Cybercab, Robotaxi Fleets, and the Economic Structure of Autonomous Mobility

1.1 The framing problem

Two claims circulate about autonomous ride-hailing, and both are wrong in the same way.

The first: autonomy makes mobility nearly free, because the driver takes 55–70% of the fare and the driver is being deleted. The second: autonomy cannot work economically, because the vehicles are expensive, the software is unfinished, and the liability is unbounded.

Both claims share a defect. They treat "the driver" as a scalar — a wage — and then either subtract it or fail to subtract it. But the human driver in a ride-hailing network performs at least four separable economic functions, only one of which is labor:

• —Labor. Perception, planning, control, and passenger interaction.
• —Distributed capital provision. The driver supplies the vehicle. Platform capex per marginal vehicle is approximately zero.
• —Liability absorption. The driver is the legally responsible operator and carries individually underwritten, statistically independent risk.
• —Demand-elasticity buffering. The driver self-deploys into peaks and self-withdraws in troughs at zero cost to the platform, because their reservation wage — not the platform's capital charge — governs the decision.

Autonomy removes function (1). It does not remove functions (2), (3), and (4); it transfers them to the fleet operator, where each is priced differently and, in two cases, worse.

1.2 The analytical stance of this paper

We adopt the permission-structure reading. A constraint that appears to be a prohibition ("autonomous ride-hailing cannot be profitable") is more usefully read as a specification: it names the exact currency that must be paid and the exact quantity. The value of the reading is that it converts vague pessimism into a list of things that can be measured and falsified.

Applied here, the question is not will robotaxis reshape the economy but: which specific physical or financial quantity has to move, by how much, before each claimed effect becomes available — and is that quantity currently moving?

This produces a different kind of answer than either the bull or bear case. Some claimed effects turn out to be nearly free. Some turn out to be priced in a currency that no amount of software improvement can pay. We attempt to say which is which.

1.3 Scope and limits

This paper is about the United States, about passenger ride-hailing, and about the period 2026–2036. It does not model freight, does not model non-US regulatory regimes, and does not attempt a valuation of any security. It uses Tesla's Cybercab as its calibration point because that program has disclosed enough specification detail to support a bottom-up model, not because the analysis is Tesla-specific — most of the results apply to any purpose-built, camera-or-lidar, high-utilization autonomous fleet.

We make no claims about intentions, competence, or good faith of any firm or individual. The failure modes identified here are structural. They would be present under any management.

Any economic model of this sector is worthless without an accurate statement of where deployment actually stands, because the gap between current scale and economically significant scale is the entire subject.

2.1 Cybercab program status The first Cybercab production unit was built at Gigafactory Texas in February 2026. Continuous production began in April 2026. [E] Tesla designed the vehicle to self-certify against existing Federal Motor Vehicle Safety Standards rather than seek an exemption, which means the 2,500-unit annual NHTSA exemption cap does not bind it, and the pending federal legislation that would raise that cap to 90,000 units is irrelevant to the program. [E] Disclosed specifications from mid-2026 regulatory filings: 48 kWh battery, 163 kW motor, approximately 3,113 lb curb weight, 418 miles laboratory range. Target price under $30,000. Two seats, no steering wheel, no pedals. [E] Management has described the assembly process as closer to consumer-electronics manufacture than conventional vehicle assembly, with a stated target cycle time under ten seconds per unit and a theoretical line capacity in the millions per year. [E] As of mid-2026, on the order of 100 production units had been observed. Cybercabs began carrying employee riders on the Giga Texas campus in July 2026. The Cybercab was not yet in public robotaxi service as of the evidence cutoff, and no date for that had been given. [E]

That last item is the single most important fact in this section and is discussed in §6.4.

2.2 Robotaxi network status Tesla's driverless service operated in six metropolitan areas — Austin, Dallas, Houston, Miami, Orlando, Tampa — plus a Bay Area operation that retains a safety driver because California regulates it under a permit class that is not a driverless deployment authorization. [E] Service hours were extended in late August 2026 to 06:00–22:00, seven days a week, with the company describing an enlarged unsupervised fleet and improved dispatch routing. [E] Independent trackers estimated approximately 85 active unsupervised vehicles across the six metros. Texas filings authorized 42. Estimates vary and Tesla does not disclose the figure. [E] Between July 2025 and March 2026, NHTSA documented 17 incidents involving the Austin fleet; six were minor contact events attributed wholly or partly to the autonomous system; two involved remote teleoperators contacting objects at low speed; no major crashes were recorded. [E] For comparison, the largest US operator ran roughly 3,500 vehicles, about 500,000 paid rides per week, and on the order of four million autonomous miles per week across eleven metros. [E]

2.3 The three-order-of-magnitude gap

The economically significant scale — the fleet size at which robotaxis measurably affect national aggregates — is on the order of 10⁶ vehicles (§5.3). Current deployment is on the order of 10² vehicles.

