What tech stack does Figure AI use?

Figure's AI training infrastructure runs on Microsoft Azure, specifically NVIDIA ND H100 GPU clusters — a commitment anchored by the Series B partnership with Microsoft ($95M investment + Azure infrastructure agreement). This is one of the most clearly confirmed aspects of Figure's stack: the company explicitly cited Azure and H100s in its Series B announcement and has maintained that partnership through Series C. NVIDIA is both a strategic investor (Series B and C) and a core technology partner via the Isaac platform.

The Helix model architecture has evolved substantially since launch. Helix (original) uses a dual-system architecture: System 2 (S2, ~7B parameters, 7–9Hz) handles high-level visual reasoning and language understanding; System 1 (S1, ~80M parameters, 200Hz) handles low-latency motor control. Helix 02 (January 2026) added a third layer: System 0 (S0), a neural locomotion controller trained on over 1,000 hours of human motion data that replaced 109,000+ lines of hand-engineered C++ code, extending control from the upper body to full-body including walking, balance, and torso coordination. All three systems run as a unified visuomotor neural network.

NVIDIA Isaac GR00T, Isaac Sim, and Isaac ROS are used for simulation, sim-to-real transfer, and robotics middleware — confirmed via the NVIDIA strategic partnership at both Series B and Series C. On-robot inference hardware is developed in-house; NVIDIA Jetson Thor-class silicon is the most credible inference platform given Figure's NVIDIA relationship and the Isaac platform dependency, though this has not been formally confirmed in a public disclosure.

The manufacturing operations at BotQ run entirely on custom software: a Manufacturing Execution System (MES), Product Lifecycle Management (PLM), Enterprise Resource Planning (ERP), and Warehouse Management System (WMS) — all built internally. This is an unusual and deliberate choice reflecting Figure's view that off-the-shelf solutions iterate too slowly and cannot handle the unique production constraints of humanoid robot assembly. The custom stack controls more than 150 networked manufacturing workstations and has enabled BotQ to achieve 80%+ end-of-line first-pass yield and 99.3% battery production yield.

For the AI data pipeline, Figure collects multimodal training data across multiple channels: human video demonstrations, teleoperation data from commercial deployments (BMW, and ongoing Figure 03 commercial customers), and tactile plus proprioception sensor inputs from the robots themselves. Helix 02's tactile sensors detect forces as small as three grams — sensitive enough to feel a paperclip — generating high-resolution haptic data streams that feed the training dataset, now estimated at approximately 20TB. The Brookfield Asset Management Series C partnership creates a new data channel: 100,000 managed residential units will serve as real-world Helix training environments, providing data from household and commercial settings at a scale no competitor has announced.

Python is the overwhelmingly likely AI development language — consistent with PyTorch, industry norms for VLA model development, and Figure's AI engineering job postings. C++ is likely used for performance-critical embedded and robotics subsystems. GTM tooling (CRM, sales engagement, marketing automation) has not been publicly disclosed and is not included here.

Figure's deep commitment to custom-built manufacturing software (MES, ERP, PLM, WMS) signals a build-over-buy posture for core operations. This is a deliberate architectural choice — vendors selling standard manufacturing software will face resistance unless they can demonstrate clear advantages in flexibility, integration with Figure's robotics systems, or speed of deployment. The more tractable entry points at the manufacturing layer are at the periphery: quality assurance tooling, supply chain visibility platforms, industrial IoT sensors, and precision manufacturing component suppliers who can meet Figure's cycle time requirements (components produced in under 20 seconds).

On the AI infrastructure side, Figure is a confirmed and growing buyer of NVIDIA GPU hardware and Microsoft Azure services — both relationships are locked in at the investor level and are unlikely to be displaced. Vendors in the MLOps, data labeling, simulation, and robotics middleware space (especially those integrated with NVIDIA Isaac or Azure's AI platform) have a natural integration angle and can potentially present as additive to these strategic relationships rather than competitive. The Brookfield data partnership opens an adjacent opportunity for facilities management and real-estate software vendors — any vendor already embedded in Brookfield's 100,000-unit residential property management stack has a warm path to the data collection workflows Figure will be running in those environments.

For enterprise software vendors more broadly, Figure's headcount of ~619 (April 2026) and $39B valuation mean it is now squarely in the enterprise buyer segment for HR, finance, legal, security, and collaboration tooling. The rapid headcount growth (from ~180 in early 2024) creates active procurement needs in people operations, workforce management, and employee experience software.