On November 5, 2026, Serve Robotics in the US announced a new business, "Physical AI Services," that offers outside companies the technology it has developed through operating delivery robots and other machines. The business provides three products: a communications platform that connects machines in the field to remote systems, a technology that selects training-relevant scenes from large volumes of operational data, and real-world data collected in the field. One of the initial customers is Bedrock Robotics, which develops autonomous driving systems for construction equipment. Serve aims to expand beyond developing and operating its own robots by providing technology to companies that build robots for different uses.
Three products for connecting, selecting, and sharing
Physical AI refers to AI that supports the perception, decision-making, and actions of robots and machines operating in the real world. Serve is now offering three products to support the development and operation of such AI. Rather than selling robots themselves or a complete autonomous driving system for every use case, it is providing technology that can be built into existing robot development and operations.
"Serve Link" is a communications platform that connects machines operating in the field with AI models, monitoring systems, and training systems on the data center side. It handles the information exchanged between field robots and remote systems. According to Serve, Bedrock Robotics has adopted Serve Link and uses it to connect machines over Wi-Fi, LTE, and 5G, as well as satellite networks. However, details have not been disclosed, such as how much control processing is handled on the cloud side and how machines behave if communications are lost.
"Serve Data Curator" processes the large volumes of operational data collected from connected machines and selects the rare situations that are useful for AI training. Rather than using every record that robots collect day to day for training, it extracts scenes likely to improve a model. The aim is to make the process from collecting operational data to feeding it back into training more efficient. But the announcement does not say whether it can automatically assign ground-truth labels, nor how much it improves training efficiency.
The third product is a catalog of real-world data that Serve has accumulated by operating its own delivery robots on sidewalks. It offers data on terrain, infrastructure, and similar subjects for developing foundation models that serve a variety of uses. What is being sold is data that can be used for AI training, not a trained foundation model itself.
The three products play different roles: connecting field machines to AI systems, selecting data for training, and providing existing real-world data. The target and function of each are summarized below.
| Product | What it handles | Main role |
|---|---|---|
| Serve Link | Field machines and data-center-side systems | Connecting to AI models, monitoring, and training systems |
| Serve Data Curator | Large volumes of operational data | Selecting rare scenes useful for training |
| Real-world data catalog | Terrain, infrastructure, and other data from delivery robot operations | Providing accumulated real-environment data |
Dividing the roles this way makes it easier for robot developers to choose the technology that fits their needs. For example, adopting only the communications platform differs in scope from entrusting external technology with selecting the collected data as well. However, this breakdown simply organizes each product's functions; it does not indicate which performs better or what effect combining multiple products would have.
Turning operating experience with its own robots into a technology base for others
According to Serve, it has deployed more than 2,000 robots across more than 40 cities and 27 hospitals. It entered the hospital robot business through Diligent Robotics, which it acquired in 2026. Sidewalk delivery robots and hospital robots differ in use and operating environment, but they share the same basic challenges: connecting field machines to remote systems and collecting and processing operational data. The new business can be seen as an attempt to sell to other companies the technology accumulated through operating its own robots.
Serve CEO Ali Kashani said in the announcement, "Building robots is just the beginning." The remark suggests the company sees business opportunity not only in developing machines but also in the technology that supports continuous operation after deployment. Alongside its traditional business of deploying its own robots to customers, Serve plans to build a business that supports other companies' robot development and operations.
The expansion does not mean withdrawing from the existing robot business, however. In an interview with TechCrunch, Kashani indicated that the company will continue its delivery and hospital robot businesses. He added that there are opportunities to sell Serve's technology to companies developing robots for farms, warehouses, and similar settings. This describes potential future customer development, though, and no concrete sales in agriculture or warehousing have been disclosed.
In the same interview, Kashani compared the move to the origins of Amazon Web Services (AWS): technology built to support a company's own business, then offered as a platform that outside companies can also use. But Serve has not adopted AWS-like pricing or compatibility, nor has it completed a general-purpose cloud service for every use. The comparison is better understood as describing a business strategy of carving out the technology that robots need regardless of their purpose and offering it to outside companies.
Bedrock Robotics' adoption, as an initial customer, makes that strategy concrete. Bedrock is not deploying Serve's delivery robots at construction sites; it uses Serve Link to connect the construction machinery it handles. This lets Serve widen the market for the technology itself, separately from increasing the number of its own robots deployed or the regions they cover. In addition to competing as a robot maker, Serve is entering the market for technology platforms that support other companies' robot development and operations.
Connectivity reliability and data versatility are the key questions
How far the new business can grow depends on whether Serve can apply the technology it built through its own robot operations effectively in different environments. Kashani told TechCrunch that maintaining connectivity even on unreliable networks is an important challenge. Supporting multiple methods, including Wi-Fi, cellular, and satellite networks, is an advantage, but that alone does not show actual communication quality or stability.
What matters, for example, is how much latency occurs on unstable connections, how often communications drop, and whether the connection can be maintained when switching between networks. Yet Serve's announcement and the TechCrunch article contain no specific figures for Serve Link's latency or availability. Nor is there any explanation guaranteeing seamless switching between connections; what has been presented so far centers on product features and use cases. Judging performance in real operating environments would require detailed specifications and test results.
For Serve Data Curator as well, being able to select data and improving AI model performance should be considered separately. A mechanism that extracts rare situations from large volumes of operational records could help decide which data to use for additional training. But neither the announcement nor the TechCrunch article gives quantitative results on how much a model for which use improves when particular scenes are selected. The existence of a data selection function does not by itself guarantee better training efficiency or model performance.
For the real-world data catalog, differences between the collection environment and where the data is used are also an important issue. Terrain and infrastructure data gathered on sidewalks will not necessarily be equally useful for developing construction machinery or hospital robots. The announcement and the TechCrunch article show no verification of how much benefit results from applying it to different environments. The scale and update frequency of the data provided have not been disclosed either, so the information available to companies assessing fit for their own uses is still limited.
Pricing, detailed contract terms, and the regions where the products will be generally available have also not been published, and it is unclear whether they can be used from Japan. To decide on adoption, companies will need to confirm these commercial terms, along with communication quality and training effects in their own operating environments. For Serve, the new business is an effort to turn experience operating its own robots into a technology asset it can sell to others. Its value will hinge on whether the communications platform and data processing technology can deliver practical results in environments different from delivery robots.