Hedra Grows Revenue 15x and Opens Its Model Factory to Build the Brains of Physical AI

Physical AI, the branch of artificial intelligence that lets machines see, move and act in the real world, has become one of the fastest-growing corners of venture capital. Startups building robots, autonomous vehicles, drones and defense systems raised $47.4 billion in the first half of 2026, more than the $41.9 billion they raised across 2022, 2023 and 2024 combined. Factories alone installed 542,000 industrial robots in 2024, with more than 700,000 a year expected by 2028.

Every one of those machines needs intelligence built for its own job. A robot that packs boxes and a camera that spots factory defects need very different models, trained on video, sensor and simulation data that is far larger than text. Building those models has required a frontier research lab, while running them in production drives a global AI inference market projected to grow from $106 billion in 2025 to $255 billion by 2030.

Hedra, an a16z-backed research company building models and infrastructure for multimodal and physical AI, launched Hedra Model Factory on 10 November 2026 to take on that work for other companies. The platform post-trains, deploys and keeps improving custom models for robotics, manufacturing, biology and defense teams. Hedra grew its revenue 15 times over the past 12 months. It now opens the research and infrastructure behind that growth to customers.

Physical AI Becomes a Funding Priority

Investors have moved into physical AI at a pace that few other sectors match. The $47.4 billion raised in the first half of 2026 was nearly 80% above the first half of 2025 across 521 deals, while Waymo's $16 billion Series D at a $126 billion valuation, Anduril's $5 billion raise at $61 billion and Shield AI's $2 billion Series G at $12.7 billion showed how large the leaders have become.

Robot foundation model companies are attracting the same appetite. Physical Intelligence, which builds general-purpose models for robots, was reported in March 2026 to be seeking $1 billion at an $11 billion valuation. These companies need large amounts of training data, compute and custom tuning, which is creating demand for infrastructure providers that can build and run models for them.

Every Robot Needs Its Own Model

The International Federation of Robotics counted 4.66 million industrial robots in operation worldwide in 2024, up 9% in a year, with annual installations more than doubling over the past decade. China installed 295,000 of the 542,000 new robots in 2024, about 54% of the total (our arithmetic), while Europe added 85,000 and the United States 34,200. Forecasts point to 575,000 installations in 2025 and more than 700,000 a year by 2028.

Each new robot, gripper and production line changes what a model has to handle. A general-purpose model trained on internet text and images rarely knows how a specific gripper reaches for a part or how a specific camera sees a weld. Gartner expects that by 2027 companies will use small, task-specific AI models three times more than general-purpose large language models, because specialized models are more accurate on domain tasks while running faster on less compute.

Inference Becomes the Main Cost

Inference is the work a model does each time it runs in production, such as predicting a robot's next move or editing an image. It is where most AI spending is heading. MarketsandMarkets projects the AI inference market will grow at 19.2% a year, from $106.15 billion in 2025 to $254.98 billion by 2030, which is about 2.4 times larger.

For physical AI the cost of inference decides whether a product works at all. A robot has to respond in real time, often on a small computer inside the machine, so a model that is accurate but slow is not usable on the factory floor. Companies that can make models both smaller and faster without losing accuracy turn that cost into an advantage, which is the gap Hedra Model Factory is built to close.

Hedra Model Factory

Hedra Model Factory is an autonomous post-training and inference platform. Post-training is the stage after a base model is built, when it is tuned on a customer's own data and tasks, while inference is the stage where the finished model runs in production. The platform combines four parts in one loop, with flexible GPU sandboxes, post-training led by Hedra researchers working with autonomous AI agents, co-optimized low-latency deployment and a data flywheel that keeps improving each model after it ships.

Teams start with compute on Hedra's inference engine, scaling training up without long-term GPU commitments. They then bring their simulation engines and data, while Hedra helps build the reinforcement learning environments and datasets they need. Reinforcement learning trains a model by rewarding good actions in a simulated or real setting. Hedra's researchers use vision-language models, AI systems that understand both images and text, to caption and score data or give feedback during training.

