TechFlow Logo
Login/ Sign up
ETH Gas
Gwei
Fear
gas
Insights into the Robotics Sector After Unitree's IPO: Hardware and Large Models Are Not Scarce, Training Data Is the Most Valuable Segment

Insights into the Robotics Sector After Unitree's IPO: Hardware and Large Models Are Not Scarce, Training Data Is the Most Valuable Segment

2026.08.19
Share
The most valuable thing in the robotics revolution is not the robots, but the data that feeds them.

Author: Shoal Research

Compiled by: TechFlow

TechFlow Editor's Note: The robotics track is heating up, but what's truly scarce might not be hardware or large models, but high-quality training data that enables robots to learn how to work. This article breaks down why the data layer could become the highest-value segment in the robotics industry, and which types of companies have the best chance to capture this cake.

The rapid development of AI in recent years has brought many lessons, the most important of which is: computer intelligence is first and foremost a function of its data. Today's frontier models and their capabilities exist because a few laboratories built massive training data engines to feed them. This engine was built on the basis of vast amounts of free text on the internet. Training data is now also a key bottleneck in the robotics field, the difference being that robotics training data requires humans to physically execute and record tasks. Robotics developers currently do not have a trainable corpus of training data, and precisely because this training data is harder to obtain, it is more valuable. In this article, we will explain why we believe the data layer is the most valuable opportunity in the robotics field today, and which type of enterprise is best positioned to seize this opportunity.

Intelligence and Data

For seventy years since the advent of computers, making a machine truly complete a task has often itself been another task. Someone must first write out every step in a sequence the machine can follow, which perfectly fits jobs originally built on some fixed, precise process. People have been doing mathematical operations, bookkeeping, and balancing accounts for centuries, but computers can execute the same set of processes faster. Automation reshaped job functions and entire industries, but largely remained limited to single tasks, and more importantly, everything a machine does must first be written out by humans.

Today, machines can write coherent prose, answer specific and open-ended questions, and turn plain language descriptions into code and deployable applications. This fundamental shift comes from changing the type of instructions given to machines and the way instructions are given. Although tasks still start with instructions, the machine's capabilities after receiving instructions have evolved significantly. More importantly, the instructions themselves have become much more intuitive and simple, especially as they can be simplified from detailed specifications to simple sentences.

Importantly, changing a machine's capabilities first requires changing their input. Machines are no longer given precise instructions on how to perform specific tasks, but are given examples of many different tasks being performed, and learn by observing these examples. Models are not taught how to write a sentence, but learn the task themselves after processing enough sentences. Models are not taught how to write code, but learn to generate code themselves after processing enough runnable code. Models are not taught how to translate two specific languages, but learn translation themselves after processing enough translated text.

To better illustrate this distinction, think of a calculator and ChatGPT. A calculator can perform large complex operations, provided these operations involve numbers and do not exceed its defined range, but it is useless if asked to write an article. Language models can be used as calculators, while also able to write articles, search the web, generate runnable code, and do much more. The key difference lies in training data. The pattern for letting a machine answer a question it has never seen before can only be learned from real examples, so the more examples an AI model is trained on, the more world structure it captures, and the more it can do. The same mechanism is about to happen with robots in the physical world, so it helps to first understand how it unfolded in the software field.

Paradigm Moment

In 2020, OpenAI released GPT-3, and a few years later ChatGPT was launched. For many people, this was the first application that truly seemed like reasoning. ChatGPT could answer various questions, write articles, translate languages, and generate runnable code, often requiring only a short prompt. These capabilities resonated strongly: ChatGPT reached one million users in five days, and within two months of launch had about 100 million monthly active users, setting the fastest adoption record for a consumer-grade application at the time.

Although ChatGPT initially ran on GPT-3.5, the breakthrough can actually be traced back to GPT-3, and GPT-3 was possible because its training data scale was unprecedented. GPT-3 is the third generation of the same underlying GPT model, with each generation's parameter count increasing by approximately an order of magnitude over the previous one. For reference, parameters are the building blocks supporting the model's core function of identifying and reproducing patterns, so the more parameters, the stronger this capability. GPT-1 had about 117 million parameters, GPT-2 about 1.5 billion, and GPT-3 175 billion.

