Don't be "fooled" by the 600 million yuan in profit. The most profitable hardware is precisely Unitree's biggest risk.

In the previous article, we wrote about "Unitree: The 'Strangest' IPO Company in History."

This time, we'll take a different perspective and reexamine Unitree from the perspective of the humanoid robot industry.

Unitree's profitability not only exceeds many people's expectations but also far surpasses that of most embodied intelligence companies.

In most industries, this would be sufficient to prove a company's competitiveness: it can make money, has high profits, and a strong brand, seemingly indicating that commercialization has been successful.

However, in the current embodied intelligence industry, this situation may need to be viewed from the opposite perspective.

The money Unitree is making today is more like a temporary pricing power rather than the realization of long - term industry competitiveness.

It has seized a very special window period:

During the window period when demand is not yet mature, technological paths have not converged, and the competitive landscape has not fully emerged, it has amplified efficiency through engineering and supply - chain advantages, and combined with the cognitive dividends brought by dissemination, temporarily achieving a seemingly "excessive" result.

But this is exactly where the problem lies.

First, the hardware form is ultimately defined by the scenario, not the other way around.

When robots enter specific scenarios as a form of productivity, the optimal solution may not be the current hardware form.

In other words, what Unitree is strongest at now may not be the most valuable in the future.

Second, Unitree's engineering - driven path has an obvious ceiling.

The development of AI over the past 70 years has repeatedly proven that all engineering miracles will ultimately be flattened by the brute - force aesthetics of "computing power + data." This will not be an exception in the field of embodied intelligence.

This is why Unitree is making money through engineering on one hand and allocating 2 billion out of the 4 billion raised in its IPO to improve its "brain" on the other.

Behind this is a very clear judgment: short - term advantages come from engineering, while long - term success still depends on the model.

01 Behind the High Gross Margin: A Temporary Victory Driven by Engineering

If we break down the embodied intelligence track, we'll find that it is still in a very early stage.

The most intuitive indicator is the demand structure.

Today, 74% of humanoid robots are still used in scientific research and education scenarios. The much - talked - about "working on the factory line" has hardly happened in reality. Instead, scenarios such as corporate tours and exhibition hall explanations account for 50% to 70%.

This means that the current mainstream demand in the industry is essentially not a productivity demand but a display demand.

The characteristic of display demand is that as long as the robot can move, walk, and operate stably for a period of time, it is sufficient. However, productivity demand is different. It requires long - term stability, replicability, the ability to replace humans, and a clear ROI calculation.

And embodied intelligence has obviously not reached the latter stage yet.

The breakthrough of large models is essentially a "data emergence" — for the first time, Internet data has been used on a large scale.

But embodied intelligence does not have this condition yet. Although methods such as remote operation, motion capture, and simulation are being used, the industry has not yet trained a general robot model that is truly viable in terms of parameter scale and performance.

So, many industry insiders have a similar judgment: Today's embodied intelligence is more like autonomous driving more than a decade ago — people can see the direction, but it may still be a long way before large - scale implementation.

How did Unitree stand out in such a stage?

Its strategy is very clear: instead of focusing on the "brain," it maximizes the more certain "cerebellum." It uses engineering capabilities and supply - chain advantages to quickly gain market share in the early market.

The most crucial step here is the quadruped robot dog.

By the first three quarters of 2025, Unitree had sold 17,946 quadruped robots.

This has brought not only revenue but also two important capabilities: one is to firmly control the supply chain, and the other is to make key components into highly universal modules. Later, these two points were directly applied to humanoid robots.

Ultimately, this is reflected in the difference in gross margin. The industry average is generally between 35% and 47%, while Unitree can achieve 63%. This nearly 20 - percentage - point difference is not essentially the advantage of a single product but the result of a complete manufacturing and supply system.

Combined with its appearance on the Spring Festival Gala, Unitree achieved almost cost - free brand expansion.

Thus, an interesting phenomenon emerged: a well - known robot company spent only 22.57 million on advertising in the first three quarters of 2025.

Looking back, Unitree's high gross margin is not difficult to understand. It is not a long - term proven profitability but a temporary pricing power formed in the early window period: When demand is not yet mature, technological paths have not converged, and the competitive landscape has not fully emerged, it amplifies efficiency through engineering and the supply chain, and combined with the dissemination dividend, it achieves an "excessive result."

But this is exactly where Unitree's problem lies.

All of Unitree's advantages are based on the premise that the industry is still in the "display demand" stage. Once the demand shifts to "productivity," will these advantages still hold?

02 The Ontology - Driven Approach Is Not the Most Favored by Capital

Although Unitree is the most successful company in the commercialization of embodied intelligence, its technological path is not considered "mainstream."

If we classify today's embodied intelligence companies according to their technological investment paths, they can be roughly divided into three categories: ontology - driven, brain - driven, and full - stack.

Unitree is a typical ontology - driven company.

Its core lies in the ontology and motion control, which is what the industry often refers to as locomotion control. In simple terms, it enables robots to walk stably, maintain balance, and even perform complex actions such as rolling, jumping, and dancing in the real world.

It also conducts research on the "brain," but the priority is clear: motion control first. Companies with similar paths include Zhuji Dynamics and Zhongqing.

On the other hand, there are brain - driven companies, represented by companies such as Galaxy Universal, Independent Variable, Xinghai Map, Qianxun Intelligence, and Zhipingfang. These companies have a very consistent consensus: prioritize model development.

