Every robotaxi story lands on the same two numbers: Tesla quotes around $0.81 per mile, Waymo around $1.40. Next to Uber and Lyft’s $2-plus a mile, the conclusion feels obvious: the driver was the cost, and now it’s gone.
Removing the driver removes only part of the cost. No driver doesn’t mean no labour. A hidden human economy, remote operators, cleaners, technicians, retrieval staff and depot staff, is still priced into every per-mile claim. Those figures are compressed bundles of assumptions.
By the end you can judge why per-mile figures move and which layers decide cost, and place any figure within the wider robotaxi cost picture and what changes when a service really scales.
Robotaxi vs human ride-hailing: which is structurally cheaper, and by how much?
Robotaxis remove the driver’s wage, the biggest line in human ride-hailing, but add remote assistance, depot labour and insurance friction, so the structural gap is not fixed. Waymo’s $1.40 and Tesla’s $0.81 are claims to interrogate; utilisation and support labour per vehicle decide “by how much”.
Human ride-hailing runs at more than $2 a mile, while owning a personal car runs about $0.77 to $1.00 a mile, and that ownership figure is the ceiling on what a robotaxi can charge. Waymo’s effective rates undercut Uber’s passenger rates, because it keeps the whole fare as platform revenue, regardless of its underlying cost.
Uber sits on both sides, aggregating human drivers and Waymo vehicles on one network, where those cars reach 25 to 30 trips a day. Utilisation closes the gap, and who owns the fleet decides who keeps the margin. A market-level contribution margin, revenue per mile minus direct variable cost, is not company profit: Waymo’s ~63% estimate in its most mature markets looks strong until mapping validation, regulatory compliance and corporate overhead land.
The first thing that moves utilisation is the empty miles a robotaxi drives between passengers.
Why is deadheading such a big deal for robotaxi unit economics?
Deadheading is the empty miles a robotaxi drives with nobody paying to be in it, and it decides whether the cost is attractive or unworkable. Empty miles burn energy and accrue depreciation and insurance with zero revenue, inflating every paid mile’s cost. With roughly 43 to 45% of Waymo’s California miles carrying no passenger, paid utilisation is the lever.
The CPUC‘s quarterly VMT reporting splits each mile into three buckets: repositioning and idle, en route to a confirmed pickup, and passenger onboard. The first two are non-revenue; deadheading is the two together. Waymo’s filings show about 53.6% of its mileage carried a passenger across the full study period; the recent plateau runs 55 to 57%, leaving 43 to 45% empty. Waymo is the only operator publishing this denominator, so California is the clearest public window.
The arithmetic is simple: cost per paid mile is total cost divided by paid miles, so every deadhead mile pushes the denominator down. Dispatch logic, fleet sizing, the operational design domain (ODD, where the vehicle is approved to drive itself) and depot location all generate empty miles. In a dense area the next rider is nearby, so there is less deadheading between rides.
Deadheading is one half of the hidden labour picture. The other half is the human still supervising every “driverless” trip.
What is remote assistance, and why do “driverless” robotaxis still depend on human operators?
Remote assistance is off-vehicle human supervision and intervention for edge cases (it is not continuous driving), and it is why “driverless” fleets still carry a human ratio. The ratio moves from roughly 6 vehicles per supervisor today toward 26 to 1 by 2035.
Waymo’s agents, in its own words, provide advice and support but do not directly control, steer, or drive the vehicle. Vehicles hand off at operational design domain boundaries, ambiguous scenarios and blocked routes, and each handoff is a labour event.
The economics live in the ratio. As one remote-assistance operator put it, the driver is no longer in the front seat, but some of the work has moved to an office. Spread one operator’s wage across six vehicles and labour cost per mile is manageable; spread it across 26 and it shrinks fast.
Beijing shows the ratio is a policy variable. China’s national standard requires one remote supervisor per three robotaxis, spreading one wage across only three vehicles. Regulators can set that number, and the cost follows.
What is the hidden human operating economy, and how do you size it?
The hidden human operating economy is the recurring labour automation never removed: charging, cleaning, maintenance, retrieval and depot handling. You size it per active vehicle by counting roles, shifts and retrieval events. Depots are the throughput constraint, and the vendors that run them reveal the cost line hidden inside “ops”.
Removing the driver removes the person in the seat; the people orbiting the vehicle remain. Autonomous vehicles can drive themselves but cannot clean interiors, run safety checks, complete repairs, manage charging or recover from a failed software update, which is why depot workflows become the utilisation bottleneck, alongside incident review, compliance, municipal relations and customer support.
Labour grows with cars and incidents rather than with revenue miles. When a car strands itself at a domain boundary, a human has to retrieve it, passenger or not.
Most of this labour is outsourced. Waymo contracts depot work to partners like Moove and Avis, and vendors such as Flexdrive, Aviari and Transdev absorb charging, cleaning and servicing for asset-light operators. That is why “ops” is a real, billable line.
How do you evaluate robotaxi unit economics before trusting a single per-mile figure?
With the hidden economy in view, evaluating a figure means decomposing it into its layers: vehicle amortisation, energy, servicing, remote assistance, deadheading and insurance, plus the stated utilisation assumption. A figure without utilisation is meaningless. The hidden lines to stress-test are remote-ops labour, insurance and depot servicing.
Every headline number compresses the same layers, which is why cost per mile alone can mislead.
Four questions do most of the work for you: what utilisation is assumed, what energy price and servicing interval, what insurance and liability, and what labour is excluded. Contribution margin is the bridge you cross between a per-mile number and profit.
