The primary data set the last few years of artificial intelligence was largely digital. Companies collected text, images, video and code to train models that operated on screens. The next iteration is moving into the physical world, where machines can learn from the work people perform and the environments they perform it in.
Robo Robotics entered that market last week with Robo-T, a two-armed mobile robot built to work at stations designed for people.
Robo’s thesis: deploy inexpensive, multipurpose robots before they are fully autonomous, use human operators to keep them working and capture useful training data from selected demonstrations and interventions. If the loop works, better models will gradually reduce the human oversight each robot requires, lowering the cost of physical labor across the fleet.
I interviewed co-founders Don Morton and Kyle Noble about how the company started and their plans for building out the Robo-T and their future platforms.
From Selling the Data to Selling a Robot
That plan grew out of Morton and Noble’s path from software to model training and, eventually, physical data.
Noble joined baby-food company Yumi as its first engineer in 2017 and brought Morton on as its second the following year. They left in 2022 to build Atlas, beginning with a browser and later developing an education and AI product.
Model tooling was still immature, so the team built its own search index, fine-tuned models and trained models internally. “We were doing all of the AI stuff that you don’t have to do today,” they told Trove, “It’s all solved now.”
That experience led them into the training-data market. They began collecting data for video-model companies and AI labs, but conversations with prospective buyers exposed a sharper need. “What we realized as we talked to labs and companies was that the robotics companies need data the most,” the team said.
Building the arms changed how the team saw the opportunity. Once they had a working version, they started talking with customers about how it could be deployed. Most wanted something that could roll up to an existing line without requiring them to redesign their processes.
That requirement shaped Robo-T. The team placed the arms on a T-shaped body and added a movable camera head, building toward the complete robot one design decision at a time. What began as an effort to understand robotics companies as customers became a robotics company itself.
Priced in Labor
Robo offers two ways to use the machine.
In the first, a customer hires Robo for labor. Robo deploys and manages their platform, using whatever combination of autonomy and remote operation is required to complete the task. The customer buys the labor, irregardless that it’s performed by a machine.
In the second, an operator buys a fleet of Robo-Ts with Robo providing the hardware and operating infrastructure, while the fleet owner runs the robots themselves.
“The customer should not have to care whether there’s a human operating or whether it’s autonomous,” the team said. “The output is just the labor.”
Today, humans remain essential to Robo-T’s output. A remote operator wears a headset and uses Quest controllers to demonstrate a task or take control when the model encounters something it cannot handle. The person keeps production moving through the failures that would otherwise stop the line.
This gives Robo a wider initial task range than its autonomous models could support alone. A person can attempt anything the robot’s body can physically perform.
Robo is candid that teleoperation isn’t unique, other deployment companies can put a human behind a robot. The differentiator is whether the economics of that human in the loop work.
Every hour a person spends driving one machine is an expense attached to that robot-hour. The economics improve only when the person shifts from continuous driver to occasional supervisor.
Robo’s Model T launch essay models a $30,000 robot amortized over two years at 240 operating hours per month, plus $1 per hour in other expenses. That creates a fixed cost of roughly $6.21 per robot-hour before teleoperation. With operators earning $25 an hour, the modeled total is $81 when three operators are assigned to each robot, $8.71 when one operator supervises 10 robots and $6.83 when one supervises 40. The exercise shows how increasing the number of robots per operator can drastically change deployment economics.
Price buys room to learn
Low-cost hardware does more than reduce the amount of the customer’s bill, Robo believes it changes the customer’s tolerance for an early product.
During the interview, the team used a $40,000 robot as an example. At that price, every period of downtime invites the same reaction: what was all that money for?
A cheaper machine gives the company more room to learn in public. Repairs and replacement parts cost less. Customers may accept early imperfections if the work remains economical and continues to improve. More customers can deploy robots, producing more chances to find failure cases and improve the system.
“The lower your price, the more accessible it is for people who have a higher threshold for working with something that isn’t perfect,” the team said.
Henry Ford’s achievement wasn’t simply producing a car that worked. It was imposing a price constraint that forced the company to improve manufacturing while putting enough cars into the world to create a larger economy around them.
Robo is attempting the same sequence with robot labor.
The company says demand following the launch exceeded what it could fulfill. They declined to disclose customer names, deployment counts or completed operating results, although they spoke at a high-level of upcoming work in healthcare, food production, and textile manufacturing.
Who owns what the robot learns?
The data loop introduces a problem that better models can’t currently solve.
A hospital or manufacturer may want robotic labor without allowing video from its facility to train a system later used by a competitor. A fleet operator may want control over the data produced by customers it sourced itself.
Robo has not publicly settled those rights yet. In the interview, Morton and Noble described a possible economic bargain rather than a finished framework, something they are actively developing.
