2026-06-18 👁️ — views

Physical AI Fleet Dispatch — How Waymo Routes 1,100 Vehicles and Why It's an Invisible Moat

Waymo's dispatch OS routes 1,100 AVs in real time — the invisible software layer that multiplies ride volume and is nearly impossible to replicate.

Article 118 in the Physical AI Benchmark Series — Physical AI Fleet Dispatch: How Waymo Routes 1,100 Vehicles in Real Time, What Dispatch Efficiency Determines About the Ride Volume Ceiling, and Why the Fleet OS Is an Invisible Competitive Moat

The most important software in a commercial autonomous vehicle operation is not the perception stack that sees the world or the planning module that chooses how to drive. It is the dispatch system — the fleet operating system that decides which of 1,100 vehicles picks up which rider, when each vehicle charges, how idle vehicles are pre-positioned ahead of demand, and how the entire fleet responds in real time to road incidents, geofence boundaries, and surge demand. This software is completely invisible to riders, who see only an ETA estimate and a vehicle arriving. But the dispatch system is the multiplier on every hardware and infrastructure dollar invested in the fleet.

A fleet of 1,100 vehicles with poor dispatch might produce 3,000 rides per day — roughly 3 rides per vehicle, with most time wasted in empty miles, charging gaps, and mismatched positioning. The same fleet with excellent dispatch can produce 6,000 or more rides per day — a doubling of revenue output from the same capital base with no additional vehicles. No other lever available to an AV operator has this magnitude of impact. The perception hardware is fixed; the routes are fixed; the regulatory zone is fixed. Dispatch efficiency is the one variable that can double or halve ride economics without changing anything visible about the fleet.

This article maps the dispatch layer as Physical AI benchmark dimension 118, analyzing what the fleet OS must optimize, how rides-per-vehicle-per-day becomes the key throughput metric, what the architecture of a production dispatch system looks like, and how Waymo’s and Tesla’s different fleet philosophies create fundamentally different dispatch challenges.

Section 1 — What Fleet Dispatch Optimizes

Fleet dispatch is a multi-objective real-time optimization problem. Unlike a navigation system that optimizes a single vehicle’s route, a fleet dispatch system must simultaneously optimize across hundreds or thousands of vehicles, balancing objectives that frequently conflict: minimizing rider wait time requires deploying the nearest vehicle, but minimizing empty miles requires holding vehicles in predicted demand zones rather than chasing each individual request.

The conflict between ETA minimization and empty-mile minimization is the central tension in fleet dispatch. A naive system that always sends the nearest vehicle will minimize individual ETAs but produce high empty-mile rates as vehicles converge on demand pockets from the wrong direction. A system that pre-positions vehicles based on demand prediction will have lower average ETAs across the fleet because vehicles are already near where riders will request rides — but any individual rider who requests a ride in an unpredicted location will face a longer wait. Production dispatch systems navigate this tension with ML models that weight ETA, demand probability, vehicle distribution, and energy cost simultaneously.

The geofence boundary problem is underappreciated. A vehicle operating near the edge of an AV operational zone cannot be sent to pick up a rider if the optimal route passes through unmapped territory. The dispatch system must track every vehicle’s position relative to the operational boundary and ensure no assignment would require the vehicle to exit the geofence during the ride — including accounting for the fact that the rider may request to be dropped off at an address that is on or near the boundary. Managing 1,100 vehicles near a complex geofence edge in real time, during a surge period when every vehicle is in demand, is a non-trivial constraint satisfaction problem.

Section 2 — Rides Per Vehicle Per Day: The Key Efficiency Metric

The ride volume a fleet can produce per day is the product of fleet size and rides per vehicle per day. Rides per vehicle per day is not determined by the vehicle hardware — every vehicle in a fully operational fleet is capable of completing rides. It is determined almost entirely by the dispatch system: how efficiently it matches vehicles to riders, how little time vehicles spend in empty transit, and how well it manages charging to keep vehicles available during peak hours.

The rides-per-vehicle-per-day figure provides a way to reverse-engineer Waymo’s dispatch efficiency from disclosed data. Waymo has reported 150,000+ weekly rides from a fleet of approximately 1,100 commercial vehicles across Phoenix, San Francisco, and Los Angeles. That implies roughly 21,000+ rides per day, or approximately 19–20 rides per vehicle per day across the full fleet — which sits in the “good dispatch” tier in the table above and well above what a poor dispatch system would produce from the same hardware.

