XDOF Could Reach a $1.2 Billion Valuation: Series B Talks, 8VC, Robot Training Data, and the Physical AI Race

XDOF has moved from a newly revealed robotics-infrastructure startup to a potential unicorn in a matter of months. The company is reportedly in late-stage discussions for a Series B financing led by 8VC at a valuation of about $1.2 billion, even though the size of the round and the final valuation mechanics have not been disclosed. The talks are not a completed financing, and neither the valuation nor the lead investor should be treated as final until the transaction is announced.

The speed of the process is notable because XDOF emerged from stealth in June 2026 after a reported $70 million Series A. The latest financing discussions are being driven by a business that has reportedly reached an annualized revenue level approaching $50 million while selling robot-training-data infrastructure to frontier AI laboratories and robotics companies. That combination places XDOF at the intersection of two capital-intensive markets: foundation-model development and general-purpose robotics.

The company is not primarily trying to build a consumer robot. Its core proposition is to solve a data-supply problem: robots need large quantities of high-quality physical-interaction data, but there is no equivalent of the open web that can be scraped at internet scale. XDOF therefore builds the systems used to collect, structure, filter, annotate, and deliver the demonstrations that robot policies learn from.

THE REPORTED SERIES B AND THE $1.2 BILLION VALUATION.

What is currently reported, what remains unconfirmed, and how the financing compares with XDOF’s recent growth.

XDOF is reportedly in late-stage talks to raise a Series B at an approximately $1.2 billion valuation, with 8VC expected to lead. The total amount of new capital has not been disclosed, and it is also unclear whether the valuation under discussion is pre-money or post-money. Those details matter because the ownership sold in the round, dilution for existing investors, and the implied price per share depend on the final structure.

The timing is unusually compressed. XDOF had not been expected to return to the market so quickly after its June financing, but investors reportedly approached the company as its commercial growth accelerated. Annualized revenue is said to be approaching $50 million, while the company has previously been reported to work with roughly 20 customers, including several frontier AI laboratories.

........

Element	Current position	Technical or financial implication
Series B status	Late-stage talks	Terms can still change before signing or closing
Valuation	About $1.2 billion reported	Would move XDOF into unicorn territory if completed near that level
Lead investor	8VC reported as lead	Not yet reflected as a completed financing on XDOF’s public materials
Round size	Not disclosed	Dilution and post-money ownership cannot yet be calculated
Prior financing	$70 million Series A reported in June 2026	The new round would follow only a few months after the previous raise
Annualized revenue	Approaching $50 million reported	A $1.2 billion valuation would equal roughly 24× that run-rate figure
Customer base	About 20 customers previously reported	Commercial concentration and contract durability remain important valuation variables

........

Using the reported annualized revenue level only as a rough reference point, a $1.2 billion valuation would represent about 24 times a $50 million run rate. That is not directly comparable with a mature software revenue multiple because XDOF’s economics combine software, robotics hardware, field operations, human data collection, annotation, quality control, and customer-specific delivery. The eventual gross margin profile is therefore as important as top-line growth.

··········

WHY ROBOT TRAINING DATA IS BECOMING INFRASTRUCTURE.

Physical AI needs demonstrations that encode actions, forces, timing, geometry, and task outcomes rather than text alone.

Large language models benefited from an enormous pre-existing corpus of text, code, images, and web documents. Robotics has no equivalent universal dataset for physical interaction. A useful robot demonstration must connect perception to action: camera frames, joint states, gripper commands, motion trajectories, task state, contact events, timing, and the final outcome all need to be synchronized well enough for a learning system to extract a reusable policy.

That creates a difficult scaling problem. A model can ingest billions of tokens without a person physically reproducing each sentence, but a manipulation dataset often requires a person, a robot, a controlled environment, calibrated sensors, and a repeatable task. Even when collection is distributed globally, the data must still be normalized across operators, hardware configurations, cameras, task definitions, and environmental conditions.

XDOF’s business is designed around this bottleneck. It combines teleoperation systems with operational data collection and annotation infrastructure so customers do not need to build every component internally. Human operators can remotely control robotic arms to generate demonstrations, while other collection methods can capture human movement and egocentric interaction data that may later be mapped into robot-relevant representations.

The value of such a supplier depends on more than raw volume. Robot data can become harmful when demonstrations contain idle motion, recoveries, inconsistent strategies, sensor drift, or action segments that technically end in success but teach an inefficient trajectory. XDOF has publicly described research in which adding more demonstrations degraded a folding policy because useful and unproductive moments were mixed together, illustrating why quality filtering and temporal segmentation can be as important as dataset size.

