach 725 million yuan, 1.325 billion yuan, and 1.985 billion yuan from 2026 to 2028 respectively, corresponding to growth rates of 108.3%, 82.8%, and 48.76%.

Of course, high market share brings not only revenue, but also changes to the business model.

In the AI era, the commercial value of simulation platforms is shifting from "selling tools" to "helping customers build long-term reusable scenario data assets and a continuously iterating simulation verification system".

In the past, simulation software was essentially project-based tool delivery: after delivering one scenario, the project was completed. But in the AI era, service has become a continuous process, relying on the enrichment of scenario data assets and the continuous iteration of the simulation verification system.

A representative case is that this year, 51Sim will collaborate with industry partners to launch the "Physical AI Factory" model, integrating computing power, software, and data. Customers will pay based on their actual usage of computing resources and data.

This means that the underlying logic of the simulation platform's business model has changed: shifting from one-time project revenue to the operation of accumulable, reusable, and subscribable scenario data assets. The richer the scenario library and the more comprehensive the long-tail coverage, the higher the customer migration cost, and the stronger the platform's network effect.

There is no doubt that the network effect of the platform will also greatly increase the value of software platforms in the Physical AI era.

Conclusion

Returning to the initial question: How much is the market consensus on AI actually worth?

Micron's answer is 17 times growth.

AI has raised the rewards of this era to an unprecedented level, but it has also greatly increased the difficulty of identifying the right direction.

Over the past two years, people have been accustomed to looking for the miracles of artificial intelligence on screens — from a single dialog box, a piece of code, to the explosive demand for computing power. But technology will eventually step out of the silicon-based greenhouse to connect with the complex, random, and unpredictable realities in the physical world.

There is still no clear answer as to who will be the next Micron. But 51World at least provides us with a viable option.

Before the consensus is formed, it may seem like just a simulation or data business. Perhaps after the consensus takes shape, the market will realize that it is the passport for Physical AI to enter the real world.

This article is from the WeChat public account "Silicon-based Observer Pro", author: Silicon-based Jun, published with authorization from 36Kr.