Every parameter leap attempted to remove the tedious but previously considered essential links in AI model training. GPT-1 proved that language capabilities learned from raw text could be transferred to specific tasks, and that the specific task data required was far less than starting from scratch. The model was first pre-trained on a large amount of unlabeled text to predict the next word, then fine-tuned for each specific task with a smaller labeled dataset.

Fine-tuning is the tricky part, because each new task still required its own round of fine-tuning and its own labeled examples, so GPT-2 tested whether fine-tuning was really necessary. The idea was that a language model trained on sufficiently diverse text itself is a generalist, capable of handling tasks never fine-tuned for it, because the training text was already full of these tasks. Testing this capability meant expanding the input scale of the GPT model, so OpenAI increased the parameter count by more than ten times to 1.5 billion, and trained on a larger, more diverse web corpus. The largest version achieved state-of-the-art results at the time on seven out of eight language benchmark tests without any fine-tuning. Simply expanding the model scale eliminated a step that was supposed to be essential.

GPT-3 pushed the same bet further: it trained 175 billion parameters on about 300 billion text tokens (tokens refer to parts of a word), these texts taken from filtered Common Crawl, WebText2, two book corpora, and English Wikipedia according to weighted proportions. As scale changed so dramatically, capabilities improved accordingly. GPT-3 could perform new tasks by simply putting a few examples directly in the prompt, completely without fine-tuning.

Although today's models' capabilities far exceed GPT-3, it remains a key node. It laid the foundation for GPT-3.5 and ChatGPT, and sparked the AI boom that reshaped the industry thereafter.

Intelligence Follows Data

Frontier laboratories are companies and research institutions that build state-of-the-art models from scratch, including OpenAI and Anthropic. They have put these models into production and delivered them into the hands of hundreds of millions of users worldwide. Thus, they attracted hundreds of billions of dollars in investment and are expected to become one of the most valuable companies globally. Their success is often attributed to recruiting researchers and possessing massive computing power. But none of this would be possible without the massive amount of data used to train the models.

As GPT-3 proved, building anything close to the capabilities of frontier AI models requires massive training data. It needs to be pointed out here that not all training data is the same. Raw text scraped from the internet is free and abundant. It can teach the model how language works, how words fit together, how arguments are organized, and where which facts appear. As the scale of scraped corpora expands, models trained on them become more capable. They eventually begin to generalize to tasks no one specifically built for them. GPT-3 could translate between languages, organize scrambled words, and do basic math, although it was never specifically trained for any of these. In fact, GPT-3 beat Facebook's leading cross-language model XLM on French-to-English translation. It was never trained on paired French-English text, relying only on a few examples in the prompt.

However, there is a reason models are not trained only on scraped text. Such models often give answers that sound convincing, regardless of whether they are correct or helpful. Improving accuracy requires human judgment on which of the model's answers are truly good. A person will see the prompt and two answers generated by the model, then mark which is better, or read the answer and correct errors, or write from scratch the answer the model should have given. Each judgment becomes a training sample, and the model is tuned to prefer outputs preferred by humans. This process is called Reinforcement Learning from Human Feedback (RLHF). These judgments are training data different from scraped text, also more valuable, and more expensive. Because each sample must be produced by a person, not collected from the web. The harder the question, the more qualified the evaluator must be. Answers about tax law can only be scored by someone who knows tax law. Therefore suppliers recruit doctors, lawyers, and engineers to evaluate outputs in their respective fields. Producing such data at frontier scale is difficult, slow, and expensive. Optimizing a model requires hundreds of people for months, costing up to tens of millions of dollars. Frontier laboratories spend billions of dollars on this annually.

Models trained on these complete work samples are better at planning tasks, executing step-by-step, and self-correcting in the process by learning from previously accumulated experience. We call these agents. Agents have become widespread productivity and learning tools. For frontier laboratories, they have become the fastest-growing product line to date. ChatGPT now serves over 900 million users weekly. OpenAI's first quarter revenue was about $5.7 billion, with a 2026 revenue run rate reaching $40 billion. Its programming agent Codex surpassed 5 million weekly active users in June 2026, more than six times its independent desktop app launch in February. Claude Code is an agent performing software engineering work, reaching an annualized revenue of $1 billion within six months of launch, exceeding $2.5 billion by February 2026. Anthropic's annualized revenue reached $65 billion in the same period, compared to about $1 billion in early 2025. In a survey by Anthropic and research firm Material of more than 500 US technical leaders, 86% said they deployed programming agents to write production code, and 42% trust them to lead development work under human supervision.

