Figure AI is planning an unprecedented datacenter buildout that could bring frontier AI supercomputer scale to humanoid robotics for the first time.
The Silicon Valley startup says it’s partnering with the British AI infrastructure giant to build the massive computing capacity needed to train its humanoid robots for jobs across sectors and households. At full scale, the project would be in the same computing class as AI clusters being built by Big Tech heavyweights like Meta, Microsoft, OpenAI, and SpaceX.
The level of infrastructure has mostly been associated with training large language and generative AI models. Figure AI is banking on the same frontier-scale approach to unlock general-purpose capabilities so robots can see, reason, move, and interact with the world just like people.
The 4-year-old robotics firm, led by serial entrepreneur Brett Adcock, says it’s making an $3.5 billion commitment for the project with potential to scale to $6 billion. The first phase will be deployed at Nscale’s campus in Barstow, TX, with initial systems targeted to go online in the second half of 2027.
“We’ve seen incredible growth with inference and agentic AI, and Figure is pushing the boundaries of AI even further,” Nscale CEO Josh Payne said in a press release. “We’re proud to be enabling the future of AI robotics together.”
Figure says the multi-year project could scale to a 100,000 NVIDIA GPUs powered by the next generation Vera Rubin platform. According to NVIDIA, a single Vera Rubin GPU can deliver up to 50 petaflops for AI inference, 35 petaflops for AI training, and 22 terabytes-per-second of memory bandwidth.
Rubin’s reported performance is up to 5x higher than NVIDIA’s Blackwell chips used in many of the world’s most powerful AI data centers today. NVIDIA connects 72 of the processors into a single-rack-size system with an estimated value around $8 million, with about $2 million of that tied to memory. In a news release, NVIDIA founder/CEO Jensen Huang called the project “the physical AI flywheel that will accelerate ethe path from models to robots in the world.”
The planned supercomputer will process a massive stream of data Figure is collecting from people all over the world as part of its recently announced Index program through a smartphone app. People Figure calls Creators can get paid to perform everyday tasks while wearing head-mounted cameras that record their movements.
Before its official launch, the program operated in stealth with 44,000 weekly active users across 108 countries collecting the equivalent of almost 5 years of human activity every day. Figure reported paying around $15 million so far at an undisclosed by-the-minute rate.
According to Figure, the raw footage runs through a five-stage data pipeline that turns millions of messy human videos into structured AI training material. The platform reportedly processes more than 35 minutes of new human task video every second. Automated systems remove lowquality and near duplicate recordings, organize the remaining footage by task, and add text descriptions explaining what the person is doing.
The cleaned footage feeds into Figures self-developed AI system called Helix, which learns the connection between what people see, the actions they take, and the tasks they’re trying to complete. Helix uses those patterns to teach figures robots to respond when they encounter similar situations in the real world. With enough varied training data, the goal is for robots to adapt on the fly to new tasks and environments.
Figure has committed to more than $1 billion over the next 12 months for data and compute.
There’s a gold rush happening right now to collect the type of footage Figure is capturing at scale. Dozens or possibly hundreds of companies are distributing specialized cameras, gloves, and other devices to get training data for robots.
A stream of reported breakthroughs increasingly seems to back the idea that scaling data and compute will work for robots like it did for generative AI. But they should also be taken with a grain of salt since the most impressive results come directly from startups trying to raise enormous amounts of money.
- Generalist AI, which operates out of San Francisco and Boston, says its latest model can learn some new physical tasks from from a single 3-to-12-second demonstration and even improvise when conditions change. The startup says the abilities emerged after the AI continuously absorbed physical interaction data for more than eight months.
- In Redwood City, the 2-year-old startup Dyna Robotics reported that its latest model more than doubled its average task success rate from 20 percent to 53 percent as it scaled from 1,000 to 1 million hours of human video. The pre-training data came from Build AI, which says it plans to expand its egocentric human video data set to 10 million hours by November 2026.