For customers developing general-purpose robots, outsourced data infrastructure can shorten the time between a model hypothesis and a new training run. Instead of building a new collection fleet, recruiting operators, designing annotation workflows, and validating every dataset internally, a laboratory can purchase a production pipeline and concentrate engineering resources on model architecture, policy learning, evaluation, and deployment.

··········

GELLO, ABC-130K, AND XDOF’S DATA PIPELINE.

The company’s technical identity combines low-cost teleoperation, large-scale bimanual datasets, and production data operations.

XDOF grew out of robotics research at UC Berkeley. Its co-founders helped develop GELLO, a low-cost teleoperation system designed to let a human operator control a robotic arm and generate demonstrations. The approach reduces the gap between human intent and robot action by giving researchers a relatively direct way to create trajectories that can be used for imitation learning and related policy-training methods.

The company has also released ABC-130K, which it describes as the largest open-source teleoperation dataset of its kind. The dataset contains more than 130,000 episodes covering about 200 complex manipulation tasks on a low-cost bimanual setup and is available under the Apache 2.0 license. Its strategic value is not only the number of episodes but the demonstration that XDOF can operate a collection pipeline at a scale materially larger than the small laboratory datasets historically common in manipulation research.

........

Data layer	XDOF approach	Why it affects model performance
Teleoperation	Human-controlled robot demonstrations using systems derived from GELLO	Produces action trajectories tied directly to robot hardware
Bimanual manipulation	Large collections on dual-arm rigs	Captures coordination problems that single-arm datasets cannot represent
Egocentric collection	Human operators wearing sensors while performing physical tasks	Expands coverage beyond tasks already instrumented on robots
Annotation and filtering	Operational tooling for structuring and cleaning demonstrations	Reduces noise, dead time, inconsistent behavior, and low-value segments
Open dataset	ABC-130K with 130K+ episodes and about 200 tasks	Provides a public benchmark and demonstrates collection capacity
Production delivery	Customer-specific data pipelines for robotics and AI labs	Connects collection volume to recurring commercial demand

........

The technical challenge is that robot datasets are not perfectly portable between embodiments. Different arms, grippers, cameras, control frequencies, coordinate systems, and force capabilities create distribution shifts. A defensible data supplier therefore needs transformation, metadata, calibration, quality-control, and task-design capabilities that make demonstrations useful across changing customer hardware rather than simply accumulating videos.

XDOF’s public positioning emphasizes full-stack infrastructure from hardware and operations through policy training. That breadth is important because the value chain is tightly coupled: poor teleoperation hardware creates poor trajectories; poor task design creates narrow coverage; poor annotation hides failure modes; and poor delivery formats increase the engineering burden for the customer.

··········

WHAT A $1.2 BILLION VALUATION WOULD SAY ABOUT THE PHYSICAL AI RACE.

The financing would price robot data as a strategic bottleneck, but the long-term economics still depend on whether the market remains externally supplied.

If XDOF closes a Series B near the reported valuation, the transaction would indicate that venture investors are assigning substantial value to the infrastructure surrounding robot foundation models, not only to robot manufacturers and model developers. The investment thesis is that physical AI will require a sustained stream of new real-world demonstrations as models expand into more tasks, environments, hardware platforms, and failure conditions.

That thesis can support a large independent supplier if data demand grows faster than individual laboratories can internalize it. A customer may prefer to own its model weights, architecture, and evaluation stack while outsourcing repetitive collection operations in the same way frontier model companies historically relied on external providers for labeling, preference data, red teaming, and specialized human feedback.

The risks are equally concrete. Major AI laboratories can build internal robotics-data organizations if the datasets become strategically sensitive. Synthetic data and simulation may reduce the amount of expensive real-world collection needed for some tasks. Hardware fragmentation can limit reuse across embodiments. Labor-intensive collection can pressure gross margins, and rapid customer concentration can make annualized revenue less durable than a diversified software subscription base.

The strongest version of XDOF’s business therefore is not a simple marketplace for human demonstrations. It is an integrated data-production system with proprietary operational knowledge, collection hardware, task design, quality-control methods, transformation pipelines, and enough throughput to deliver datasets faster and more reliably than customers can build internally.

A completed Series B at roughly $1.2 billion would not prove that this model has already won. It would show that the market is beginning to price high-quality physical-interaction data as one of the scarce inputs in the transition from digital AI systems to machines that can manipulate the real world. XDOF’s next stage will be measured by whether its reported revenue growth converts into repeatable margins, larger customer deployments, broader embodiment coverage, and a data advantage that persists as robotics models and simulation systems improve.

·····

FOLLOW US FOR MORE.

·····

·····

DATA STUDIOS

·····

[datastudios.org]