This is the organizing fact of the paper. Three orders of magnitude separate demonstration from economic significance. The paper's structure is a decomposition of what must be paid, and in what currency, to close each order.

Gap From → To Binding currency Assessed in 10² → 10³ ~100 → ~1,000 vehicles Regulatory jurisdiction count; remote-operator ratio §4.4, §4.1 10³ → 10⁴ ~1,000 → ~10,000 Insurance underwriting of correlated risk §4.2 10⁴ → 10⁵ ~10,000 → ~100,000 Fleet capital financing; charging infrastructure §5.2, §5.4 10⁵ → 10⁶ ~100,000 → ~1,000,000 Battery cycle life; urban congestion feedback §5.1, §4.3

Note that the currency changes at every step. This is why single-constraint arguments — "it's a software problem," "it's a regulatory problem" — systematically mispredict. The binding constraint is a function of scale.

3.1 The naive subtraction

A US ride-hailing trip prices at roughly $2.00–2.50 per passenger mile. The driver's gross take is roughly $1.10–1.40 of that, before the driver's own vehicle, fuel, insurance, and depreciation costs, which they bear personally.

The naive argument subtracts the driver's take and concludes that the fare can fall to approximately $0.80–1.10 while preserving the platform's margin.

This is wrong in both directions simultaneously, and the two errors do not cancel.

3.2 Error one: the driver was also the capital

In a human ride-hailing network, the marginal vehicle costs the platform nothing. The driver finances it, insures it, maintains it, parks it, and absorbs its depreciation. The platform's balance sheet carries software and a brand.

Autonomy converts this to a capital-intensive fleet business. The platform must now buy every vehicle, build depots, install charging, and carry the depreciation. The correct comparison is therefore not fare minus driver take but fare minus driver take plus fleet capital charge plus fleet operations.

The magnitude is not small. A fleet capable of serving 10% of US vehicle miles requires on the order of $100–200 billion of vehicle capital alone (§5.2), against a US ride-hailing industry whose total annual revenue is a fraction of that. The industry's capital intensity increases by roughly an order of magnitude at the moment the driver is removed. [T]

3.3 Error two: the driver was also the shock absorber

Ride-hailing demand has a peak-to-trough ratio within a day of roughly 3:1 to 5:1, and pronounced weekly and seasonal structure.

Human networks absorb this at zero platform cost. A driver whose expected hourly earnings fall below their reservation wage simply stops driving; the marginal cost of an idle driver to the platform is zero. Surge pricing recruits latent supply within minutes.

An autonomous fleet cannot do this. An idle robotaxi still depreciates, still occupies a parking space or a charging stall, still carries insurance and financing cost. Utilization variance converts directly into cost, where previously it converted into unpaid driver time. [T]

This is the first of the four currencies. It has a specific implication that the naive model misses: an autonomous fleet is sized either for peak (and idles expensively off-peak) or for average (and loses peak demand to competitors). Neither is free. The optimal fleet is smaller than peak demand, which means robotaxi networks structurally cannot fully serve peak and will exhibit either surge pricing or wait-time rationing at peak indefinitely — not as a transitional artifact of small fleets, but as an equilibrium property. [C]

3.4 Error three: the driver was also the liability sink

Treated separately in §4.2, because it is the most underappreciated and the most quantitatively severe.

3.5 The corrected accounting identity

all expressed per paid passenger mile, with each term inflated by the deadhead ratio where it scales with total rather than paid miles. The distinction between the two brackets is the paper's principal organizing device: the first bracket sets a floor no amount of operational learning can breach; the second is where all genuine improvement occurs.

4.1 Currency one: utilization variance against capital intensity

The unit-economics model. We define a duty cycle by service window $H$ (hours/day), paid-occupancy fraction $u$, effective door-to-door speed $v$ (mph, inclusive of stops and urban delay), and deadhead ratio $d$ (empty repositioning and pickup miles as a fraction of paid miles).

Four regimes, at 355 operating days per year:

• —Regime H u v d Paid mi/day Total mi/day Paid mi/yr Total mi/yr
• —Observed 2026 12 0.40 14 0.45 67 97 23,900 34,600
• —Base 16 0.55 16 0.40 141 197 50,000 70,000
• —Mature 18 0.70 17 0.35 214 289 76,000 102,700
• —Ceiling 20 0.78 18 0.30 281 365 99,700 129,600

The Ceiling regime is not a forecast. It is an upper bound constructed to bound the argument: a vehicle serving paid passengers 78% of a twenty-hour window with only 30% deadhead is at or beyond the best plausible dispatch efficiency, and the analysis below asks what remains true even there.

