Anvil's most recent round was a $5.5M seed closed in April 2026, co-led by Matter Venture Partners and Humba Ventures, with participation from DNX Ventures, Vivek Sodera, Spacecadet Ventures, and Position Ventures.

Before the seed, Anvil raised a $1M pre-seed from Matter Venture Partners in 2025, shortly after the company was founded.

Total capital raised stands at $6.5M across both rounds.

Anvil sells composable robot devkits and a standardized software stack for teams building physical AI applications. The problem it targets is time to first data: assembling a working robot system from scratch, including arms, cameras, controllers, compute, teleoperation software, and training pipelines, typically consumes five to six months and several engineers before a team can collect a single useful data point. Anvil packages those components into a ready-to-run kit.

The two main robot embodiments are OpenYAM, a desktop manipulator for lighter tasks, and OpenARM, a bimanual 7-degrees-of-freedom arm for more dexterous manipulation. Both ship with the Anvil Devbox, a prebuilt Ubuntu-based workstation that runs robot controllers, inverse kinematics, teleoperation, data collection, and safety guardrails out of the box.

Teleoperation is the primary workflow. A researcher or engineer uses Meta Quest controllers, and Anvil's software maps their hand movements into robot motion at up to 90 Hz over a wired USB connection. Grip buttons mirror motion, triggers open and close grippers, and A/B buttons start and stop recording episodes. Anvil frames this Quest-based approach as an alternative to traditional leader-follower setups, which it says can create operator fatigue and kinematic mismatch over long data-collection sessions.

Recorded demonstrations are saved as MCAP files on the Devbox and organized into sessions and episodes through a local web app. Anvil's tooling converts recordings into LeRobot v3.0 format for training, and the stack supports ACT, Diffusion, SmolVLA, Pi0, Pi0.5, and fine-tuning of models like NVIDIA GR00T and Pi 0.6, with integrations for ROS2 and Foxglove for inspection and replay.

The workflow runs from unboxing through teleoperation, episode recording, data export, policy training, offline evaluation, and inference deployment back onto Anvil hardware. The company says teams can complete that loop on day one. Store prices run from $1,900 for OpenYAM or the Devbox alone, to $5,600 for OpenARM 2.0, $8,770 for the OpenARM Quest Teleop Kit, and $13,920-$15,120 for the OpenARM Leader-Follower Kit.

Anvil sells to research labs, startups, and enterprise robotics teams through a direct, product-led model. Customers configure and purchase devkits through Anvil's online store, with transparent pricing and without long integration engagements. Every kit ships with an enterprise support package bundled in, so the initial hardware sale also establishes a support relationship.

The monetization logic today is hardware-first: revenue scales with systems shipped and average selling price per configuration. Because most customers buy full devkit bundles rather than individual components, blended ASPs cluster in the $9K-$11K range, well above entry-level component prices. Accessories such as grippers, camera modules, wheeled lift bases, and additional Quest kits add to the base system sale.

Gross margins on hardware are currently in the 20-30% range, reflecting Anvil's Taiwan-based manufacturing operation, CNC and sheet metal assembly costs, and relatively low volumes. The company is its own manufacturer rather than a pure design house, which adds operational complexity but gives it control over availability, BOM choices, and iteration speed. If volume increases, that margin profile should improve.

The expansion path is hardware to software. Each devkit sale installs Anvil's controls stack, data pipeline, and training workflow as the team's default physical AI environment. Once a team's demonstrations, datasets, and deployment recipes are built around Anvil's tooling, adding more arms or upgrading embodiments is far easier than replatforming, creating a path to future software, fleet monitoring, and services revenue without a separate sales motion.

Anvil competes in an emerging physical AI enablement layer where vendors are trying to own the developer workflow from robot setup through policy deployment. The field includes open-source ecosystems, integrated kit vendors, and industrial incumbents converging on the same buyer.

Trossen Robotics is the closest direct competitor, selling packaged hardware for teleoperation, dataset recording, and model deployment across single-arm, stationary bimanual, and mobile configurations. Trossen markets around the ALOHA ethos of low-cost ML hardware and has integrations with both Physical Intelligence and Hugging Face LeRobot, competing on the same core promise as Anvil.

Trossen's advantage is breadth: a wider menu of embodiments can capture more lab budgets and keep customers in-house as needs evolve. Anvil's counter is tighter productization, faster shipping from its own manufacturing operation, and a more opinionated end-to-end workflow.

