Serve Robotics Sells Its Fleet Stack, but Does Not Say to Whom

Serve Robotics says its new Physical AI Services business already powers "thousands of vehicles" across a dozen industries, and the number to dispute is the denominator. The company has deployed more than 2,000 of its own robots; whatever sits between 2,000 and "thousands" belongs to someone else's fleet, and neither the release nor the coverage gives that number.

Serve, the Nasdaq-listed sidewalk delivery company, announced the business on November 5, alongside third-quarter results. It sells three products built for its own robots to other companies that build machines. Co-founder and CEO Ali Kashani calls it "the AWS moment where we are taking the technology we've built for us", a summary of a TechCrunch interview.

The three products

The release names three products. Each one packages a piece of infrastructure that a delivery robot company has to build before its robots can leave a depot.

Product	What it does	Source of the data
Serve Link	Connects field machines to models, monitoring and training systems; bonds 4G, 5G, Wi-Fi, satellite and private cellular	Serve's own fleet operations
Serve Data Curator	Turns petabytes of sensor data into searchable fleet data and finds edge cases for training sets	Customer and Serve fleets
Data Catalog	Street-level terrain and infrastructure data	Serve's sidewalk deployments

According to Unite.AI's account of the release, Serve Link also handles task management, unit-status monitoring and remote assistance when a machine hits an exception it cannot resolve. The catalogue, Serve says, holds millions of hours of street data collected in ways that respect local laws and bystander privacy.

The company is not new to the business of selling robots at scale. It was spun out of Uber's 2021 acquisition of Postmates and went public in 2024, per the TechCrunch-derived summary, so the infrastructure it is now selling has been tested against a real delivery fleet.

One named customer

The launch customer is Bedrock Robotics, which builds autonomous systems for heavy construction equipment. Bedrock uses Serve Link to connect its equipment over Wi-Fi, LTE, 5G and satellite. It is not putting Serve's delivery robots on a construction site; it is buying the connectivity layer.

Serve says the services already support vehicles in construction, mining and autonomous trucking, plus its own sidewalk fleet. No other customer is named in the release or in the TechCrunch-derived summary. Pricing, availability dates and contract terms are not given anywhere either.

That makes the claim hard to size. A connectivity product sold per vehicle and one sold as an engineering contract are different businesses, and nothing published says which one Serve has.

Why a delivery company would sell this

Reasoning in a chain: a sidewalk robot depends on a cellular connection, Serve operates more than 2,000 of them in more than 40 cities, and so it has had to build for bad networks. Kashani says as much: "You need to be able to manage really unreliable networks."

A construction robot or an autonomous truck faces the same problem at a different price per failure. That is the credible part of the pitch. The less credible part is the AWS analogy, because the data catalogue is built from sidewalk deployments, and it is my judgment, not a sourced finding, that a mining haul truck gets little from kerb-level street data.

The company is also changing who it delivers for. TechCrunch's reporting, as summarised by TechBytes, describes sidewalk delivery shifting from Uber toward DoorDash and Wonder. Serve's release says its robots reach about 3 million people and support delivery for more than 4,000 restaurants across the United States and Australia. It also acquired Diligent Robotics in 2026, which added an indoor hospital robot, and says its robots operate in 27 hospitals.

What the sources do not say

Revenue from Physical AI Services is not disclosed. The TechCrunch-derived summary gives no financial impact and no scope for the Bedrock contract, and it describes a T-Mobile arrangement to bring AI computing closer to robots without saying whether T-Mobile is a customer. Availability and rollout dates for the three products are also missing, so a buyer cannot yet tell whether Serve Link is a shipping product or a programme for selected partners.

The market data attached to the press release page showed Serve at $2.93 a share and a market capitalisation of about $256 million at the time we fetched it. A company that size has an obvious reason to want a software line.

The data pitch is aimed at a problem others are attacking differently. Berkeley's agentic-robotics work asks whether agents can program industrial robots without training data, and Nvidia's GR00T 2 is another route to robot models. Neither tells a buyer why sidewalk data is the right data, and Serve has not said who has bought the catalogue.

What would change my read

The variable is the count. If Serve reports the number of non-Serve machines connected through Serve Link, and that number is in the hundreds rather than the thousands, "AWS moment" describes a pilot programme. If the number is in the thousands and Serve discloses revenue per vehicle, the claim is sized and can be argued with. Third-quarter results were published on the same day as the launch; the next reported quarter is the first place a number could appear.