The cost stack. With vehicle-plus-allocated-infrastructure capital of $35,000 (Base and Mature) or $27,000 (Ceiling), and energy consumption of 0.18 kWh/mile in urban duty:

• —Component Base Mature Ceiling
• —Capital amortization $0.140 $0.077 $0.034
• —Energy $0.030 $0.024 $0.018
• —Maintenance, tires, cleaning $0.084 $0.068 $0.052
• —Insurance $0.180 $0.100 $0.050
• —Remote operations $0.212 $0.047 $0.010
• —Depot / field operations $0.050 $0.035 $0.020
• —Software, compute, connectivity $0.030 $0.020 $0.015
• —Total per paid mile $0.726 $0.371 $0.199

The Base figure of approximately $0.73 sits in the same neighborhood as third-party estimates published in 2026, which placed Tesla's operating cost near $0.81/mile against roughly $1.43 for the leading lidar-based operator and roughly $1.71 for conventional human ride-hailing. [E] The model is therefore calibrated, not free-floating.

Remote operations is the dominant early lever. At a 1:15 operator-to-vehicle ratio, remote supervision is the largest single line in the Base stack — larger than the vehicle itself. At 1:50 it collapses to $0.047. This single ratio moves total cost by $0.165/mile, more than any other parameter in the model. [T]

This has a direct consequence for interpreting deployment news. Fleet size and city count are weak evidence about progress; the operator ratio is strong evidence. A fleet that doubles while the operator ratio holds constant has demonstrated capital availability, not autonomy. Neither Tesla nor its competitors routinely disclose this ratio, which is itself informative about which number the operators consider load-bearing. [C]

4.2 Currency two: correlation, not safety

This is the result we consider most underweighted in public discussion.

The formal structure. Consider a fleet of $N$ vehicles with per-vehicle annual liability loss $L_i$

. An insurer prices at expected loss plus a risk load proportional to the portfolio's standard deviation contribution.

Human fleet. Driver errors are approximately independent. Write 𝐿 𝑖 ∼ ( 𝜇 ℎ , 𝜎 ℎ ) L i

∼(μ h

,σ h

), iid. Portfolio standard deviation is 𝜎 ℎ 𝑁 σ h

N

, so the per-vehicle risk contribution is 𝜎 ℎ / 𝑁 → 0 σ h

/ N

→0. Premium per vehicle converges to the actuarial mean 𝜇 ℎ μ h

.

Autonomous fleet. Every vehicle runs identical software. Decompose the loss into a common factor and an idiosyncratic residual:

=μ a

+βZ+ε i

,Z∼(0,σ Z

),ε i

∼(0,σ ε

) iid

where $Z$ represents a fleet-wide event: a software regression, a perception failure on a newly common road configuration, an over-the-air update with an unanticipated interaction, a sensor supply-chain defect. Portfolio standard deviation is $N_{2_{{\beta }^{2_{{\sigma }_{Z^2}}}}}+N{\sigma }_{{ϵ}_2}$

, so per-vehicle:

⟶

𝛽 𝜎 𝑍

0 N σ portfolio

= β 2 σ Z 2

+ N σ ε 2

⟶βσ Z

0

The per-vehicle risk load does not vanish with fleet size. It converges to a positive floor. [T]

The insurance cost of an autonomous fleet is therefore lower than a human fleet only if:

𝜇 ℎ − 𝜇 𝑎 > 𝜆

𝛽

𝜎 𝑍 μ h

−μ a

λβσ Z

that is, only if the mean safety improvement exceeds the price of the correlation. [T]

What this means in practice. Four consequences follow, and none of them is about whether the software is good.

Safety improvements have diminishing insurance returns. Driving ${\mu }_a$

toward zero eventually stops helping, because the binding term is $\lambda \beta {\sigma }_Z$

, which is a function of fleet homogeneity, not of accident rate. The remedy is heterogeneity, and heterogeneity is expensive. Staged rollouts, version pinning across sub-fleets, and geographic software diversity all reduce $\beta$ — and all forfeit the fleet-learning advantage that is the central argument for the camera-only, data-scaling approach. There is a direct trade-off between the learning rate and the insurability of the fleet, and it has not, to our knowledge, been priced publicly. [C] Self-insurance does not escape the problem, it relocates it. A fleet operator retaining risk on its own balance sheet still faces $\beta {\sigma }_Z$

; it has merely moved the correlated tail from an insurer's capital to its own. For a firm whose equity valuation is itself levered to the same autonomy thesis, the correlation between the liability tail and the firm's cost of capital is positive — the worst loss arrives precisely when raising capital is most expensive. The insurance line in §4.1 is the least reliable line in the table. We modeled $0.180 → $0.100 → $0.050. The theory above says the decline should flatten, not continue. If it flattens at, say, $0.12, the Ceiling case never reaches $0.20/mile regardless of everything else.