The initial rapid growth of leading AI applications stemmed from breakthroughs brought by their application to new fields. Every task agents can currently execute happens on computers. Although a large part of today's world runs on computers, the world and most of its labor is physical.

The Next Foundation of AI is the Physical World

In the first half of 2026, global robotics startups raised $18.8 billion. In comparison, the full year 2025 was $15 billion, exceeding the full-year record of $14.1 billion set in 2021. SoftBank agreed to acquire ABB's robotics division for $5.4 billion, stating physical AI is its next frontier. Meta acquired Assured Robot Intelligence to train its own physical AI foundation models. Unitree was recently approved for an IPO in Shanghai. Deployment has unfolded alongside deals. For example, BMW began using humanoid robots for pilot production in German factories. Japan Airlines uses humanoid robots for baggage handling at Tokyo Haneda Airport.

However, robots have been working in factories since 1961. At that time Unimate robotic arms were installed on General Motors die-casting production lines in Ewing, New Jersey. Between 1966 and 1972, Stanford Research Institute created Shakey, the first mobile robot capable of reasoning about its own actions, rather than executing fixed sequences. Waseda University completed WABOT-1 in 1972, the first full-sized humanoid robot.

So, why have funding and attention suddenly risen now?

Robots now have something to think with. Robots must decide what to do before taking action. Making decisions requires a large amount of detail about how the physical world works. Language models absorbed this information from internet text, so robots can now use it as a brain. But language models process software and output text, while robots process the physical world, and their output is action. Vision-Language-Action (VLA) models emerged as a solution bridging the physical world and the software world. It receives what the robot camera sees and instructions received, returning actual joint movements. In practice, this means robotics developers no longer need to write custom code for each unit, letting them execute only one task at a time. They can be closer to operating an AI application, giving tasks in simple natural language, and then calculating actions in real time.

Robot bodies have become cheaper. Robots observe through cameras and LiDAR. LiDAR emits laser pulses and times how long they take to return, thereby building a map of the surrounding environment. Ten years ago, LiDAR units were largely considered too expensive. But mobile phone manufacturers and car manufacturers subsequently began buying the same sensors in bulk. This drove down prices for everyone including robotics companies. The cost of motors and joints also declined respectively, because Chinese manufacturers entered the component field previously produced in small batches by a few Japanese and German companies. Humanoid robots manufactured with Chinese components now cost about $35,000. Unitree's G1 humanoid robot sells for $13,500.

Stronger intelligence and cheaper hardware are crucial for robots and their capability progress. But like AI models, everything first depends on underlying training data. To understand why this data is so valuable, let's first review what a robot actually is.

Understanding Robots

By definition, a robot is an autonomous machine capable of perceiving the environment, thinking and making decisions, and executing actions in the physical world. Simply put, it is a computer with a body.

Almost all robots are composed of the same components, each corresponding to a biological analogy of the human body. The skeleton is usually metal, composed of rigid segments called links. Links are connected by joints, which facilitate movement. Sensors constitute the machine's sensory organs, including proprioceptive sensors, used to measure the position of the robot's own links; and exteroceptive sensors, used to measure the environment the robot is in. Actuators are most commonly electric motors, acting as the robot's muscles, applying force to links through transmission devices (such as gears), to increase force, while proportionally reducing speed, allowing limbs to produce slow and powerful movements. Copper wires power the actuators, and signals are transmitted back through the data bus. The data bus is a shared line, each joint and sensor connects to it, every message marked with the address of the sending component. Power either comes from the wall via cable, like most factory arms, or from batteries carried by the robot.

Feedback loops connect these different physical components, enabling the robot to actually perform physical operations, record the process, and improve learning. More accurately, robot software receives information from sensors, processes information to decide what to do next, and sends instructions to actuators. Actuators apply force to objects in front, sensors subsequently read the results, and the loop closes. Roboticists call this the perception-planning-action framework: figure out what is happening, decide what to do, make it happen, and learn from each observation.

