Barclays believes that mass deployment of general-purpose humanoid robots is more likely to arrive around 2035, rather than the much-anticipated 2030 timeline.

In a research report released on September 18, Barclays offered three key judgments on the current humanoid robot investment frenzy: the primary bottleneck lies in intelligence, not hardware; the computing power demand will be centred on edge computing rather than data centers; and the first wave of Physical AI disruption might not be humanoid in form at all. The report also delivers a detailed analysis of investor concerns regarding scale, compute demand, and robot design.

The First Debate: Mass Deployment in 2035, Not 2030

Barclays noted in its report that many demonstrated capabilities of current humanoid robots still rely on pre-programmed routines, remote teleoperation, or narrow automation tailored for specific scenarios, and they remain far from achieving true general autonomy. The core bottleneck, according to Barclays, is not hardware but intelligence; the AI models required for perception, reasoning, and action are still immature. On the hardware side, the industry faces a chicken-and-egg problem between scale, cost, and capability; without scale, costs cannot drop, and without sufficient intelligence, commercial value cannot be proven, making it difficult to break through all three simultaneously.

This dynamic closely mirrors the commercialisation cycle that autonomous driving has gone through over the past decade or so. Consequently, Barclays suggests that earlier-stage investment opportunities lie in computing power, data, and the model layer, the foundational infrastructure needed to unlock the GPT moment for humanoid robots. Large-scale economic deployment of hardware, meanwhile, will have to wait for the intelligence layer to break through first.

The Second Debate: Computing Power Demand is Modest for Data Centers, But Critical for Edge

Humanoid robots' demand for computing power needs to be assessed from two distinct dimensions: centralised data center compute and distributed edge AI compute. On the data center side, compute is primarily used for two things: simulation and synthetic data generation for training and testing strategies, and training and post-training of foundation models to build intelligence from perception to action. However, inference, the component where a robot perceives, decides, and acts in real-time in the real world, must run on dedicated edge processors within the robot itself due to strict constraints on latency, power consumption, reliability, and safety. This means that the inference load for humanoid robots will not translate directly into surging data center demand the way agentic AI does. For data centers, humanoid robots represent a small incremental addition; for edge computing, they could be a major driver.

Still, in an industry where AI data center capacity is already tight, even small increments carry significant weight. More importantly, the demand for computing power will expand ahead of hardware deployment. Developers require massive computing resources for simulation and model training long before robots are widely deployed. A telling example is the partnership between Figure and the emerging cloud service provider Nscale. The multi-year agreement carries an initial investment of about $3.5 billion, with a target scale exceeding $6 billion, and could support up to 100,000 NVIDIA Vera Rubin GPUs. Estimated at roughly 300 to 330 megawatts of IT capacity based on fully configured racks of approximately 3.0 to 3.3 kilowatts per GPU, this represents a substantial incremental demand on data centers during the development of the brain alone, even before humanoid robots have scaled.

The Third Debate: The Human Form May Not Be the Optimal Solution for Physical AI

Barclays argues that the first wave of Physical AI disruption will likely not be humanoid in form. The bank also says that its "too early" stance specifically refers to fully autonomous, general-purpose humanoid robots. Before that milestone, task-specific AI robots are already moving across a wide range of industries. Collaborative robots, autonomous mobile robots, AI-powered drones, and quadruped robots are entering the workforce through dirty, dull, and dangerous tasks, covering sectors such as commerce, industry, and defence.

Currently, Amazon has deployed more than one million robots in its operations, spanning mobile drive units, AMRs, and AI manipulation systems, all allocated by task and form, with not a single humanoid among them. Atoms, under the leadership of Uber founder Travis Kalanick, recently completed a $1.7 billion funding round focused on specialized industrial robots and Physical AI systems for mining and transportation.

So why pursue the human form at all? Barclays points to two primary reasons. First, the human world is built around the human body. Stairs, doors, tools, workbenches, and production lines are all designed to human scale, and a humanoid robot can operate directly within existing environments without the need to retrofit infrastructure for machines. Second, general-purpose humanoid robots offer the possibility of a multi-task platform, where a single robot can move materials, operate tools, and inspect equipment. The value lies in a single platform spanning multiple scenarios, rather than outperforming specialised robots in any single task.

However, when the task and environment are clearly defined, dedicated solutions are often faster, cheaper, and safer. In a flat warehouse, wheels are more practical than legs; on uneven terrain, four legs are more stable than two; and for repeated operations, a specialised gripper is more precise than a five-fingered hand. Ultimately, choice of form is dictated by the specific use case.