Barclays elaborates on the "three major controversies" of humanoid robots: When will large-scale deployment be carried out? Is the demand for computing power experiencing a sharp surge? Is the "humanoid form" a necessary design?

Barclays' analysis states that the bottleneck lies in intelligence rather than hardware, the computing power pull is concentrated at the edge rather than in data centers, and the first wave of disruption brought by Physical AI will most likely not be humanoid at all.

Barclays believes that the large-scale deployment of general-purpose humanoid robots may not happen until around 2035, instead of 2030.

In its latest research report released on September 18, Barclays put forward three judgments on the humanoid robot investment boom: the bottleneck lies in intelligence rather than hardware, the computing power pull is concentrated at the edge rather than in data centers, and the first wave of Physical AI disruption will most likely not be humanoid at all.

In the report, Barclays also conducted a detailed analysis around the issues that investors are most concerned about, such as scale, computing power, and form.

01

Controversy 1: Large-scale deployment in 2035, not 2030

Barclays wrote in the report that many of the humanoid robot capabilities demonstrated in the current industry still rely on pre-programming, remote control or narrow-domain automation for specific scenarios, with a significant gap from real general-purpose autonomy.

The institution believes that the core bottleneck is not in hardware, but in intelligence — the capabilities of AI models required for perception, reasoning and action are not yet mature. On the hardware side, there is a "chicken-and-egg" dilemma between scale, cost and capability: no scale means no low cost, and insufficient intelligence makes it impossible to prove commercial value, making it difficult to achieve breakthroughs in all three at the same time.

This pattern is highly similar to the more than ten-year commercialization cycle that autonomous driving has gone through.

As a result, Barclays believes that earlier investment opportunities lie in the computing power, data and model layers — that is, the underlying infrastructure needed to unlock the "GPT moment" for humanoid robots. The large-scale economical deployment of hardware will have to wait for the first breakthrough in the intelligence layer.

02

Controversy 2: Computing power demand — limited increment in data centers, the edge is the main battlefield

The pull of humanoid robots on computing power needs to be viewed from two dimensions: centralized data center computing power and distributed edge AI computing power.

On the data center side, computing power mainly serves two types of work: simulation and synthetic data generation (for training and testing strategies), as well as the training and post-training of foundation models (to build intelligence from perception to action). However, reasoning — the link where robots perceive, make decisions and act in real time in the real world — must be completed on dedicated edge processors inside the robot, due to strict constraints on latency, power consumption, reliability and security. This means that the reasoning load of humanoid robots will not be directly converted into incremental demand for ultra-large-scale data centers like Agentic AI. For data centers, humanoid robots represent a small increment; for edge computing, they may be a major driving force.

However, in an industry where the production capacity of AI data centers is already tight, even the increment carries weight. More critically: computing power demand expands ahead of hardware deployment. Before the large-scale implementation of robots, developers need massive computing power for simulation and model training.

The cooperation between Figure and new cloud service provider Nscale is a typical example. This multi-year agreement has an initial investment of about 3.5 billion US dollars, with a target scale exceeding 6 billion US dollars, and can support up to 100,000 NVIDIA Vera Rubin GPUs. Estimated based on the full-configured rack capacity of about 3.0-3.3 kilowatts per GPU, it corresponds to an IT capacity of about 300-330 megawatts — at the stage when humanoid robots have not been launched on a large scale, only developing the "brain" has already constituted a considerable incremental demand for data centers.

03

Controversy 3: "Humanoid" is not necessarily the optimal solution for Physical AI

Barclays believes that the first wave of Physical AI disruption will most likely not be humanoid.

What Barclays calls "too early" specifically refers to fully autonomous general-purpose humanoid robots. Before that, AI robots for specific tasks have already been put into operation in multiple industries. Collaborative robots (cobots), autonomous mobile robots (AMRs), AI drones, quadruped robots — these robots of various forms are cutting into "dirty, dull, dangerous" job types, covering many scenarios in commerce, industry, national defense and other fields.

At present, Amazon has deployed more than 1 million robots in its operations, covering mobile drive units, AMRs and AI control systems, which are divided by tasks and forms, none of which are humanoid. Atoms, a company under Travis Kalanick, the founder of Uber, recently completed a $1.7 billion financing, focusing on dedicated industrial robots and Physical AI systems in the mining and transportation sectors.

Then why pursue humanoid forms? Barclays' analysis points to two main logics:

First, the human world is built around the human body. Stairs, doors, tools, workbenches, and production lines are all based on human dimensions. Humanoid robots can directly enter the existing environment without the need to renovate the infrastructure for robots.

Second, general-purpose humanoid robots offer the possibility of a multi-task platform — the same robot can transport materials, operate tools, and inspect equipment. Its value lies in one platform covering multiple scenarios, rather than outperforming dedicated robots in any single task.

However, when tasks and environments are clear, dedicated solutions are often faster, cheaper and safer. Wheeled robots are more practical than bipedal robots in flat warehouses, quadruped robots are more stable than bipedal robots on uneven terrain, and dedicated grippers are more accurate than five-fingered hands in repetitive operations. The choice of form is ultimately determined by specific scenarios.

This article is from the WeChat Official Account "Hard AI", the author is Hard AI focusing on technology production and research, and 36Kr is authorized to release it.