Open problem. We have no empirical estimate of $\beta {\sigma }_Z$

for any deployed autonomous fleet, because no fleet has yet operated at a scale and duration sufficient to observe a fleet-wide correlated loss event. The 17 documented Austin incidents are far too few, and too idiosyncratic, to identify a common factor. This is a genuine open problem and it gates the 10³ → 10⁴ transition. [O]

4.3 Currency three: congestion, deadhead, and the cruising trap

The parking-versus-cruising inequality. An idle robotaxi in a dense urban core has two options: occupy paid parking, or circulate. Circulating at low speed costs energy plus wear — on the order of $0.05–0.08 per mile, or roughly $0.30–0.50 per hour at typical congested-core speeds.

Metered or garage parking in any dense US core exceeds this by a wide margin. Therefore:

cost cruise ≪ cost park

⟹

privately optimal policy is to cruise cost cruise

≪cost park

⟹privately optimal policy is to cruise

Zero-occupancy cruising is the private optimum wherever urban parking is priced above a few dollars per hour. [T] This result is not new — it is a known consequence of removing the driver's own time from the parking decision — but its economic implication is often stated too weakly. It is not a nuisance. It is a negative externality that feeds back onto the operator's own cost structure, because cruising vehicles reduce $v$, and $v$ enters the revenue equation linearly.

The congestion fixed point. Let $V$ be total urban vehicle miles and $v(V)$ the speed function, declining in $V$. Fleet capacity in paid miles is $N\cdot H\cdot u\cdot v(V)$. Falling fares raise demand; rising demand raises $V$; rising $V$ lowers $v$; lower $v$ reduces per-vehicle throughput; restoring throughput requires more vehicles, which raises $V$ further.

With a ride-hailing price elasticity around $-1$, halving the fare roughly doubles demanded miles. If that doubling depresses effective speed by 20%, the fleet required rises by a factor of $2\mathrm{/}0.8=2.5$, not 2. Congestion is a convex tax on scaling, and it is paid in fleet capital. [T]

The policy lever is unusually sharp. The entire cruising trap disappears if curb and parking prices are set below the cost of cruising — the opposite of the direction most cities have moved. Alternatively, per-mile pricing on zero-occupancy vehicle miles internalizes it directly. This is a case where the permission-structure reading yields an unusually precise specification: the constraint is not "robotaxis cause congestion," it is "zero-occupancy miles must be priced above $0.05–0.08 per mile or the private optimum fills the road." That is a number a city council can act on. [C]

4.4 Currency four: regulatory serialization

The manufacturing story and the deployment story have different mathematical forms, and the mismatch is the program's most likely near-term bottleneck.

Manufacturing at a sub-ten-second cycle time is a parallel process: one line produces units at a fixed rate, and additional lines add linearly. A single line at that cycle time and realistic overall equipment effectiveness supports on the order of two million units per year. [E]

Deployment is a serial process. Each state has its own authorization regime; the observed pattern is jurisdiction-by-jurisdiction, with California specifically requiring a permit class the company had not obtained as of the evidence cutoff, which is why the Bay Area operation retains a safety driver despite a vehicle count reported in the hundreds — larger than the entire driverless Texas fleet. [E]

Therefore the binding constraint on deployment is the count of authorized operational design domains, not the count of manufacturable vehicles. [C] A fleet's economic value is min ( vehicles , authorized demand ) min(vehicles,authorized demand), and as of the evidence cutoff the second term is smaller by orders of magnitude. This is the structural reason a company can simultaneously (a) begin production of a purpose-built robotaxi and (b) not place it in public service — the two facts are not in tension, they are what serialization looks like.

An important corollary: the FMVSS self-certification route removes a federal constraint (the 2,500-unit exemption cap) that was never the binding one. It is real and it matters for the 10⁵ → 10⁶ transition. It does nothing for the 10² → 10³ transition, which is gated by state-level driverless deployment authority. [C]

The purpose of this section is to derive limits that hold independently of any operational assumption, so that later claims can be checked against them rather than against optimism.