This choice is directly reflected in the allocation of R & D resources. For example, Xinghai Map allocates 80% of its R & D investment to the "brain," with 30% for data, 40% to 50% for computing power, and only 20% for hardware. Galaxy Universal has even only developed a wheeled chassis, and almost all other resources are invested in model R & D.

Meanwhile, these companies generally choose wheeled robots instead of bipedal ones.

The reason is not complicated. In the stage when demand is not clear, the most core ability of a robot is not "how human - like it walks" but "whether it can complete tasks." Compared with mobility, the upper - body operation and decision - making abilities are more crucial.

The third category is full - stack companies, which focus on both the "brain" and the ontology. Typical representatives include Zhiyuan Robotics, as well as Tesla and Figure overseas.

From the preference of the primary market, betting on the "brain" has almost become an industry consensus. There are two very intuitive side - confirmations of this.

First, it's Unitree's own valuation.

Unitree's post - investment valuation in the latest round is about 12.7 billion. Considering its demonstrated profitability and the popularity of the track it is in, this valuation is not very high.

Even earlier, before its appearance on the Spring Festival Gala, its valuation had been stagnant for a long time and was once only half of Zhiyuan's.

Second, it's the common choice of unicorns.

There are already 8 domestic embodied intelligence companies with valuations over 10 billion, including Galaxy Universal, Zhiyuan Robotics, Unitree, Xinghai Map, Zhipingfang, Qianxun Intelligence, Independent Variable, and Xingdong Jiyuan. The vast majority of them clearly emphasize a model - first strategy.

Why are investors more willing to bet on the "brain"?

Because the biggest bottleneck in current embodied intelligence lies not in hardware but in software.

The ultimate goal of humanoid robots is not to be automated equipment in a specific scenario but to have universality and be able to complete different tasks in different environments like humans.

Without a basic model, robots cannot understand the physical laws of the real world and thus cannot handle truly complex tasks.

Take the simplest example: when a robot picks up an egg, the hardware is responsible for executing the action, but it is the model that determines how much force to use without breaking the egg.

This is why the real focus of companies such as Google, NVIDIA, Physical Intelligence, and Galaxy Universal is the world model — enabling AI to understand the physical world.

More importantly, the evolution logic of software and hardware is completely different, which directly determines their value differences in the capital market.

Although hardware has bottlenecks, they are mostly linear. It can become cheaper and more mature through engineering accumulation and continuous iteration. For example, in the past, industrial - grade dexterous hands might cost hundreds of thousands of yuan, but now Tesla has reduced the cost of a single hand to about $6,000.

However, the bottlenecks of software are non - linear. No one knows when a breakthrough similar to ChatGPT will occur. But once it does, the improvement in the generalization ability of robots is likely to unlock a large number of scenarios.

So, the problem is clear. Before the intelligent level has a significant leap, simply expanding application scenarios will make it difficult to achieve a real ROI and even more difficult to form a stable business closed - loop.

03 Hardware Is Determined by the Scenario: This Is Unitree's Real Risk

At this point, someone might ask: since the model has not yet emerged, why not focus on the ontology first?

Of course, it's possible. And in the long run, embodied intelligence will definitely be vertically integrated — doing both model development and hardware manufacturing.

However, the problem is that focusing on the ontology first is essentially locking in an answer in a stage where variables are not yet determined.

There are two key constraints here.

First, the hardware form is ultimately defined by the scenario, not the other way around.

In the display demand stage, the humanoid form, running, jumping, and dynamic abilities are all meaningful because they are more intuitive and visible.

But once it enters the productivity stage, the evaluation criteria will completely change. The robot is no longer judged by "what it can display" but by "whether it can enter the process, whether it can replace humans, and whether the ROI can be calculated."

At this time, the optimal solution may not be the current hardware form.

Enterprises may not necessarily need a humanoid robot that can run and jump. They are more likely to choose a wheeled solution with lower cost and higher stability. Compared with complex motion abilities, the more frequently used abilities may be grasping, assembly, and sorting, which may not seem as cool.

In other words, once the scenario converges, the hardware will be redefined.

Second, the progress of the model is likely to rewrite some of the abilities that currently rely on engineering optimization.

This can actually be understood through "The Bitter Lesson." The development of AI over the past 70 years has repeatedly proven that many paths relying on manual design and engineering skills will ultimately be replaced by more general methods that rely on computing power and data.

For example, AI engineering methods have evolved three times in the past two years: Prompt engineering, Context engineering, and Harness engineering.

The same applies to embodied intelligence.

Many of the seemingly strong control abilities today are still essentially the result of engineering optimization. But once the model's ability improves, these abilities may be re - covered by end - to - end learning.

In other words, some of today's advantages may only be solutions for the transitional stage.

Considering these two points together, we can see Unitree's structural risks.

Most of its current shipments still go to scientific research and display scenarios, essentially verifying "whether the technology can work." However, the real business world only cares about two things: whether the demand has converged and whether the ROI can be achieved.

As of now, neither of these two things has happened.

This means that Unitree's current advantages are more like "temporary" rather than "long - term."

Once there is a breakthrough in model ability, the scenario begins to converge, and the hardware is redefined, the current path may not continue smoothly.

From proving that robots can run to proving that they can really make money, Unitree's real test has just begun.

This article is from the WeChat official account "Silicon - based Observation Pro," written by Silicon - based Jun and published by 36Kr with authorization.