Serve Robotics Launches Physical AI Services to Sell Fleet Connectivity and Training Data
Serve's fleet connectivity and real-world training data are now available to other robot companies
Serve Robotics launched Physical AI Services on November 5, opening the connectivity and training-data infrastructure behind its own delivery robot fleet to other companies building autonomous machines. The move positions the San Francisco company — best known for its sidewalk food-delivery robots — as a potential infrastructure provider across construction, mining, and autonomous trucking, while Bedrock Robotics, a $4.1 billion autonomous-excavation startup, became the first named customer.
The announcement arrived alongside Serve's third-quarter 2026 financial results, which showed revenue of $2.835 million, up 313% year over year from $687,000 in the same period in 2025. Daily delivery volume tripled from a year earlier as new partners came online, and the company raised its full-year 2026 revenue guidance to $12 million from a prior range of $9 million to $10 million.
Three Products Built From Five Years of Fleet Operations
Physical AI Services consists of three distinct offerings that Serve assembled through a series of acquisitions and in-house development since its 2021 spin-off from Uber.
The first product, Serve Link, provides reliable connectivity between robots operating in the field and the off-board computing systems — cloud-hosted AI models, monitoring infrastructure, and training pipelines — that increasingly do the heavy work of perception and decision-making. The underlying technology comes from Voysys AB, a Swedish teleoperation company acquired by Serve in 2025 for approximately $5.75 million. Voysys's system achieves glass-to-glass latency as low as 50 milliseconds by combining proprietary bandwidth regulation, advanced video compression, and multi-link redundancy across Wi-Fi, LTE, 5G, and satellite networks simultaneously — so a lost cellular signal does not break the robot's connection to its remote operator or its cloud model.
The second product, Serve Data Curator, addresses a bottleneck that every robot maker faces at scale: most of the data a large fleet generates is uneventful, but the rare, difficult edge cases are what improve a model's reliability. Serve says its curation system automatically identifies those high-value moments from the continuous sensor stream, allowing training runs to focus on the examples that actually matter rather than processing hours of unremarkable footage.
The third offering is a real-world foundation-model data catalog drawn from Serve's deployments in more than 40 cities. The data covers the kind of complex, unstructured outdoor environments — sidewalk obstacles, curbs, pedestrian flows, variable weather — that simulation environments struggle to replicate convincingly. It is explicitly intended to give robotics companies training data for navigating human environments rather than controlled warehouse spaces.
Why Bedrock Chose to Buy Rather Than Build
Bedrock Robotics, founded in 2024 by veterans of Waymo's autonomy program, raised $600 million in Series C funding in October 2026 at a $4.1 billion post-money valuation. Its Bedrock Operator is a reversible, same-day retrofit kit that installs on existing excavators and enables fully autonomous earthmoving without permanent equipment modification. The company has conducted what it describes as the construction industry's largest supervised autonomous excavation test, at a 130-acre manufacturing site in Phoenix.
Despite deep autonomy expertise and substantial capital, Bedrock chose Serve Link for its connectivity layer rather than building the infrastructure itself. That decision is the most informative detail in the launch announcement. Construction sites face some of the same multi-network connectivity challenges as city sidewalks — unpredictable cellular coverage, obstacles blocking line-of-sight links, the need for a human operator to intervene if the autonomous system encounters an unexpected situation. By using an existing, battle-tested connectivity stack, Bedrock can focus engineering resources on its core excavation autonomy rather than rebuilding teleoperation infrastructure.
The pattern has a precedent in cloud computing: companies that could theoretically run their own servers frequently chose AWS or Azure instead, once those platforms demonstrated reliability and offered a clear cost advantage. Serve is making a similar bet that robot makers will reach the same build-vs.-buy calculation.
A Guidance Raise With Important Context
The $12 million full-year guidance Serve announced on November 5 represents a meaningful recovery from August 2026, when the company cut its revenue outlook from $26 million to $9 million to $10 million after Uber Eats delivery volumes fell below expectations. The Q3 rebound was driven by new delivery partners, including DoorDash, whose partnership revenue grew nearly 50% sequentially in the second quarter, and by branding revenue that grew nearly 80% sequentially in the third quarter. Recurring revenue reached approximately 69% of total Q3 revenue, a figure the company says reflects the structural shift toward software services.
Operating losses remain substantial. Serve recorded a GAAP net loss of $59.2 million in Q3, with $198 million in cash and marketable securities on hand as of September 30. The company has not disclosed pricing, contract terms, or financial projections for Physical AI Services, leaving the platform's contribution to future revenue entirely speculative for now.
The Broader Physical AI Infrastructure Race
Serve is entering a market where the competitive lines are still forming. NVIDIA dominates the physical AI development stack through its Isaac robotics SDK, Cosmos world foundation models, Omniverse simulation platform, and Jetson edge processors, with partnerships across major industrial robot manufacturers. Field AI offers an API-first autonomy layer for robots operating in unstructured environments, positioning itself closer to the model and decision-making portion of the stack. Viam provides cloud-based fleet management and teleoperation for general robotics deployments.
Serve's specific positioning — real outdoor environment data from public deployments, plus a connectivity stack field-validated at multi-thousand-vehicle scale — sits between those layers. It does not compete with NVIDIA's simulation and model development tools, and it offers more domain-specific depth than a general cloud robotics platform. Whether that niche is wide enough to support a durable infrastructure business depends on how quickly other robot makers decide that building their own connectivity and data pipelines is not worth the investment.
The company's October international expansion into Australia and its hospital healthcare footprint through Diligent Robotics, acquired in January 2026, give the data catalog geographic and environmental breadth beyond U.S. city sidewalks. The practical question Physical AI Services must answer in its first operating year is whether the environments where Serve's fleet has actually driven produce training data relevant enough to accelerate robots that dig excavation trenches, haul mining ore, or navigate long-haul highway routes. The Bedrock partnership is the first real-world test of that claim.