5.1 The battery-cycle ceiling and the $0.20/mile target

Deriving the irreducible floor. Strip the cost stack to the three terms that cannot be driven to zero by learning — capital amortization, energy, and mechanical maintenance — and set insurance, remote operations, depot labor, and software all to zero:

• —Regime Lifetime total miles Irreducible floor ($/paid mi)
• —Base (5 yr) 350,000 $0.235
• —Mature (6 yr) 616,000 $0.169
• —Ceiling (8 yr) 1,037,000 $0.122

The Base regime cannot reach $0.20/mile even if insurance, remote operations, depot labor, and software are all free. [T] The floor is $0.235. This immediately establishes that the quoted target is not a claim about operational efficiency. It is a claim about asset life.

The translation into cycles. At 0.18 kWh/mile against the disclosed 48 kWh pack:

Lifetime total miles Energy throughput Equivalent full cycles 350,000 63,000 kWh ~1,310 500,000 90,000 kWh ~1,880 616,000 111,000 kWh ~2,310 1,037,000 187,000 kWh ~3,890

The full Ceiling cost stack in §4.1 lands at $0.199 per paid mile — essentially exactly the quoted target. It gets there only by simultaneously assuming an 8-year asset life, 78% paid occupancy over a 20-hour window, a 1:200 remote-operator ratio, $0.05/mile insurance, and a $27,000 all-in vehicle-plus-infrastructure cost.

And that regime demands approximately 3,900 equivalent full cycles from the pack.

This is the paper's central quantitative claim. [T] For context, a well-managed LFP cell under gentle duty is typically warranted in the 2,000–4,000 cycle range to 70–80% retained capacity; a fleet vehicle under frequent DC fast charging, high ambient temperature in the Texas and Florida markets where deployment is concentrated, and deep daily depth-of-discharge sits at the unfavorable end of that envelope. Shallow opportunity charging helps; heat and fast charge rate hurt.

We therefore state the target's true content precisely:

$0.20 per mile is achievable if and only if the pack sustains roughly 3,900 equivalent full cycles under fleet duty, or the vehicle receives a mid-life pack replacement whose cost is small relative to $27,000, or energy consumption is materially below 0.18 kWh/mile in real urban service.

Each of these three is measurable. None is a software question. [C] The permission structure is explicit: the currency is cell degradation, and the price is roughly 3,900 cycles.

5.2 The capital-formation ceiling

For a fleet serving 320 billion paid miles per year — 10% of approximately 3.2 trillion annual US vehicle miles, the threshold commonly invoked in valuation arguments:

• —Regime Fleet required Vehicle capex @ $30k Annual energy Average power
• —Base 6.40 M $192 Bn 80.6 TWh 9.20 GW
• —Mature 4.21 M $126 Bn 77.8 TWh 8.87 GW
• —Ceiling 3.21 M $96 Bn 74.9 TWh 8.54 GW

Two observations.

First, vehicle capital of $96–192 billion is a national-scale financing problem, not a corporate one. It is roughly an order of magnitude above the annual capital expenditure of the firm most likely to attempt it. It cannot be funded from operating cash flow on any plausible timeline. It requires either sustained equity issuance or the development of a securitization market in which robotaxi cash flows are packaged as asset-backed paper — a market that does not currently exist and whose creation requires exactly the actuarial history that §4.2 says is missing. [C] The financing constraint and the insurance constraint are the same constraint viewed from two directions.

Second, note the weak dependence of energy on regime. Total energy is set by miles driven, which is set by the demand target, not by fleet efficiency. Improving utilization shrinks the fleet and the capital bill; it barely touches the electricity bill.

5.3 The grid interface

Approximately 75–81 TWh/year against US generation of roughly 4,200 TWh is on the order of 1.8–1.9% of national electricity. Average power draw of 8.5–9.2 GW.

This is significant but not exceptional — comparable to, and currently smaller than, the load growth attributed to data-center construction. More importantly, the temporal profile is favorable rather than adversarial. Ride demand peaks in the morning commute, the evening commute, and late night. Charging must therefore occur in the mid-day trough and the pre-dawn trough — precisely the periods of solar surplus and minimum system load respectively.

A robotaxi fleet is, to first order, a fleet of grid-connected batteries with strong economic incentive to charge off-peak. The correct default expectation is that robotaxi charging is grid-supportive, not grid-stressing, provided depot charging (not opportunistic public fast charging) dominates. [C] The failure mode is the opposite case: if fleets rely on public DC fast charging colocated with retail, charging follows convenience rather than price and the load lands on the evening peak. This is an infrastructure-siting question with a clean answer, and it is one of the few places in this analysis where the favorable outcome is also the cheap one.

5.4 Production versus authorization, restated numerically

At two million units per year of line capacity, the Base-regime fleet for 10% of US VMT represents 3.2 years of total dedicated production. At the observed 2026 rate of roughly 10² units deployed, it represents something else entirely.

