Figure Commits $3.5B+ for 100,000 GPUs to Train Home Robots

Figure, the humanoid robotics company, is committing $3.5 billion initially — scaling beyond $6 billion — to deploy up to 100,000 NVIDIA Vera Rubin GPUs, in partnership with neocloud provider Nscale. The announcement came via X on September 3, 2026: "Bringing a robot into every home demands compute at an unprecedented scale." Initial deployment is targeted for Barstow, Texas, starting the second half of 2027.

It's a striking number for a company whose product is a physical robot, not a chatbot — and that's exactly the point worth unpacking: this deal is a reminder that the compute race isn't just a language-model story anymore.

TL;DR

Why a robot company needs a frontier-scale training cluster

The instinct to question "why does a robot need 100,000 GPUs" is reasonable, and the answer is that this compute almost certainly isn't for the robots themselves. Figure's robots run inference locally or at the edge, on hardware built for real-time motor control, not on a data-center cluster. The 100,000 GPUs in this deal are for training — specifically, training the foundation models (Figure's Helix line, which explainx.ai covered when two Helix-02 robots coordinated to tidy a bedroom from a single learned policy) that eventually get deployed to run on the robot itself.

Training a model that controls a physical body is arguably a harder compute problem than training a language model, not an easier one. A humanoid robot's policy has to learn continuous motor control across dozens of degrees of freedom, from video, simulation, and teleoperation data, and generalize that learning to real-world physics that a simulator can only approximate. That's the compute Figure is buying — the ability to iterate on foundation models at a pace and scale that matches what frontier AI labs already do for language models, applied to embodied intelligence instead.

What Vera Rubin actually is, and why the platform generation matters

NVIDIA's Vera Rubin is the compute architecture generation succeeding Blackwell — pairing NVIDIA's custom "Vera" CPU with "Rubin" GPUs in a full-rack, liquid-cooled platform built for both training and inference at frontier scale. Committing to Vera Rubin specifically, more than a year before deployment, is a bet on where the compute frontier will actually be by the time the hardware ships — not a purchase of currently-available capacity.

This is the same generational bet explainx.ai covered when NVIDIA partnered with Ilya Sutskever's Safe Superintelligence for a roughly 10x compute scale-up on the same platform. The pattern worth noticing: Vera Rubin commitments are no longer confined to the handful of labs building general-purpose frontier language models. A robotics company making the same architecture bet, at a comparable scale, is a concrete signal that the compute buildout race has a second front — physical AI — running in parallel to the one everyone's used to watching.

The Nscale piece: why a neocloud, not a self-built data center

Figure isn't building this data center itself — Nscale is. Nscale is a neocloud: a company whose business is building and operating GPU infrastructure at scale, then renting that capacity to AI labs and companies that need frontier training clusters without taking on the data-center construction, power procurement, and cooling engineering themselves.

For a robotics company, this division of labor makes sense on its own terms. Figure's core expertise is robot hardware and foundation-model training; a data center in Barstow, Texas, with the power and liquid-cooling infrastructure a 100,000-GPU Vera Rubin deployment requires, is a different discipline entirely. Partnering with a neocloud is the faster, lower-execution-risk path to that scale of compute than standing up the physical infrastructure from scratch — the same logic explainx.ai has covered from the labor side, where AI companies are hiring electricians and carpenters by the thousands specifically because this kind of data-center buildout has become its own industrial-scale undertaking, separate from the AI research it ultimately supports.

What's confirmed versus what's still ahead

Worth stating plainly: the dollar figures here are a stated commitment from Figure's own announcement, not a completed build. Initial deployment isn't targeted until the second half of 2027 — more than a year from this announcement — and "up to 100,000 GPUs" and "plans to scale beyond $6 billion" are both explicitly forward-looking language, not a confirmed final total. What's confirmed today is the partnership itself, the initial $3.5 billion commitment, the platform generation (Vera Rubin), and the target location (Barstow, Texas). The execution — actual GPUs racked, actual models trained at that scale, actual robots benefiting from the resulting foundation models — is the part to watch for over the next 12-18 months.

Related on explainx.ai

• Figure Helix-02: two humanoid robots collaborate to tidy a bedroom
• NVIDIA x SSI: Ilya Sutskever's lab gets Vera Rubin and a 10x compute bet
• AI companies are hiring electricians and carpenters by the thousands
• NVIDIA's $500 billion compute becomes a Wall Street asset class
• Gemini Robotics 2: whole-body intelligence
• 1X Neo: 25-DOF hands and the Physical API
• AI chip architectures — GPU, TPU, Trainium, Cerebras, Groq guide

Primary source: Figure (@Figure_robot) on X, September 3, 2026; Figure and Nscale Sign Strategic Partnership — figure.ai.

This post reflects Figure's public announcement as of September 3-5, 2026. Deployment is targeted for the second half of 2027; final GPU count and total investment are stated as "up to" and "beyond" figures respectively, and may change before deployment begins.