Figure AI has identified computing power as the primary constraint holding back humanoid robots from reaching homes worldwide. The company just signed a multiyear partnership with infrastructure provider Nscale that includes an initial $3.5 billion compute commitment, with potential expansion beyond $6 billion. The deal underscores a fundamental shift in the robotics industry: building better robots is no longer the hardest problem. Getting enough computing power to train and deploy them is.

Why Is Computing Power Suddenly the Bottleneck?

Figure's founder and CEO Brett Adcock put it plainly: "To bring humanoid robots to every home in the world, we are largely constrained by data and compute." This statement reflects a maturation in the field. The robots themselves are becoming capable enough to perform real work. What's missing is the infrastructure to train them at scale and the computational horsepower to run their AI models in the field.

The Nscale partnership addresses this head-on. Nscale will provide power, computing infrastructure, and software to coordinate workloads across what could eventually reach 100,000 NVIDIA GPUs. Initial deployments are targeted for the second half of 2027 in Barstow, Texas. The companies plan to train Figure's models on NVIDIA Vera Rubin infrastructure, validate them in NVIDIA Isaac Sim simulation software, and deploy them on NVIDIA GPUs inside the robots themselves.

Nscale is also becoming a shareholder in Figure and its preferred compute provider, cementing a long-term relationship. This isn't just a vendor deal; it's a strategic partnership that signals both companies believe computing capacity will define who wins in humanoid robotics over the next several years.

What Does This Mean for the Broader Robotics Race?

Figure's computing challenge sits alongside other infrastructure gaps that are slowing the entire industry. While Figure focuses on data and compute, other companies are grappling with a different constraint: the physical components that make robots move. Apptronik CEO Jeff Cardenas recently warned that the United States lacks the domestic supply base needed to manufacture humanoid robots at scale, particularly for gears and actuators.

Actuators, which combine motors, gearboxes, and electronics, can represent up to 60 percent of a robot's bill of materials. Cardenas argued that while the U.S. remains ahead in artificial intelligence, China has built a strong industrial base directed toward robotics manufacturing. "It's not just enough to have the software," he said. "We've also got to have world-class hardware to go along with it." This means Figure's computing ambitions will only matter if companies like Apptronik can solve the manufacturing puzzle on the hardware side.

Cardenas

How Are Companies Building the Data Foundation for Humanoid Training?

• Figure's Index Program: Figure is collecting recordings of people performing everyday activities to train its Helix robot AI models, turning human demonstrations into training material without requiring robots to perform every task themselves.
• Tesla's Berlin Factory Initiative: Tesla reportedly equipped selected Grünheide factory workers with wearable cameras to capture their movements and work sequences, providing practical training data grounded in real production tasks.
• Boston Dynamics' Manufacturing Focus: Boston Dynamics is training its Atlas robot at a new facility at Hyundai's Georgia plant, concentrating on automotive parts sequencing and manufacturing workflows.

These approaches reflect a broader industry recognition that training data is as critical as computing power. Figure's Index program has grown to over 69,000 weekly active users, showing how companies are scaling data collection across distributed networks rather than relying solely on robots to generate their own training material.

Tesla's Berlin initiative, first reported in July, demonstrates how existing manufacturing operations can contribute to humanoid development. Workers' demonstrations of tool handling, component orientation, and work sequences provide valuable examples that AI systems can learn from, even if the robot's physical form differs from a human's.

The distinction matters: collecting footage of humans is one step in training a robot, but it's not a complete solution. A robot has different joints, hands, and physical constraints than a human, so its learning system still needs to translate observations into actions it can execute reliably. However, starting with human demonstrations is faster and cheaper than having robots learn through trial and error alone.

What's the Timeline for Deployment?

Figure's partnership with Nscale targets initial GPU deployments for the second half of 2027 in Barstow, Texas. This timeline suggests that Figure expects to have trained models ready for deployment within the next year, assuming the infrastructure comes online as planned. The company is already testing its Helix 2.5 robot in homes, so the computing infrastructure deal appears designed to accelerate the transition from pilot programs to broader deployment.

Apptronik's CEO offered a longer-term perspective on the industry's development cycle. He compared humanoid robotics to personal computing in the early 1980s, describing the industry as being at the beginning of a roughly 40-year cycle. His expected progression runs from heavy industry into retail and healthcare, then broader domestic use. He predicted humanoids would start entering homes in the next few years, while qualifying that doing the full range of household chores people want would take longer.

The gap between Figure's near-term computing ambitions and the broader industry's longer development timeline reflects two different challenges converging. Figure is solving the compute problem to enable rapid scaling of robots that already work. The rest of the industry is still solving the manufacturing and capability problems that will determine whether those robots can actually be built and deployed at the scale Figure envisions.