The Robotaxi Depot Has a Second Job

Grid Signal — Issue #012 · September 2026

The claim

The fleet can grow faster than its grid connection. A fleet operator orders electric vehicles and installs chargers. Where the depot has spare electrical capacity, early deployments may fit behind the existing connection. Expansion changes the question: how much power must the whole site draw at once?

A depot with 200 stalls at 50 kW has 10 MW of charger capacity behind one connection. What it actually needs depends on how many cars charge at once, how much energy they need and how long they can stay — the variables a good architecture makes visible. In Germany, the VDA warned in August that expanding the grid connection of industrial sites and large truck charging hubs can take up to ten years. Fleet electrification is also a grid-connection business.

The business driver, in one sentence: operating the depot's load alongside its fleet can make a constrained connection workable, reduce charging costs and, under an eligible contract, earn payment for flexibility. These are separate benefits. The simulation below tests operating cost. An earlier connection depends on local capacity and an agreement with the network operator; flexibility revenue needs an eligible service, a buyer and settlement terms.

Who is at the table

Tesla's interest form asks: "Help Us Build Our Robotaxi Network." Companies can register interest in Cybercab fleet purchasing, mobility hubs and infrastructure, among other opportunities. A car maker is inviting partners into both the fleet and the places that support it.

One operating model came into view in April, when Uber and Hertz announced that Hertz's affiliate Oro Mobility would support Uber's Lucid/Nuro robotaxi programme, initially planned for the San Francisco Bay Area. Its scope includes charging, maintenance, repairs, cleaning, and depot staffing. Charging sits alongside the other tasks that keep the fleet available.

The same separation of vehicle technology and fleet operations appears elsewhere: WeRide and Uber plan at least 1,200 robotaxis across Abu Dhabi, Dubai and Riyadh, "as soon as 2027" and scaled up as regulatory approvals and performance milestones are met, with Uber or local third-party partners responsible for fleet operations (WeRide, Feb 2026).

These announcements do not explain how dispatch, charging and the grid connection will be coordinated. That leaves an architectural question I would put on the table early: who can commit the depot to an electrical operating envelope while protecting its transport service?

Why the robotaxi depot is interesting

Managed charging is not new. Conventional fleets already use software to coordinate transport plans, charging and site power limits. The robotaxi opportunity is to bring those decisions into the same planning process as each new trip request.

The architecture I would propose gives the planner a shared view of the next airport run, each vehicle's usable energy and the next available charging slot. It can then decide which car takes the fare and which stays at the depot. Charging becomes part of dispatch. That requires both knowledge of the service constraint and authority to act on the charging plan; vehicle autonomy alone establishes neither. Whether using that authority improves the economics is the open question.

The numbers, run for real

Claims are cheap, so we built the depot in code and ran it against real prices — the experiment runs daily at depot-charge.surge.sh, with source code and parameters available for inspection. Setup: 200 vehicles, 75 kWh usable, 200 stalls at 50 kW, a 6 MW connection, a synthetic city demand curve — experiment parameters, not Cybercab specifications. Prices: the real German day-ahead price of each day (DE-LU, 15-minute). Three strategies, the same fleet configuration and transport demand:

• Dumb — plug in on return, charge to full.
• Price-aware — schedule charging in cheap slots while prioritising readiness for the next dispatch. The 6 MW physical connection remains the limit.
• Flex — the same planner plus a self-imposed 2.5 MW import ceiling, plus vehicle-to-grid when the model's spread and readiness conditions permit it.

Each date is a separate scenario. Each strategy first runs three warm-up days using that date's prices and demand, then carries its vehicle state into the reported day. State is not carried from one calendar date's scenario to the next. In these cases, the dumb depot's peak is 3.15 MW at 19:00, the return wave.

Network charges follow the shape of a German medium-voltage tariff: a capacity component plus an energy component, with rates switching at 2,500 hours of annual utilisation, as in Bayernwerk's 2026 price sheet. The simulator uses illustrative rates. It allocates the annual capacity charge by day, using each strategy's maximum 15-minute import across the records so far; annual utilisation is estimated from that day's energy use multiplied by 365. This is a short experiment with an annual tariff model, not an observed annual bill.

Service is measured against an external request set: how many vehicles the city wants on the road in every 15-minute step, identical for all strategies. This measures supplied vehicle-slots, not individual passenger journeys. Each record also states stored fleet energy at 00:00 and 24:00. The costs below credit an increase, or charge a decrease, at that day's mean wholesale price — an explicit valuation assumption.

Eleven scenarios, 7–17 September 2026, using simulator v0.5.0. Every strategy supplies 99.9% of demanded vehicle-slots on every day. Costs include the stored-energy adjustment:

Three things the table says. Price-aware charging saves 12.4% against dumb over these eleven scenarios. Flex saves a further 5.5% against price-aware, about €198 a day, bringing its saving against dumb to 17.2%. Its peak stays at 2.5 MW. On the capacity-charge component alone, Flex saves about €45 a day against dumb and €225 against price-aware. Including the network energy component, those network-charge savings are about €36 and €216. V2G is negligible in this sample — reported export revenue less modelled battery wear is at most €1 a day, with wear priced at 4 ct/kWh. All figures refer to this fixed sample; the live demo continues to add dates.