Perception and action are hardware-based functions, so how well they run depends on available components and their cost. Planning is software-based, largely depending on the availability and quality of training data.

Understanding Robotics Training Data

Robots learn how to complete tasks from videos of task execution. Each video pairs two synchronized streams: 1) what cameras and sensors record; 2) what motor commands were issued at the same moment. With enough such pairs, models can begin to correlate, from seeing tasks being completed to completing them themselves.

Programming this process is fine in single fixed tasks, because movements can be written once and repeated. Think of a robot arm stamping on a factory line, it encounters the same panel at the same position at the same moment. But any task with some variation and uncertainty (most things in life are often so) makes this more complex. Especially in real time there are many subconscious decisions, most can only be learned through experience and intuition. No one wants to think about how much force to pinch a shirt, or how much force to hold a cup to pick it up. Therefore, rather than trying to write this subconscious knowledge into specifications, robotics developers capture it by recording real-time execution of tasks.

The robot's model (its brain) needs to answer a simple direct question: based on what cameras and sensors show at this moment, what should the motors do next? Camera images and readings from robot joint and finger sensors are shown together with the exact motor commands produced by skilled operation at each moment. There is a large amount of video content on the internet, but most lack these two specific requirements needed for robot learning. Today there are several different methods to produce and collect robotics training data.

Teleoperation: Humans directly control the robot, executing tasks through the robot's body, using motion tracking devices or controllers that can mirror the robot's arms. The system records what the robot sees, what joint sensors measure, and what motor commands the operator's actions produced.

Motion Capture: Humans wear tracking markers or sensor suits and gloves to perform tasks, generating precise 3D trajectories of body and hands, then remapped to robot joints.

First-person View Video: Head-mounted cameras record humans doing tasks with their own hands, software subsequently processes the footage, estimating hand and finger positions frame by frame, and converting them into equivalent robot actions.

Synthetic Data: Simulating robots attempting tasks in simulated environments, each attempt recorded as a training sample.

It needs to be pointed out that models trained only on demonstrations inevitably encounter ceilings, because most demonstrations only show tasks proceeding smoothly, models never see errors, and never learn what to do when encountering errors. A common solution is to turn data production itself into a loop, formally called DAgger (Dataset Aggregation) in robot academia, which is a type of interactive imitation learning. Humans demonstrate to train the initial model, then the robot attempts the task itself. When it makes mistakes, human operators take over via teleoperation, correct the errors, and hand control back to the robot, while corrected content is incorporated into the robot's training data for future reference. Robotics teams usually aim to eventually let robots practice on their own, their attempts scored according to a set of defined rules, to reinforce successful behaviors. However in practice, currently most work remains in early stages, with humans closely involved in this loop.

The product of the training loop is what we can call the robot "brain": a trained model, capable of receiving reports from cameras and joint sensors at a certain moment, and returning subsequent motor instructions. Today's models are sold in three levels, each level corresponding to different degrees of robot autonomy.

The first level is skill models, also the narrowest models. It executes a single task through running the training loop, with slight environmental variations. It sounds like factory line robots designed for single repetitive actions. But skill models learn to execute their single task under different conditions. For example executing tasks in different positions or different lighting. Second is autonomous models. It chains different skill models into complete work, and handles decisions between each step, calling each individual skill model as needed. The third level models are today's widely recognized "frontier", namely Vision-Language-Action models, abbreviated as VLA. It is a single training system, combining vision, understanding ordinary language instructions and taking action. This design draws from language models, indeed using them in several leading implementations. A large vision-language model, the same type of system behind chatbots, handles slow high-level reasoning for understanding scenes and planning actions. Another faster component translates plans into continuous motion instructions multiple times per second.

Because VLA inherits general knowledge from internet-scale pre-training, it arrives knowing things no one taught it. This changes the way new capabilities are added. Skill models know nothing outside their task. So new tasks mean starting from scratch, one demonstration after another. VLA can receive tasks in ordinary language. It already has some grasp of objects and goals. So its demonstrations only need to provide physical techniques, without needing to provide techniques, objects and task concepts all at once.