The manufacturing capability is not the constraint and will not become the constraint before approximately 2030 under any deployment trajectory consistent with observed regulatory throughput. [C] Analysts and commentators who track production milestones as the leading indicator are tracking the wrong series.

6.1 The aggregate is small; the incidence is not

Industry employment in US taxi and limousine services — a classification that captures gig ride-hailing drivers — stood at roughly 1.56 million in 2026, having grown at approximately 5.8% compound annually over the preceding five years. [E] Separately, BLS projects overall employment of taxi drivers, shuttle drivers, and chauffeurs to grow 9% from 2025 to 2035, much faster than average, with roughly 55,900 annual openings. [E]

That BLS projection deserves attention: the official labor-market forecast for the decade in question expects the occupation to grow, not shrink. This is a live, dated, testable disagreement with the displacement thesis and we treat it as such in §8.

Aggregate driver earnings at risk, at $1.20 per paid mile across roughly 50 billion annual ride-hailing miles, are on the order of $60 billion per year, or approximately 0.20% of US GDP. [T]

Against this, the consumer-surplus gain from a fare decline from roughly $2.10 to $1.00 per mile, at unit elasticity, is approximately $69 billion per year, or 0.23% of GDP. [T]

These are the same order of magnitude. The first-order macroeconomic effect of robotaxi substitution in the ride-hailing channel is a transfer from a concentrated group of workers to a diffuse group of consumers, plus a modest genuine surplus creation of perhaps 0.03–0.05% of GDP. It is not a productivity revolution. It is a redistribution with a small efficiency dividend attached.

Anyone forecasting trillion-dollar economic effects is forecasting the private-car-ownership channel, whether or not they say so. That channel is §7.

6.2 The absorber claim

Here is the effect we consider most likely to be missed entirely.

Research on rideshare entry has found that job displacement — separation from an employer undergoing a sharp employment decline — raises the rate of entry into the taxi and limousine sector by approximately 23%, and that each additional year of rideshare platform presence in a metropolitan area increases the likelihood of entry into that sector by roughly 65% relative to the mean. New entrants are more likely to combine self-employment driving income with wage and salary income. [E]

The economic reading: gig driving functions as an automatic stabilizer. It is a low-barrier, instantly-available, hours-flexible income floor that absorbs displaced workers between formal jobs. Its value to the macroeconomy is not the wage — the wage is low — it is the speed and elasticity of absorption. It converts what would be unemployment spells into partial-income spells.

Automating ride-hailing therefore does something that headline employment figures will not capture: it removes a counter-cyclical absorption mechanism from the labor market, at the precise moment when other automation may be increasing the flow of workers who need absorbing. [C]

We flag the second-order character of this claim deliberately. If AI-driven displacement raises the inflow into transitional work while autonomy simultaneously removes the largest transitional-work sink, the two effects compound rather than offset. Whether this is quantitatively important depends on parameters nobody has estimated. But it is the kind of interaction that is invisible to any analysis that counts jobs rather than modeling flows, and it is testable (§8, F7).

6.3 Distributional geometry

The gains and losses have different shapes.

Losses are concentrated: a specific workforce, disproportionately immigrant, disproportionately without post-secondary credentials, geographically concentrated in exactly the dense metros where robotaxi deployment is economically viable first. The overlap between "best robotaxi market" and "highest driver employment density" is close to complete, by construction — both follow trip density.

Gains are diffuse and, in per-dollar terms, mildly progressive: lower fares benefit non-car-owning urban households most, and non-car-ownership correlates with lower income.

Net welfare is positive. Net distribution is ambiguous. Geographic concentration is unambiguous and severe, and it is the dimension least addressed by aggregate analysis. A metro where 2% of the workforce drives does not experience a 0.2%-of-GDP national transfer; it experiences a local labor-market shock concentrated in a few zip codes over a few years. [C]

6.4 New employment: honest accounting

Autonomy creates roles: remote operations, depot maintenance, cleaning, fleet logistics, charging infrastructure, mapping and validation, insurance and safety analytics.

But §4.1 established that the entire economic argument for autonomy requires driving the remote-operator ratio from 1:15 toward 1:50 and beyond. The new jobs are created in inverse proportion to the profitability of creating them. [T] A fleet that employs many remote operators is a fleet that has not achieved the economics that justify its existence. This is not a criticism; it is a structural statement that should discipline any forecast projecting large remote-operations employment as an offset. Such employment exists only in the regime where the business does not yet work.

Depot, cleaning, and maintenance employment does not have this property and scales with fleet size roughly linearly. It is also lower-wage and lower-flexibility than driving, and it is location-fixed rather than schedule-flexible — it does not reproduce the absorber function described in §6.2.