In Robo’s proposed bargain, customers that contribute data to a shared autonomy model would benefit when the model improves. Their robots would require less teleoperation, lowering the hourly cost and improving deployment margins. Robo is also considering revenue sharing if the data can be packaged and sold to another buyer.
The team offered a hypothetical example in which an hour of video had $10 of resale value and participants in the chain received a share.
Robo’s flywheel is built around a physical-world scaling law: the more work its robots perform, the more demonstrations and useful interventions it collects, and the more capable its models become. Within an industry, data from one job trains similar tasks across the fleet. Across warehouses, factories and hospitals, recurring patterns in movement, perception and recovery strengthen the models beneath every deployment.
Each deployment expands the training set for the next. As the fleet grows, skills compound across tasks and industries, reducing human intervention and driving down the cost of robot labor.
They’ll need to find a framework that is mutually beneficial to their partners but also enables them to compound the work they’re completing and data they’re collecting to generate better models and robots.
The Manufacturing Bottleneck
Besides the data collection issue, Robo faces another constraint common to American robotics companies: once a product finds demand, scaling becomes as much a manufacturing problem as an AI problem. The models must improve, but the company also has to source components, assemble machines and expand production without losing the cost advantage that made deployments possible.
Morton and Noble said data and model training have a clearer development path than manufacturing. For Robo, the two are linked. It needs more robots in the field to produce data and improve autonomy, but it also needs a cost-effective supply chain across the United States and allied countries to build that fleet. That’s something the company will still need to solve as it scales, but for now it’s focused on fulfilling the demand it generated from the launch and starting to build out the flywheel.
The Path Forward
There is a clear playbook for improving models than for manufacturing large numbers of inexpensive robots. Robo’s answer is to begin the feedback loop immediately.
The team pointed to Unitree’s own history: make robots, sell them, collect feedback and repeat. The only way through the manufacturing pain is sale, feedback and improvement. They need to build enough robots, start them working with actual customers, and work towards a product that turns from primarily human operated to mainly automated.
From their small office in the Arts District today, Robo is trying to prove that the path to autonomous labor is not to wait for it. Deploy imperfect machines, work with real customers in complex environments, and continuously iterate through real work
Partnerships & Deals
Transactions are grouped by the primary part of the AI supply chain they support.
Compute, Energy & Infrastructure: Power, data centers, chips, servers, networking, storage and cloud capacity.
Data, Knowledge & Retrieval: Data rights, proprietary datasets, indexing, search and grounding systems.
Models, Training & Developer Tools: Models, post-training systems, agent platforms, security tools and the software used to build AI.
Enterprise Deployment & Distribution: Integrations, channels and services that bring AI into organizations.
Industry Applications & Workflows: AI products built for specific business functions and vertical markets.
Physical AI & Robotics: Embodied systems, autonomous machines and robotics platforms.
Science & Healthcare: AI-enabled biology, medicine, diagnostics, agriculture and clinical systems.
Compute, Energy & Infrastructure
Anthropic reportedly signed a six-year, roughly $45 billion compute agreement with Nscale covering 460 megawatts at its West Virginia campus. The companies haven’t confirmed the reported terms.
Cisco expanded its Secure AI Factory with Nvidia to include rack-scale systems from Supermicro, with channel availability scheduled for October.
Andreessen Horowitz raised a $1.1 billion Machine Age Fund focused on AI infrastructure and hardware, including chips, memory, networking, edge systems, appliances and robots.
Emerald AI raised a $150 million Series A at a $1.05 billion valuation to build software that lets AI data centers operate as flexible grid resources.
Navitas agreed to acquire AI data-center power startup Claros for approximately $232.8 million in cash and stock, including milestone consideration.
Lambda closed a $926 million senior secured term-loan facility to fund GPU infrastructure for a committed investment-grade customer deployment.
Lambda separately secured $1 billion of short-dated private debt to buy Nvidia chips that will be leased to Microsoft, according to Bloomberg.
Lancium partnered with Nvidia across a 4-gigawatt leased AI-factory portfolio and a development pipeline exceeding 15 gigawatts, while Nvidia made a strategic investment. The pipeline isn’t the same as contracted deployed capacity.
SuperX signed a supply agreement with Ezisight and received an initial order for 128 Nvidia B300 AI server clusters, with delivery planned for the fourth quarter.
SCX.ai selected DDN as an infrastructure partner for what the companies describe as Australia’s largest sovereign AI inference cloud.
Kasm Technologies expanded its Intel partnership to run private local AI workspaces on Xeon 6 processors with AMX acceleration.
Data, Knowledge & Retrieval
Keenable raised a $26 million seed round to build web-indexing, search and retrieval infrastructure for AI agents.