Cheaper hardware helps only one layer. The legacy Jaguar I-PACE platform runs $150,000 to $200,000 a vehicle, which is why Hyundai and Zeekr platforms matter. Falling sensor and compute prices lower the capital layer; the operating layer, where deadheading and labour concentrate, keeps its own cost.
Most per-mile figures are company claims, and independent verification is limited.
“No driver” was never “no labour”. The people charging, cleaning, maintaining, retrieving and supervising the vehicle are still there, on shift somewhere, and every per-mile figure is a bundle with that hidden human economy priced inside it.
The operating layers, deadheading and support labour, move the number more than the capital layer where cheaper sensors and compute land. Before you trust any figure, ask what utilisation it assumes and what human ratio it prices in. That is the difference between copying a benchmark and reasoning about cost, and carries through to the complete picture of robotaxi scale.
Frequently Asked Questions
How does a robotaxi’s unit economics change as vehicle hardware gets cheaper?
Cheaper sensors and compute lower the capital layer of cost per mile, and little else. Goldman Sachs Research estimates vehicle depreciation falling from roughly $0.35 to about $0.14 per mile by 2035 as platforms such as Zeekr and Hyundai cut hardware cost. But that gain does not touch the operating layer, where deadheading, remote assistance and depot labour concentrate. Cheaper hardware makes an efficient operation cheaper; it does not make an inefficient one viable.
Why does Beijing’s 1:3 remote-supervisor rule matter for the economics of scale?
Beijing’s rule requires one remote supervisor for every three robotaxis, and that cap turns a policy choice directly into per-mile labour cost. At 1:3, one operator’s wage is spread across only three vehicles, far worse than the roughly 6:1 ratio in use today or the 26:1 target for 2035. It shows the vehicles-per-supervisor ratio is not purely an engineering outcome: regulators can set it, and the cost follows.
What is contribution margin, and how is it different from a robotaxi’s actual profit?
Contribution margin is revenue per mile minus direct variable costs, before enterprise overhead. Waymo’s estimated ~63% in mature markets looks strong, but it is not company-level profit, because central validation, mapping, compliance and corporate overhead still land on top. Treat contribution margin as the bridge between a per-mile figure and profitability, not as the profit itself. It tells you whether each additional paid mile contributes.
Does personal car ownership cap how cheap robotaxis can get?
Yes. Personal car ownership, at roughly $0.77 to $1.00 per mile all-in, is the alternative riders compare against, and it sets a ceiling on what robotaxis can charge. If a robotaxi’s all-in cost cannot undercut driving your own car in a given city, the addressable demand shrinks. That is why per-mile figures are usually benchmarked against both human ride-hailing (~$2+/mile) and ownership, not ride-hailing alone.
Where can I check a robotaxi operator’s real empty-mile numbers?
The clearest primary source is CPUC quarterly VMT reporting in California, which records mileage with and without a passenger onboard. Waymo’s filings have shown roughly 43 to 45% of its California vehicle miles carrying no passenger, a passenger-onboard share near 53.6% across the study period. Checking the raw filings lets you verify deadhead claims rather than accepting a press figure.
Who actually does the charging, cleaning and servicing for a robotaxi fleet?
Much of this labour is outsourced, not run in-house. Vendors such as Moove, Aviari and Flexdrive absorb charging, cleaning, servicing and fleet handling for asset-light operators, and partnerships with companies like Avis and Transdev do similar work. That is why “ops” appears as a real, billable cost line rather than an overhead footnote: someone is paid for every vehicle washed, charged, inspected and retrieved.
Why is insurance so hard to price into robotaxi unit economics?
Insurance and liability sit in the per-mile stack but resist a clean number because the risk pool is thin and the data immature. With few robotaxis, limited claims history and unresolved questions about who is liable in an incident, insurers price uncertainty rather than experience. That makes insurance a genuine line item most per-mile figures either estimate loosely or quietly exclude, which is why you should ask what liability cost a quoted figure assumes.
What happens to cost when a robotaxi gets stranded and needs retrieval?
Retrieval is a real labour event, not a rounding error. When a vehicle stops at an operational design domain boundary, gets blocked or faults, a human must reach it, recover it and return it to service, and that cost is incurred whether or not a passenger is onboard. Sizing these retrieval events per week, per vehicle, is part of the hidden human economy that never shows up in a headline cost-per-mile claim.
How many paid trips per day does a robotaxi need to make the maths work?
There is no universal number, but higher daily paid trips spread fixed asset and labour cost across more revenue miles. Waymo vehicles on Uber reportedly reach 25 to 30 trips per day, closing the gap with human ride-hailing’s higher utilisation. The point is not a magic threshold but direction: every extra paid trip lowers cost per paid mile, so trips per vehicle per day is a better operating metric than the headline figure.
Are robotaxi per-mile figures independently audited, or are they company claims?
For the most part, they are company claims, which is why they should be treated as claims to interrogate rather than settled benchmarks. The figures for Waymo (~$1.40) and Tesla (~$0.81) bundle assumptions about utilisation, energy price, servicing and excluded labour that are rarely disclosed together. Independent verification is limited, so the practical discipline is to demand the assumptions behind any number before using it.
Will remote-assistance jobs shrink as the vehicles-per-supervisor ratio improves?
The labour per mile falls as each supervisor oversees more vehicles, but the jobs do not simply vanish. Ratios moving from roughly 6:1 today toward 26:1 by 2035 cut remote operator cost per mile substantially, yet larger fleets also multiply the total number of supervisors, and edge cases still require human judgement. The hidden human economy shrinks per vehicle even while it persists across the network.
