Figure AI CEO Brett Adcock believes scaling humanoid robots will eventually demand more data and computational resources than large language models (LLMs), the AI systems powering tools like ChatGPT. In a recent interview, Adcock outlined a four-stage roadmap for humanoid development and explained why the intelligence-scaling phase will require an unprecedented investment in training data and computing power.

What Makes Humanoid AI Different From Language Models?

The distinction between training a language model and training a humanoid robot comes down to the physical world. Language models learn patterns in text; humanoid robots must learn to perceive their environment, make decisions, and execute precise physical movements in real time. Adcock separates this challenge into distinct phases, each with different resource requirements.

According to Adcock's framework, the first two stages focus on building capable hardware and developing an AI control architecture that can translate observations and instructions into sustained physical work. These stages require prolonged engineering effort but are fundamentally different from the scaling challenge that follows. As Adcock explained, giving an inexperienced team an enormous budget would not automatically produce a working humanoid, any more than it would guarantee a successful orbital rocket.

Why Does Scaling Humanoid Intelligence Cost More Than Scaling LLMs?

Once the basic architecture works, Adcock's thesis centers on a critical insight: investment in training data and compute can translate into a wider range of successful behaviors. This is where humanoids diverge sharply from language models. Humanoids must learn from video recordings of human physical tasks, not text. The quality, relevance, and volume of this data directly determine how well a robot can perform in new environments and situations.

Figure's response to this data challenge is Index, a project designed to collect recordings of people performing physical tasks. During his RoboStrategy interview, Adcock revealed that Figure initially tried purchasing data from outside providers but found the quality insufficient for its needs. The company built its own collection operation instead, emphasizing that spending more on data collection is only useful if the resulting material actually helps the robot learn.

"Spending more on collection is therefore only useful if the resulting material helps the robot learn," Adcock explained, noting that the sensors, the relationship between human demonstrations and robot behavior, and the work of cleaning and checking the data all matter equally.

Brett Adcock, CEO at Figure AI

How to Understand Figure's Four-Stage Roadmap for Humanoid Development

• Stage One: Capable Hardware: Building a machine that meets necessary requirements for speed, torque, dexterity, runtime, weight, and cost. This stage focuses on the physical engineering foundation.
• Stage Two: AI Control Architecture: Developing systems that can translate observations and instructions into sustained physical work. This requires prolonged engineering effort and cannot be rushed with funding alone.
• Stage Three: Expanding Capabilities: Using investment in training data and compute to expand what the robot can do reliably across more tasks and environments. This is where the massive data and computing requirements emerge.
• Stage Four: Manufacturing Scale: Growing production alongside increased usefulness. Commercial viability depends on building large numbers of machines only when their capabilities justify customers buying them.

Adcock's latest prediction makes the scale of his expectations explicit: humanoids will eventually require more training data and computation than large language models. However, he did not provide a quantitative comparison to establish that forecast.

The sequencing of these stages matters strategically. Adcock emphasized that production engineering should not be postponed until AI is finished. Instead, the commercial argument is that building large numbers of machines only makes sense when their capabilities justify customers buying them. This means manufacturing must grow alongside the robot's expanding usefulness.

What Evidence Supports Adcock's Scaling Thesis?

Adcock presented Figure's Helix 2.5 robot as an early way to measure how additional data changes performance. He acknowledged that the release had not yet achieved his wider ambition for long-duration autonomous work in arbitrary homes. The real test will come from future releases and whether performance improvements continue across more tasks, environments, and longer periods of operation.

Adcock described Figure as being near the beginning of the intelligence-scaling stage. This means the company is still in the early phases of the data and compute investment cycle. The question for investors and observers is whether the improvements continue as Figure invests more heavily in training data collection and computational resources.

Adcock's latest post raises the size of the bet significantly. His prediction that humanoids will eventually require more data and compute than LLMs suggests that the robotics industry is entering a phase of investment comparable to or exceeding the billions spent on large language models. The evidence for whether this thesis holds will come from what the robots can consistently do in real-world environments over time.