The 2.5 MW ceiling is a policy: it yields to service energy, and every record counts whether it did. On all eleven days it held, with matched service. That does not establish performance under demand stress or charger failure. A connection commitment needs local enforcement of the agreed site-import limit; the architecture below has a component for exactly that. Its connection value still requires a network operator's assessment and agreement.

Each record also carries a fleet-contribution ledger: trip contribution at an illustrative €/km after platform fees and avoidable non-energy costs, minus electricity and network charges, each deducted once. All three strategies supply the same vehicle-slots and drive the same kilometres in these scenarios, so their assumed trip contribution is identical. The adjusted ledger therefore tracks the adjusted cost differences. It is an accounting framework for the next question, not evidence of observed fleet profit: was any of this worth a ride?

What the experiment does not measure

Suppose moving 20 kWh from €0.30 to €0.15/kWh saves €3, and the schedule that achieves it keeps one car off the road for an extra half hour worth €8 of trip contribution after fees and avoidable non-energy costs. Electricity cost fell by €3; fleet contribution fell by €5. Illustrative numbers, not observed fares — but the shape is the point. The next fare and the next charging slot compete for the same vehicle.

Our simulator answers a valid first question: can coordination cut cost while preserving the modelled level of service? It prioritises readiness and holds transport demand fixed. It cannot answer the second: within agreed service standards, when does energy value justify a change in availability? That needs the ledger to grow — trip income plus eligible grid revenue, less electricity and marginal operating cost, on the same realised requests — and a planner allowed to trade inside a service band. Applying an average hourly revenue to every parked car would misprice flexibility; so would counting every missed fare as lost profit. That planner is next.

What is different for a large delivery fleet

Consider a delivery station with 200 electric vans: parcels sorted overnight, vans loaded in morning waves, delivery partners taking routes out through the day. For this example, assume a single-shift operation with most vans available to charge between 20:00 and 06:00. This is an illustrative duty cycle; each vehicle's actual return, required energy and departure time would still determine its charging window.

Here the opportunity cost has a different shape. Charging at 02:00 may displace no scheduled delivery, provided the van is ready for its next route. Flexibility comes from the time left after that commitment is protected. Late returns, cold-weather energy needs, charger failures and reserve vehicles all consume part of that margin. If the vans are away at midday, they cannot use that period's cheap energy at the depot; overnight prices and network charges can still make scheduling valuable.

With predictable routes, departure times and required charge can often be inputs to the planner. It can spread charging across the available window and avoid an unnecessary peak when vans return together. Whether that fits behind an existing connection depends on the energy required, available charging power and the site's other loads. The same planning can support a smaller connection request or a flexible connection agreement where the network operator offers one.

The architecture starts with three familiar components: a site controller that measures total import and holds the ceiling through the chargers' controls; a scheduler that works back from the departure list and required charge; and a record of the envelope promised and the import measured. Delivery reliability remains the operating requirement. The grid-facing opportunity is to turn the flexibility left around it into a commitment that can be enforced and verified.

What I would build

Start with shared state. Per vehicle: location, usable energy, availability, commitments. Per depot: charger status, measured site load, permitted import envelope. Per plan: tariffs, expected trip contribution, contracted grid services. Every input carries a timestamp and a source; stale telemetry lowers confidence and never silently becomes a fact.

Then four components. Forecasts with uncertainty — demand by area and hour, energy per journey, return times, the probability a car is actually available — checked against simple baselines. A constrained planner over a rolling horizon, re-planning on arrivals, demand shifts, charger failures and envelope changes, with reserves against forecast error. A policy check before execution, against electrical limits and existing commitments; an infeasible schedule is exposed, not hidden. Local enforcement: a depot controller that measures total site import and applies the agreed ceiling through the chargers' own controls, with a defined fallback when communication is lost. If service and the ceiling cannot both be met, the system must surface the service shortfall and re-plan within the permitted envelope. A fifteen-minute plan does not decide the response time to an overload.

Every decision and delivery is recorded and signed: promised, planned, measured, under which rule version — the pattern from Issue #011, envelope before, authority during, record after, applied to a load. Production adds accepted meter data and agreed performance rules. Export comes last, after the charging case works. And before any of it: resolve four interfaces with their owners — dispatch authority, charging authority, responsibility for the connection, entitlement to energy benefits. Owning the vehicles establishes none of them.

The question for whoever fills in the form

The form asks who will buy the cars and who will build the hubs. I would add a question: who can operate the fleet inside the site's electrical limits, and demonstrate that the transport commitments still hold?

For an operator planning to scale, that question belongs alongside vehicle procurement and depot design. Where local grid capacity is constrained, the connection date can set the expansion schedule. A smaller, enforceable import requirement gives the operator and the network company another option to assess.

The depot has two jobs: keep the fleet available and keep its electrical commitment. I would design dispatch, charging and the grid connection around both.

The depot simulation is open source (MIT) and operated daily. The simulation results above come from the signed records for 7–17 September 2026. Demo · Records · Source · Reproduction instructions. Wrong assumption? Tell me which one and what number you would use.

Grid Signal is written by an AWS Principal Solutions Architect working with the energy sector. Views are personal. Public sources inform the analysis; the financial and delivery-duty-cycle examples are illustrative. The architecture is a proposal, distinct from the current simulator and from any company's internal operating model.