Data Bottleneck

Every method of producing this kind of data is subject to the same fundamental constraints. For data to exist, a physical execution must first happen once. A demonstration recording time is as long as the task execution time. Occupying one robot, work site and operators throughout. When language models need more data, AI laboratories turn to more web scraping and hiring more labelers. When robots need more data, solutions require more robots, more operators, more warehouse labor hours.

Models can indeed transfer some general capabilities across tasks. But reliably executing new tasks still requires demonstrations of that specific task. Each task has variations, such as different objects, layouts, lighting etc. They all need their own coverage. Therefore it is hard to imagine demonstration demand slowing down in the short term. Models may eventually learn new tasks from few examples like language models. But as GPT-3 and its successors showed, capability leaps can only come from exponential growth in training data and inputs.

The Big Opportunity for Robots Lies in the Data Layer

Robotics startups raised more in the first half of 2026 than in all of 2025. 2025 itself was record-breaking. The gap between enthusiasm and reality, the gap between robot demand and current actual capabilities remains large. Simply put, robot demand is huge, but today's robots are not good enough. Robotics developer Unitree shipped over 5,500 humanoid robots in 2025. This is the highest shipment volume globally disclosed by the company. Iconic industry deployments like BMW, Mercedes-Benz, Amazon are all pilots, involving only dozens of machines. Industrial buyers need machines available 95-99% of the time. Current humanoid robots need charging or human intervention after working 30-90 minutes. Robot hands cannot reliably handle varying objects. Their brains, although frontier laboratory models have improved greatly in past years, are still confused by differences in physical reality and unexpected complexity.

The first three generations of GPT models proved that the most effective way to narrow capability gaps is to give models more high-quality data. The limitation here is that data needed by language models is already largely sufficient. This thanks to humans recording and writing history for several centuries. Data needed by robots is multi-modal, time-synchronized and physically collected. This means it cannot be scraped like text, must be manufactured by humans operating robots and recording equipment. Labeling also cannot be parallelized like text labeling. Because judging whether a plate is placed stably or unstably requires physical intuition, not pattern matching. So labeling teams cannot simply scale to thousands of remote workers.

All robot makers, whether robotics companies or model laboratories, need this data. But there are no shortcuts to obtaining it. Content in public datasets rarely transfers to specific robots. Internal production data means operating warehouses, fleets, operators, and bearing all related costs. Similar situations have occurred in the software field. Frontier laboratories need human judgments and expert-written examples not on the internet. And suppliers producing this data (such as Scale, Surge, Mercor) grew into a category of about $8.5 billion revenue.

The Path Forward Lies in Vertical Integration

A useful comparison is chip manufacturing. Designing chips and manufacturing chips are two businesses. Design is intellectual property. And factories etching designs onto silicon wafers cost billions of dollars, only breaking even when near full production. So most chip companies no longer own factories. Nvidia, Qualcomm and Apple all design chips, but none manufacture themselves, handing manufacturing to TSMC. TSMC does not design own products, now manufacturing almost all the world's most advanced chips.

The robotics data layer is very valuable, but not easy to occupy, let alone succeed. Demonstrations recorded on a specific robot body and their associated datasets cannot transfer to another robot. This means data suppliers must rebuild products for each client. As mentioned earlier, what can be taught by demonstrations alone is also limited. Because demonstrations only show tasks executed correctly. Models still need to learn what to do when errors occur. Operators often need to intervene during training. Clients looking for robotics training data have little interest in raw data. They care more about the ability to continuously produce data when robots change. This requires 1) robots for recording, 2) real work these robots need to do, 3) pipelines feeding each failure back into the next model. Now there is a unique opportunity. By owning these three components, an enterprise can own the robotics training data layer.

Case Study: Eastworlds

A team actively seizing this emerging opportunity is Eastworlds. It is the new robotics division of Virtuals Protocol. In fact, Eastworlds is a neodeployment robotics laboratory and accelerator. It serves robotics and embodied AI startups. The team operates in Kuala Lumpur, possessing the most Unitree G1 robots in Southeast Asia (30+). The team established research collaborations with universities like Carnegie Mellon University and Nanyang Technological University.