The revenue implication of dispatch efficiency compounding is significant. The difference between 12 rides per vehicle per day (poor dispatch) and 25 rides per vehicle per day (excellent dispatch) is more than 2x revenue output from the same fleet capital. At 1,100 vehicles and a $12 average fare: poor dispatch produces approximately $52.8K per day; excellent dispatch produces approximately $110K per day. The dispatch software is effectively worth tens of millions of dollars per year in incremental revenue — and unlike vehicle hardware, it does not depreciate.

The comparison to human rideshare drivers reveals a structural advantage. A human Uber driver in a dense city achieves 15–25 rides per day — but with no central coordination, no demand prediction, and hours lost to personal breaks and end-of-shift. An AV fleet operating 24 hours per day, centrally dispatched with demand prediction, can in principle exceed the per-vehicle output of human drivers even without faster driving, purely through better fleet coordination.

Section 3 — Dispatch Architecture: What the Fleet OS Must Do

A production fleet dispatch system for 1,100 vehicles is a real-time distributed system operating under sub-second latency constraints. It ingests a continuous stream of vehicle state updates, ride requests, and environmental data; it runs optimization algorithms continuously; and it issues routing and assignment commands that vehicles must execute with minimal delay.

The state tracking requirement alone is formidable. 1,100 vehicles each sending position, speed, SoC, and passenger state at 10 Hz produces 11,000 messages per second in steady state. During surges — when demand spikes and vehicles are in rapid transit between pickups — message rates could be higher. The dispatch system must process this stream, maintain a consistent real-time world model, and issue optimized assignments, all within latency constraints tight enough that the vehicle can execute the assignment before conditions change materially.

The charging scheduler deserves particular attention as an underappreciated complexity. Unlike a human rideshare driver who simply plugs in when their shift ends, an AV fleet has no natural off-shift periods. The vehicles are capable of operating 24 hours per day, but batteries must be recharged. Scheduling 1,100 vehicles to charge in staggered windows that avoid peak demand periods, while keeping each vehicle’s SoC above a safe operational floor, is a scheduling problem that does not have an obvious greedy solution. A naive approach — charge every vehicle to full before returning it to service — will consistently drain vehicles from service precisely when demand peaks, because high demand periods are the same periods when vehicles have been running longest.

The anomaly detection component has a direct impact on public safety perception. When a Waymo vehicle stops unexpectedly in an intersection, the visual impact is significant — it creates traffic delays and generates social media coverage. A dispatch system with fast anomaly detection can identify the vehicle, dispatch remote assistance within minutes, and typically resolve the situation before it becomes an extended incident. Slow anomaly detection allows minor software confusion events to escalate into multi-hour blockages. The 2024–2026 period saw Waymo invest heavily in this component, reducing the frequency and duration of public vehicle stalls.

Section 4 — Tesla’s Network Architecture: Different Design Philosophy

Tesla’s entry into the commercial AV market introduces a fundamentally different fleet dispatch challenge. Where Waymo operates a company-owned fleet under complete centralized control, Tesla’s robotaxi network is designed around a hybrid model: Tesla-owned Cybercab vehicles in high-density zones, supplemented by consumer-owned Model Y and Model 3 vehicles that opt into the network when the owner does not need the car. This architecture creates advantages in fleet density and capital efficiency, but introduces dispatch constraints that Waymo does not face.

The consumer-owned vehicle dispatch problem is genuinely novel in the history of fleet logistics. A traditional fleet operator has full control over every vehicle in service — when it operates, where it goes, when it charges, and when it is retired. Tesla’s hybrid model accepts consumer vehicles into the network on the owner’s terms: when the owner does not need the car. This creates a fleet with unpredictable availability patterns. A dispatch system that has learned to expect a specific vehicle in service on Tuesday mornings may find that vehicle unavailable because the owner has a doctor’s appointment. At scale, these individual availability fluctuations may average out into predictable aggregate patterns — but the dispatch system must be designed to handle them.