Repodo raised a EUR 8.2 million pre-seed round to build a data and development platform for physical AI.
Dun & Bradstreet connected its Commercial Graph to Google Cloud’s Gemini Enterprise through MCP, grounding credit, onboarding and underwriting agents in verified business data.
Dun & Bradstreet brought its Commercial Graph to Perplexity through MCP connectors for research, risk and growth workflows.
Reuters made its multilingual news archive, dating to 1987, available through Snowflake Marketplace for enterprise AI and analytics.
Models, Training & Developer Tools
Nvidia reportedly agreed to acquire Hugging Face for $12.9 billion, although neither company has announced a transaction and reports conflict over whether an agreement was signed.
Deep Cogito raised a $43 million Series A to expand reinforcement-learning post-training for open-weight and enterprise models.
Alice raised $140 million from Apax Digital, bringing total funding for its autonomous security-testing and remediation platform to $280 million.
Stability AI raised a $76 million Series B from strategic entertainment investors to support image, video, audio and 3D models for media production.
Runable raised a $21 million Series A at a $65 million post-money valuation to build agents that can operate and grow businesses.
Embedd raised a EUR 2.3 million pre-seed round to build chip digital twins and agents that automate embedded-software integration for physical AI systems.
Enterprise Deployment & Distribution
Clearlake formed a portfolio-wide partnership with Google Cloud covering AI infrastructure, enterprise data systems, Gemini Enterprise, custom models and cybersecurity.
Arga Labs raised a $10 million seed round to build enterprise agents that learn how a business operates from its data and workflows.
Google Cloud and Verizon formed a strategic partnership to scale enterprise AI across customer-service, network and employee workflows.
Bain partnered with Anthropic to help clients deploy Claude in regulated and operationally complex enterprises.
Salesforce expanded its Anthropic partnership through Claudeforce, a package of Claude integrations and 37 prebuilt sales skills.
Enabled Intelligence selected Seekr for National Geospatial-Intelligence Agency and other defense and intelligence programs.
Sayari partnered with DataExpert to distribute sovereign AI tools for economic-security and investigative work across European governments.
Industry Applications & Workflows
Owner raised $240 million at a $2.3 billion valuation to expand AI agents for ordering, marketing and back-office work at independent restaurants.
Socure raised $156 million at a $5.2 billion valuation and acquired agentic-AI startup Fravity.
Instinct raised a $250 million Series B at a $2.5 billion valuation, bringing the consumer AI company’s total funding to $350 million.
Standard Metrics raised a $20 million Series B from customers to expand AI-assisted portfolio monitoring and fund administration.
eComID raised a $17 million seed round to expand AI-assisted returns, resale and product-lifecycle infrastructure for retailers.
Ringg raised a $10 million Series A to expand its voice AI platform into multimodal customer workflows.
Neno raised a EUR 6.6 million seed round to build AI-native accounting, banking and finance workflows for European small businesses.
Euler raised a $4.3 million seed round for its agentic partner-management platform after bootstrapping to profitability.
Volve raised a $3 million seed round to expand AI systems that analyze construction bids, scopes and procurement documents.
Itoflow raised a $2.5 million pre-seed round to bring AI-assisted analysis and workflows to investment managers.
Descartes acquired Tai Software for approximately $100 million in cash, adding an AI-powered transportation-management platform for freight brokers.
Physical AI & Robotics
Generalist raised nearly $200 million in an extension that values the robotics-model company at $3 billion and brings its Series B to roughly $600 million.
XPeng’s Dogotix raised more than $900 million at a post-money valuation above $6.3 billion for its embodied-intelligence and robotics business.
Corvus Robotics raised $20 million to expand autonomous inventory drones that scan warehouses without fixed infrastructure.
Mara raised $7 million to develop low-cost autonomous systems for detecting and defeating FPV drones.
Motion raised a $2 million pre-seed round to build a humanoid robots-as-a-service platform for European industrial customers.
Science & Healthcare
Adaptyv Bio raised a $40 million Series A to expand automated wet-lab infrastructure for testing proteins designed by AI systems.
Lupin Dental raised a EUR 15 million Series A for a supervised robotic system that prepares teeth for veneers under a dentist’s control.
Legato emerged from stealth with $12 million to develop AI-powered hearing glasses combining directional audio, speech enhancement and computer vision.
CropX acquired portable spectroscopy company SCIO to connect crop-quality measurements with its AI agronomy platform.
Monod Bio granted SignalChem a non-exclusive license to AI-designed protein technologies for custom discovery assays.
HiRO and Differentia Biotech signed an MOU to apply AI-powered simulation to clinical-trial design.
PENTAX Medical expanded its AI-assisted colonoscopy partnership with MAGENTIQ EYE from the United States to EMEA, JAPAC and Latin America.