7.1 Where the large numbers actually live

US consumers spend on the order of $1.2 trillion annually on personal vehicles — purchase, financing, fuel, insurance, maintenance, parking. This is roughly twenty times the ride-hailing channel. Every forecast of transformative macroeconomic effect is, whether stated or not, a forecast about this number.

7.2 The two crossovers

Average all-in cost of personal vehicle ownership, at roughly $11,000 per year:

Annual mileage Average ownership cost 6,000 $1.83/mi 9,000 $1.22/mi 12,000 $0.92/mi 15,000 $0.73/mi

At a robotaxi fare of $1.00/mile, substitution is rational for households driving under roughly 11,000 miles per year and irrational above it. At $0.70/mile it becomes rational for nearly all households. The ownership-displacement thesis has a sharp price threshold near $0.70–0.75 per mile, which lies above the Mature cost of $0.371 and above the Ceiling cost of $0.199 — meaning it is financially feasible, with margin, in the mature regime. [T]

But there is a second crossover, and it is the one that binds.

A household that already owns a car does not compare robotaxi fares to average ownership cost. It compares them to perceived marginal cost — fuel and incremental wear, on the order of $0.20–0.25 per mile — because the capital is sunk and does not feel like a cost per trip. Against $0.20/mile perceived marginal cost, no robotaxi fare that covers its own costs can win.

Therefore robotaxis cannot displace an existing vehicle. They can only prevent the acquisition of the next one. [T]

Substitution operates exclusively at the ownership decision margin: household formation, vehicle replacement, a second-car decision, a young adult's first-car decision. It does not operate continuously across the installed base.

7.3 The decadal time constant

This converts the macro question into a durable-goods turnover question with a hard arithmetic answer.

The US light vehicle fleet numbers roughly 292 million units against new sales of roughly 16 million per year. Maximum annual fleet replacement is therefore approximately 5.5%, and turning over half the fleet takes approximately 9.1 years — and that is the upper bound, assuming every single replacement decision goes to robotaxi substitution, which is absurd. [T]

At a more plausible 20% substitution rate among households reaching a replacement decision in dense metros only, the penetration into total personal vehicle miles after a decade is in the low single digits of percent.

Conclusion. The macroeconomic effect of autonomous mobility through its largest channel is rate-limited by the physical turnover of a durable capital stock, and that rate limit is independent of software quality, cost per mile, regulatory speed, or manufacturing capacity. A step-change in technology produces a step-change in the price signal and a decade-long ramp in the quantity response. [T]

This is, in our view, the most robust single conclusion in the paper, because it does not depend on any contested parameter. It follows from the size of the fleet and the rate of new sales.

7.4 The parking and land channel

The same time constant governs, more severely, the land-use channel. Urban parking represents an enormous quantity of centrally located real estate. Its release is frequently cited as a major economic dividend.

Parking structures have service lives of 40–60 years. Zoning revision, structural conversion, and redevelopment financing operate on 5–15 year cycles even after demand has fallen. The land dividend is real and probably large, and essentially none of it arrives before the late 2030s. [C] Forecasts that book it earlier are mistaking a change in demand for a change in the built environment.

Stated with dates and thresholds, so that the paper can be scored rather than reinterpreted. Each is designed so that a specific observation would force a specific claim in the Ledger to be downgraded.

Prediction Resolution date Falsifies

F1 Cybercab enters public (non-employee) robotaxi revenue service in at least one metro 30 Jun 2027 If earlier than 31 Dec 2026: C-4 (regulatory serialization is binding) weakens F2 No operator publicly discloses a remote-operator-to-vehicle ratio better than 1:50 with independent verification 31 Dec 2028 T-3 (operator ratio dominates early cost) — if a 1:200 ratio is verified early, the Base stack is too pessimistic F3 Total US driverless (no safety driver) Tesla fleet remains below 10,000 vehicles 31 Dec 2027 C-6 (three-order gap closes slowly); a breach implies the 10²→10⁴ currencies were mispriced F4 Commercial auto insurance cost per robotaxi mile does not fall below $0.08 for any fleet above 5,000 vehicles 31 Dec 2029 T-5 (correlated-risk floor). A verified sub-$0.08 rate at that scale falsifies the $\beta {\sigma }_Z$