The SoC problem for consumer vehicles is particularly significant. A consumer who owns a Tesla Model Y and occasionally puts it in the robotaxi network may not maintain the same charging discipline as a fleet operator. A vehicle that enters the network at 30% SoC is significantly constrained in dispatch — it can complete only a limited number of rides before it must return to charge, and it may need to be sent to a charging station at an inconvenient time for the fleet. Waymo’s charging scheduler can plan charging for every vehicle in its fleet because it controls the charging. Tesla’s dispatch system must infer the SoC state of each consumer vehicle and work around owner charging behavior it cannot fully control.

The potential network effect advantage of the Tesla model is real but depends on achieving sufficient consumer vehicle enrollment. A city in which 10,000 consumers enroll their Tesla vehicles in the network, making them available during typical 9-to-5 work hours, creates a fleet density that Waymo cannot match without purchasing 10,000 vehicles. If the dispatch system can manage the availability unpredictability effectively, the Tesla model could achieve lower ETAs at far lower capital cost per ride than a fully company-owned fleet model.

Section 5 — The Dispatch Efficiency Benchmark

Dispatch efficiency can be measured along several operational dimensions. For a public company or investment analysis context, these metrics translate directly into ride economics, unit economics, and ultimately the path to fleet profitability. For a technology comparison, they reveal the maturity and sophistication of the fleet operating system — which is one of the most difficult software systems to build in the AV industry and one of the least visible.

The empty mile rate is one of the most useful dispatch efficiency benchmarks because it is directly observable from vehicle GPS data and directly translates to cost efficiency. An empty mile is a mile driven without a paying passenger — the vehicle is consuming energy, accumulating wear, and not generating revenue. In a well-optimized fleet, the vehicle is never idle: it is either completing a ride, en route to a pickup that a demand prediction model identified before the rider even requested the ride, or in a planned charging window. An empty mile rate of 15–20% means that for every 5 miles of total fleet driving, 1 mile is empty transit. An empty mile rate of 30% means the fleet is operating far below its potential, with vehicles spending nearly a third of their driving time generating no revenue.

Vehicle utilization rate is the complementary metric. A vehicle that is in revenue service 65–75% of its available hours is spending the remaining 25–35% in charging, depot maintenance, cleaning cycles, or anomaly recovery. The 65–75% target reflects the structural non-negotiable costs of operating any vehicle — charging cannot be eliminated, only scheduled efficiently. Improving utilization above 75% likely requires either faster charging (reducing the time each vehicle spends at the charging station) or better charging scheduling (concentrating charging in demand troughs).

The invisible nature of the dispatch system is precisely what makes it a durable competitive moat. A new entrant to the AV market can, in principle, acquire the same sensor hardware as Waymo. It can license HD mapping data. It can hire engineers who have worked on AV perception. But it cannot quickly replicate a dispatch system that has been optimized against millions of real rides, calibrated on real demand patterns in specific cities, and tuned through years of operational feedback. The algorithms that predict where demand will arise in Phoenix at 6 AM on a Tuesday during baseball season, the models that balance ETA against empty miles in specific geofence configurations, the anomaly detection systems trained on thousands of edge cases — all of this accumulated operational learning is embedded in the fleet OS. It is not disclosed. It is not patentable in most of its specifics. It compounds with every additional ride the fleet completes.

At 150,000 rides per week in mid-2026, Waymo is accumulating dispatch intelligence at a rate that no competitor currently matches. Each ride is a data point: where did the rider request pickup, where did they go, how did the vehicle routing perform, how accurate was the demand prediction, how efficient was the charging schedule? The fleet OS learns from each of these data points, compounding its dispatch efficiency in ways that are real but invisible. This is the moat. Not the vehicles. Not the maps. The software that routes 1,100 vehicles in real time, invisibly, at a level of optimization that takes years to build and cannot be purchased off the shelf.

Note: All figures labeled “(est.)” are derived from publicly available information, engineering estimates, and industry reporting as of mid-2026. Waymo does not publicly disclose dispatch system specifications, empty mile rates, or vehicle utilization data; estimates are directional. This article does not constitute investment advice.

Sources

• Waymo One operations — Waymo ↗
• Fleet management optimization — IEEE Transactions on Intelligent Transportation ↗
• Rideshare dispatch systems — Uber engineering blog ↗
• Tesla robotaxi network — Tesla ↗
• AV fleet operations analysis — McKinsey Center for Future Mobility ↗

Tags