Eastworlds' business model revolves around robotics data collection and new deployment. Data collection covers teleoperation and gripper data, video, proprioception, motion capture, dataset auditing and quality control, and policy improvement datasets, including human corrections and associated recovery data. Besides data collection, Eastworlds also focuses on new deployment, which is putting robots into commercial scenarios where robots have never been used. Machines perform paid work under teleoperation, while the work itself generates records. Clients purchase labor, while generating data for Eastworlds.

Every deployment produces data, usually unstructured, and full of edge cases because it comes from real operations. This data is collected in the form of egocentric teleoperation data, multi-modal sensory input and task execution feedback, and flows into what the company calls the embodied AI data lake. In the short term, this data will retrain models running the fleet, letting each robot fail less on current work. In the long term, accumulated datasets will become training material for next-generation VLA and world action models, driving the G1 fleet from teleoperation towards full autonomy.

Eastworlds also operates an incubator, providing funding for teams passing audit, and letting them use the robot fleet, facilities and teleoperation stack within one month. Each team pursues their own commercial use cases. According to Teng, there are currently seven teams resident, covering entertainment, oil & gas, groceries and hotel management, planning to accommodate over twenty teams. Each team runs their own deployment on the shared fleet, exposing their own failures and repair solutions. Learning outcomes and data all flow back to the data lake, applying the same mechanisms driving the Virtuals agent ecosystem to robots.

Eastworlds integrates Virtuals Protocol's on-chain infrastructure to handle economic issues of data operations. Eastworlds adopts a distributed business model, involving thousands of hours of recording and teleoperation, completed by many people in many places, quality needs to be verified before payment. Hiring such labor requires establishing compensation and supervision systems in each relevant geographic region. Another method is using incentives, i.e., publishing tasks and rewards, contributors can join from anywhere, payment arrives quickly after verifying delivery. Enabling this mechanism at scale requires programmable, global, low-cost payment rails. Virtuals already runs such rails for software AI agents on Base and Solana, used for fundraising, trading and programmatic settlement. Eastworlds applies the same design to robotics training data, using token incentives to recruit and pay distributed labor.

The market for this data is already paying. According to Teng, buyers pay about $3 to $10 for one hour of egocentric recording, and $50 to $100 for one hour of teleoperation. These rates are higher than typical hourly wages in most parts of Asia. Traditionally, token incentive activities tend to collapse after new participants stop entering, because no one outside the system is buying anything. But incentives here are backed by paying clients, i.e., laboratories buying data. Therefore as long as data demand exists, the fund pool can be sustained.

Comprehensively, Eastworlds' roadmap is roughly as follows: Today sell data and new deployments, because buyers are already paying for both. Utilize these paid deployments to collect content phased acquisition cannot produce, especially correction data. Then utilize the growing data lake to train models, starting from single-task models, gradually moving towards more general models. Each of these models sells for far more than raw data. Eastworlds operates at the data layer, but its moat is everything needed to continuously produce data, including G1 fleet, operators, incubated teams, on-site deployments and underlying payment rails.

Risks and Important Considerations

Of course, there are some risks and important considerations to remember regarding scaling the robotics data layer. The deepest objection to this category is that teleoperation may ultimately not bring robot autonomy. A robot needing an operator to complete a hotel shift anyway has people involved, just adding a robot and its related costs. Teleoperation suppliers and robot manufacturers believe this is only a temporary phase when activity variation is small and failures repeat. Models will indeed continue improving until operators completely stop intervening. Tasks that initially made robots difficult, i.e., tasks where objects, layouts and failure modes constantly change, are precisely tasks operators today cannot completely leave. These tasks also produce the most valuable data. This is why suppliers can remain profitable in a world where full autonomy never truly arrives.

Simulation threatens demand for recorded demonstrations, and recorded demonstrations are exactly what data suppliers sell. Operators manually recording produce about 5 to 50 clips per hour, while simulators can run millions overnight at near zero cost. What simulation still does poorly is contact, because once two objects touch, simulated physics and real physics diverge. This is why robots capable of grasping items cannot be trained on simulation data alone. Teams are narrowing this gap by randomizing conditions in simulators, such as lighting, textures, friction, so models do not overfit to one kind of rendering. They also adjust simulator physics settings until simulated arm movements match real arms.