- The Pittsburgh-based startup Skild AI, which has raised more than $2 billion to date, recently reported similar results. After scaling its training data from 1,000 to 100,000 hours, its new model could watch a single video of a task it had never seen before and then attempt it itself. Success rates on unseen tasks reached 66 percent.
- Back in Silicon Valley, Sunday Robotics reported that its Memo humanoids achieved a 99.1 percent success rate, folding unfamiliar laundry in homes it had never seen before after scaling data collected from people wearing custom grippers.
Figure says it tried using robot training data from outside vendors, but it lacked the diversity and quality its Helix system needs. The company says controlling its own data pipeline means it can collect much larger and more varied data sets specifically for its hardware.
Figure has been sharing demos showing a broad set of autonomous capabilities since launching Helix in early 2025. Its Figure 03 humanoids have been shown doing long horizon kitchen work, fine manipulation, tidying up a mock living room, teaming up to reset a bedroom, taking turns on a logistics conveyor setup during an 8-day livestream, and working as a small fleet at BMW’s plant in Spartanburg, SC.
Adcock has also claimed that the Figure 03 was acting fully autonomously during its March 2026 visit at the White House, where it walked alongside First Lady Melania Trump, addressed attendees, greeted them in multiple languages, and walked out.
Most recently, Figure shared footage of its flagship humanoid climbing up and down a short ladder, coordinating its hands and feet rung by rung while maintaining balance. Adcock said the capability was developed for a real customer deployment where the Figure 03 must climb to a mezzanine, perform its work, and climb back down at the end of its shift.
The Figure 03, introduced in October 2025, stands 173 cm (5 feet 8 inches). The fully electric humanoid weighs about 62 kg (136 lbs), operates for about 5 hours on a single battery charge, and comes equipped with five-fingered dexterous hands. The robots are not yet available for general purchase, though hundreds of units have been built so far for internal use and pilot deployments.
Figure has now publicly committed at least $4.5 billion across its data and compute initiatives, potentially scaling to more than $7 billion if its Nscale agreement reaches its stated maximum. Exactly how Figure plans to pay for all of this remains unclear.
The company has raised around $2 billion from investors so far, including more than $1 billion from its 2025 Series C round that valued it at $39 billion. Funding its aggressive expansion will require another enormous private round, substantial debt or infrastructure financing, or maybe an IPO.
Though Figure has not indicated it plans to go public, the idea is becoming more realistic. Adcock has already taken one capital-intensive startup public with Archer Aviation, while Unitree’s blockbuster debut on the Shanghai STAR Market has shown that public investors are willing to pour billions into the humanoid race.
Figure is scaling its computing infrastructure to compete with Tesla, which already operates massive AI supercomputers to train both its autonomous driving systems and Optimus humanoid robots. Elon Musk has said he intends to unveil the highly anticipated third generation Tesla bot by the end of 2026. It’s the first iteration intended for mass production.
The EV maker has been installing a dedicated production line at its Fremont factory designed to eventually scale to a million units per year. While Musk has described high-rate manufacturing as the greatest challenge to bringing Optimus to market, Adcock argues that the harder problem is intelligence.
Musk has compared mass-producing Optimus to building a car, only far more complicated, while Adcock likens humanoid robots to smartphones. He describes them as standardized hardware whose real value comes from increasingly capable software and artificial intelligence.
The two also differ on China. Musk has said his only real competitors in humanoid robotics are in China because of its increasing strength in AI, engineering, and mass manufacturing. Adcock, however, has been dismissive of Chinese humanoid firms, saying they’re far behind figure in the autonomous intelligence needed to make humanoids genuinely useful.
Both Tesla and Figure plan to eventually sell their humanoids to the general public for around $20,000, but it’s unclear when either will go to market. One robot that’s available for purchase today is this robotic duck from Hugging Face, which was just acquired by NVIDIA for $13 billion. Read all about why the Microduck has captivated the robotics world right here.