Serve's Strategy: From Delivery Robot to Physical AI Platform

Serve Robotics, known for its fleet of autonomous sidewalk delivery robots serving meals in U.S. cities, announced last week the launch of Physical AI Services, a new business line aimed at transforming the company into a technology provider for other robot-building companies. The move marks a phase transition in a market that is moving from prototypes to scale infrastructure.

From operations to platform

Serve didn't invent this strategy out of thin air. Since 2021, when it was spun off from Uber's acquisition of Postmates, the company has built a fleet of nearly 2,000 delivery robots operating across 44 cities in 14 states, with quarterly revenues jumping from US$176,000 in Q1 2024 to US$3.2 million in Q2 2026. In January 2026, it acquired Diligent Robotics, manufacturer of Moxi — a hospital delivery robot that has already completed over 1.25 million deliveries in more than 25 hospitals.

But robots delivering food in Los Angeles or transporting medications in hospitals don't sustain a billion-dollar business. The physics are simple: for every dollar spent on delivery, Serve needs to keep human operators monitoring, loading, and maintaining a fleet that's physically distributed. It's a model with thin margins and growing logistical complexity.

Physical AI Services represents the leap: licensing the technology Serve built to operate its own robots. Three products launched simultaneously.

Serve Link: connectivity for robots

The first product, Serve Link, connects robots in the field to AI systems running in data centers. Instead of packing heavy computation into each robot — which means expensive hardware, high energy consumption, and battery limitations — Serve proposes delegating offline intelligence processing to the cloud, where more powerful models can run without compromising the device's battery or weight.

CEO and co-founder Ali Kashani described the problem frankly: "Any robotic application generally needs to connect the robots to some offline intelligence, whether it's to collect data or do exception handling or whatever. It's the connectivity piece that has to be really high performance. You need to be able to manage really unreliable networks and still reliably connect, and that's what we solve."

The AWS analogy isn't just rhetoric. Just as Amazon discovered it needed scalable servers to run its own online store before renting capacity to other companies, Serve discovered it needed robust connectivity to operate 2,000 robots in the real world before offering the same service to third parties.

Data as product

The second product is a data curation tool that helps companies find edge cases in the massive volumes of real-world data a robot collects. Instead of manually reviewing hours of video and LiDAR, the algorithm automatically identifies rare situations — a pedestrian crossing outside the crosswalk, a vehicle stopped in the middle of the street, adverse weather conditions — and labels them for model retraining.

The third product is a foundation model data catalog based on the real experiences of Serve's sidewalk robots. It's the largest publicly available collection of data from robots operating in uncontrolled human environments: sidewalks, intersections, parks, shopping centers. Companies training computer vision or autonomous navigation models can use these data as a training base.

Why it matters

The Physical AI market — the intersection between artificial intelligence and physical objects that perceive, reason, and act in the real world — was valued at US$450 billion by 2030 by Serve itself in its investor presentation. But value distribution in this market remains undefined. Who will be the NVIDIA of this ecosystem? Who will be the Amazon Web Services?

The answer, apparently, might be a company that started by making sidewalk delivery robots. Bedrock Robotics, a startup developing autonomous systems for construction equipment, is already a customer of Physical AI Services. T-Mobile also announced a partnership to bring AI computing closer to robots, simultaneously with the third-quarter earnings results.

Serve's strategy is ambitious because it places the company in direct competition with established giants. NVIDIA with its Isaac stack, Amazon with its logistics robotics services, Google with ROS (Robot Operating System). But Serve has an advantage that none of these companies possess: it operates the largest commercial fleet of autonomous sidewalk robots in the United States, and each day of that operation generates data that no simulation can reproduce with fidelity.

The risk is real. Serve hasn't achieved consistent profitability — Q2 FY2026 revenues were US$3.2 million, which is impressive growth but at a still-small scale. Competition in Physical AI includes companies with decades of industrial robotics experience and bottomless resources. The market has yet to prove that third-party companies will trust their robotic operations to a startup that began by delivering pizza.

But the lesson from previous platforms is clear: AWS was born from Amazon.com's own internal need, not conceived as a product from the outside in. The SaaS and autonomy licensing model that Serve announces as a long-term horizon could generate margins of 50% or more in recurring revenue, far superior to delivery margins. If the technology works — and the Diligent data, with 15 times faster perception speed in Moxi 2.0, suggest that it does — Serve could be building the infrastructure layer upon which the next generation of robots will be built.

Sources: TechCrunch, Serve Robotics Investor Deck, Serve Robotics Q3 2026 Earnings

✓ Independent sources cross-checked and verified before publishing