Prices for robotics training data itself have also declined. According to SVRC's "2026 State of the Robotics Industry", one hour of high-quality teleoperation data was about $340 in early 2024, dropped to $136 in Q4 2025, and to $118 by March 2026. Decline driven by two factors. First, master-slave devices dropped below $2,000, setups where operators move small replica arms and robots follow no longer require custom hardware. Second, software like DROID and LeRobot can now format operators' raw recordings directly into training-ready datasets, without manual cleaning. Both are still improving. Therefore, suppliers selling recorded hours as products are entering a market where prices fell about 66% in two years.

Competing robotics data suppliers moved towards vertical integration almost simultaneously with Eastworlds. Adamo sells hosted teleoperation, with control software below 40 milliseconds latency, letting robotics companies avoid building operations themselves. XDOF founded in October 2024, expanded from simply selling recorded data to cleaning, tools and labeling. Its founders' judgment was that only selling recorded data is a dead end. Physical Intelligence, Lux, iMerit and DataX Power all sell some version of hosted collection plus配套 tools. DataX Power conducts projects in Southeast Asia, same operational region as Eastworlds. However, none of them operate their own paid commercial deployments. Correction data comes from commercial deployments, this is the part of the feedback loop that cannot be replicated by hiring more operators for warehouses. The collection layer is indeed crowded, the deployment layer is thin.

Robot manufacturers are naturally a harder problem. Because several already produce training data themselves. Figure operates its own teleoperation fleet in BMW deployments. Tesla Optimus project is built on the same loop: deploy robots, record failures and retrain. These operators already own hardware, deployments and client relationships. This means they accumulate data faster than anyone buying data.

But the data they accumulate is just data of single robots doing one company's work. A supplier serving twenty clients in hotels, farms and warehouses. The diversity of data it collects, no single manufacturer fleet can produce. Figure and Tesla are also the two best-funded robotics teams globally. Dozens of robot manufacturers below them cannot fund operations of such scale. This is exactly the market independent suppliers can actually sell to.

Paying contributors with tokens avoids establishing wages and compliance in every country where someone records one hour of video. But this ties contributor income to token price, not to work done. Therefore hiring is hardest during drawdowns, and that is precisely when operations most need stable output. Legal handling of paying workers this way is also undecided in most relevant jurisdictions. Eastworlds' answer is that client revenue funds incentives, not new token buyers. This is of course the correct structure, because it means the pool does not rely on new participants arriving. As long as clients continue, this structure holds. A contributor base priced in volatile assets has a second failure mode, which wage-based competitors do not have.

Eastworlds is building a pickaxe business for the emerging robotics field, specifically around the training data layer. The software version has already gone the full distance: frontier laboratories scrambling to build state-of-the-art models, spending hundreds of billions. Companies selling them training data Scale, Surge and Mercor, grew into billions of dollars enterprises, making money regardless of which client wins.

That boom has spread from software to the physical world. Robots have brains borrowed from language models, and cheap bodies mass-produced mainly by Chinese manufacturers. Every serious laboratory and robot manufacturer is chasing the same kind of machine capable of autonomously completing physical work. What none of them can skip is physical work examples available for learning. These examples cannot be scraped, mostly must be produced by humans operating robots, one recording session at a time. Aggregated public datasets help models start. But buildings where robot work occurs, hardware running, and their specific failure modes must be re-recorded. So demand updates with each deployment, not satisfied once.

Suppliers selling only raw demonstrations are easily replaced and squeezed. Because recordings transfer poorly between robots, and few clients often account for most revenue. A supplier simultaneously owning robot fleets, hiring operators, managing deployments. It brings every on-site failure back to the next model. It becomes an enterprise clients find hard to leave or rebuild. It does what originally required five suppliers. This is exactly what Eastworlds is building. Its Unitree fleet now conducts paid pilot work in hotels and malls. Operators intervene when autonomy fails. Both ultimately become training data, provided in the form of demonstrations and corrections, flowing into the data lake. Although data is today's robotics technology biggest bottleneck, surrounding infrastructure remains highly fragmented. Now there is a unique opportunity for enterprises to vertically integrate and own as much of the robotics data layer as possible. Eastworlds is well-positioned to seize this opportunity, but only time will prove whether it succeeds.