Hugging Face LeRobot is not a hardware vendor, but it increasingly acts as a platform substitute by owning a standard workflow for robot learning. LeRobot supports SO-101, OpenArm, Reachy 2, Unitree G1, and other embodiments, includes browser-based calibration and teleop via LeLab, and has NVIDIA integrating Isaac and GR00T directly into the stack.

For Anvil, LeRobot is both a distribution channel and a commoditization risk. As the open software stack becomes more complete, Anvil has to justify its value through integration quality, support reliability, and time-to-first-policy rather than component aggregation. OpenArm presents a related thin-layer risk: Anvil packages OpenARM into its kits, but as OpenArm 2.0 expands into a broader reproducible research stack with an evaluation cell and passive teaching device, Anvil's overlay has to offer more than convenient bundling.

ABB, FANUC, Universal Robots, and KUKA are incorporating NVIDIA simulation and edge AI into production robot stacks. FANUC supports ROS2, Python, and digital twins via Isaac Sim; ABB is integrating Omniverse into RobotStudio; Universal Robots launched an AI Accelerator built on Jetson AGX Orin and Isaac libraries.

These incumbents control procurement relationships, safety certifications, and enterprise trust in ways startup devkit vendors cannot easily replicate. If physical AI development becomes good enough inside their own ecosystems, enterprise customers in regulated or uptime-sensitive environments may stay with approved industrial vendors rather than adopt a startup stack.

Anvil's current wedge is horizontal, selling manipulation infrastructure to teams building physical AI, but its expansion logic runs in three directions: up the software stack, deeper into specific verticals, and outward geographically as the physical AI developer base globalizes.

Anvil already bundles controls, data collection, training wrappers, and deployment into its devkits. The next step is to separate and monetize those layers: experiment tracking, fleet monitoring, policy governance, rollback tooling, and production-grade observability are capabilities the platform already touches in early form.

NVIDIA's March 2026 Physical AI Data Factory Blueprint frames the industry around generating, augmenting, and evaluating training data at scale, which validates this direction. A team that collects demonstrations on Anvil hardware and trains policies through Anvil's pipeline is a natural buyer for managed data services, hosted evaluation, and deployment controls, without Anvil needing to create a new product category.

Anvil sells horizontally today, but its composable architecture fits a market where many customers need mostly standard, partly custom systems. Food and beverage, light manufacturing, logistics, and biohazardous handling are verticals where structured workflows and fixed workstations concentrate near-term automation demand, and where legless manipulation platforms have a cost and engineering advantage over humanoid alternatives.

The research-to-enterprise conversion path is another expansion vector. Universities and early-stage startups are top-of-funnel buyers; as their projects move into production pilots, Anvil can follow with purpose-built cells, custom integrations, and higher-ACV enterprise packages. The IFR's World Robotics 2025 data shows professional service robot sales growing across logistics, inspection, and maintenance, categories fragmented enough to reward a modular platform over a single fixed-form robot.

Anvil's kits are built around globally shared tooling, MCAP, ROS2, LeRobot, Foxglove, which makes the product shippable without heavy local integration work. The company already ships to 60+ countries, and its Taiwan-based manufacturing operation gives it a China-plus-one alternative for teams that need to source outside mainland China.

That supply chain posture is increasingly strategic. Anvil aggregates demand across hundreds of customers, lowering the risk for Taiwanese, Korean, and Japanese component partners to invest in actuator and sensor production that would otherwise require guaranteed downstream buyers. As that non-China supply base matures, Anvil gains both margin leverage and a differentiated sourcing position for enterprise and defense-adjacent customers facing procurement restrictions on Chinese-origin components.

Open-source commoditization: As Hugging Face LeRobot expands hardware support, NVIDIA deepens its Isaac and GR00T integrations, and upstream embodiment projects like OpenArm add teleoperation and evaluation workflows, a sufficiently motivated team can reproduce much of Anvil's current integration layer with open-source components, which could compress differentiation to support quality and time-to-value rather than a proprietary technical layer.

Hardware margin ceiling: With gross margins currently in the 20–30% range and a cost structure tied to Taiwan-based CNC manufacturing, international freight, and tariff handling, Anvil's path to a durable business depends on either scaling volume fast enough to drive meaningful BOM leverage or layering in higher-margin software and services revenue before hardware pricing pressure from lower-cost Chinese competitors like Unitree intensifies.

Research-to-production gap: Anvil's value proposition is strongest when it gets small teams running quickly and cheaply, but converting early devkit buyers into large recurring customers requires physical AI applications to graduate from prototype to production deployment, a transition that most robotics projects have historically failed to complete at speed, leaving Anvil exposed to a top-of-funnel that grows faster than its installed base converts into durable platform revenue.