Deployment runs on low-latency compute on premises or at Hedra's nearest partner data center. The work ranges from distilling world action models, which predict how a scene will change after an action, so they run faster than real time, to fitting a vision-language-action model onto an edge device inside a robot. Once a model is live, Hedra researchers analyze its production performance, retrain it and release only versions that pass the customer's own evaluations.

For companies without their own research lab, Hedra researchers also set up data capture, evaluations and the full loop from training to serving. Teams that already have a model can bring it to run on Hedra's inference engine. The offer gives companies the capabilities of a frontier AI lab without having to build one.

How Recursive Self-Improvement Works

Recursive self-improvement describes AI systems that get better at improving themselves. The idea is already speeding up how frontier models are built. Hedra applies it to model making, with researchers working alongside autonomous AI scientists that run experiments and iterate on training runs while every deployed model feeds data back into the next version.

Real-world data makes the loop harder and more valuable. Video, sensor and simulation data are orders of magnitude larger than text, while the simulators used for reinforcement learning often need GPUs just to run at scale. Hedra's aim is a factory that gets faster with every model it produces, so each customer project adds to the methods and data that improve the next one.

Speed on the Same Hardware

Hedra's clearest proof point comes from image editing. Working with Krea 3, an image generation and editing model, Hedra's team cut the time for a 2K image edit from about 22 seconds to 3.6 seconds on the same four AMD Instinct MI355X GPUs in under a week. That is about 6.1 times faster and 84% less time per edit (our arithmetic). Krea 3 now runs on MI355X GPUs on Hedra.

The gain came from distillation and inference optimization. Distillation trains a smaller or faster model to copy a larger one. Hedra replaced the original model running in bf16 precision at 30 sampling steps with a distilled version that needs only 14 steps, 53% fewer (our arithmetic). The same methods are what Hedra now offers for robotics and manufacturing models, where response time decides whether a system can work in real time.

Physics and Robots in 2027

Video models that respect physics are the training ground for robots, since a model that understands how objects fall, slide and collide can help a robot plan its actions. Hedra retrained its H3 video model for physical consistency. The new version scored 34.9 on Google DeepMind's Physics-IQ benchmark, against 30.3 for H3 Max, a 15% gain with comparable visual quality (our arithmetic). The benchmark tests 396 real-world videos across 66 scenarios in mechanics, fluids, optics, thermodynamics and magnetism, while the best score in its original 2025 paper was 24.1.

An autonomously distilled version of the model rendered a five-second clip in under three seconds on AMD MI355X GPUs, faster than real time. In a separate simulation study, Hedra fitted a robot running a trained policy with a new gripper and restored task success from near zero to near 100% with no retraining, by matching the new gripper's reach and grasp point in the wrist camera view. Hardware changes like this happen constantly in factories, so fixing them without a full retraining cycle saves weeks of work.

From Character Video to Physical AI

Hedra launched in 2024 under founder and chief executive Michael Lingelbach, who has spent his career on reinforcement learning for visual and physical systems, first at Stanford and NVIDIA. The company first became known for Character-3, a model that generates talking and moving characters for video. By May 2025 its creator platform had more than 3 million users and more than 10 million videos made.

Hedra has raised $44 million, including a $10 million seed round from Index Ventures, Abstract Ventures and a16z speedrun and a $32 million Series A in May 2025 led by Andreessen Horowitz's infrastructure fund. Its research team brings together talent from Stanford, Berkeley and UIUC with experience at NVIDIA, Meta Superintelligence Labs, Fireworks AI and Salesforce, while it has published research in generative video and embodied AI at NeurIPS, CVPR and ICLR.

Availability and Outlook

Hedra Model Factory is available from today at the new hedra.com. Teams can work with Hedra researchers to build and deploy a custom model for a problem their general-purpose model cannot solve, start with flexible GPU sandbox capacity without long-term commitments or serve an existing model on Hedra's inference engine.

Physical AI will need thousands of specialized models, one for each robot, camera, production line and lab instrument, while each one has to be affordable to build and fast enough to run in the real world. Hedra grew revenue 15 times in a year building and running models of its own. With Model Factory it is offering that capability to the companies racing to put AI into the physical economy, aiming to become the factory behind that shift.

Don’t forget to like and share the story!