floor as economically material F5 At least one fleet-wide correlated incident occurs — a single software or sensor cause producing ≥5 same-week incidents across ≥2 metros 31 Dec 2029 If no such event occurs by then at fleets >5,000 vehicles, T-5's $\beta$ is smaller than argued F6 Published advertised robotaxi fares in any US metro do not fall below $0.70/paid mile sustained over one quarter 31 Dec 2029 T-9 (ownership crossover unreached); a breach opens the ownership channel earlier than §7 allows F7 In metros with >5% robotaxi share of ride-hail trips, the elasticity of gig-driver entry with respect to local unemployment shocks declines measurably relative to matched control metros 31 Dec 2030 C-11 (absorber removal). Requires CBSA panel data; the test is available now in principle F8 BLS employment for taxi drivers, shuttle drivers and chauffeurs is higher in 2030 than in 2026 Annual, to 2030 Direct test of §6.1's noted disagreement. If employment grows as BLS projects, the displacement thesis is wrong on timing F9 No US metro implements zero-occupancy-mile pricing or curb pricing calibrated below cruising cost 31 Dec 2029 C-8 (cruising trap). If implemented and deadhead falls measurably, the policy lever is confirmed F10 Robotaxi-backed asset securitization (ABS) issuance remains below $5 Bn cumulative 31 Dec 2029 C-10 (financing constraint binds at 10⁴→10⁵). Large early issuance means the actuarial-history gap was surmountable F11 No fleet vehicle publicly demonstrates >500,000 miles on an original battery pack with >70% capacity retention 31 Dec 2031 T-7, the central claim. A verified 700k+ mile pack would confirm the $0.20/mile regime is physically available

F11 is the decisive one. If it resolves against us, the cost floor in §5.1 moves and much of §7 accelerates.

Tags. T = theorem, identity, or arithmetic consequence of stated premises. E = empirical, sourced to disclosure or published data. C = conjecture, defensible but not established. O = open problem, not currently resolvable with available data.

ID Claim Tag Depends on Falsifier E-1 Cybercab entered continuous production April 2026 at Giga Texas; ~10² units observed by mid-2026; not in public service at cutoff E Company disclosure, trackers — E-2 Tesla driverless service in 6 metros; ~85 active unsupervised vehicles per independent trackers; Texas filings authorize 42 E Trackers, state filings — E-3 Cybercab self-certifies to FMVSS; NHTSA 2,500-unit exemption cap does not apply E Company statements — E-4 Disclosed specs: 48 kWh, 163 kW, 3,113 lb, 418 mi lab range, sub-$30k target E EPA filings — E-5 Third-party 2026 operating-cost estimates: ~$0.81/mi Tesla, ~$1.43 lidar operator, ~$1.71 human ride-hail E Analyst publication — E-6 US taxi/limousine sector employment ~1.56 M (2026); BLS projects +9% for the occupation 2025–2035 E Industry data, BLS F8 E-7 Job displacement raises rideshare-sector entry ~23%; each year of platform presence raises entry likelihood ~65% E NBER research — T-1 Removing the driver raises industry capital intensity by ~1 order of magnitude T §3.2 — T-2 Utilization variance converts to cost under autonomy where it converted to unpaid driver time under human operation T §3.3 — T-3 Remote-operator ratio is the largest single cost lever in the Base regime ($0.212 → $0.010) T §4.1 model F2 T-4 Fleet-wide idle cruising is the private optimum where parking exceeds ~$0.05–0.08/mi T §4.3 F9 T-5 Per-vehicle insurance risk load converges to $\beta {\sigma }_Z>0$ for a homogeneous AV fleet, not to zero T §4.2 derivation F4, F5 T-6 Base regime cannot reach $0.20/mi even with insurance, ops, depot, and software all set to zero (floor $0.235) T §5.1 — T-7 $0.20/mi requires ~1.04 M lifetime miles ≈ ~3,890 equivalent full cycles of the 48 kWh pack T §5.1 F11 T-8 Aggregate: ~0.20% of GDP driver income at risk vs ~0.23% consumer surplus gained — same order T §6.1 — T-9 Ownership substitution threshold ≈ $0.70–0.75/mi average; but perceived marginal cost ~$0.20/mi blocks substitution for existing owners T §7.2 F6 T-10 Max US fleet replacement ≈5.5%/yr; ≈9.1 yr to turn over half — hard decadal time constant on the ownership channel T §7.3 — T-11 New remote-operations employment scales inversely with the profitability that justifies the fleet T §6.4 — C-1 Robotaxi networks cannot fully serve peak demand in equilibrium; surge or wait-rationing is structural, not transitional C T-2 Sustained peak parity with human ride-hail C-2 There is a direct, unpriced trade-off between fleet learning rate and fleet insurability C T-5 Verified low premiums at high homogeneity C-3 Fleet size and city count are weak progress signals; operator ratio is the strong signal C T-3 F2 C-4 Binding deployment constraint is authorized ODD count, not manufacturing capacity C E-1, E-2, §5.4 F1 C-5 FMVSS self-certification relieves the 10⁵→10⁶ constraint, not the 10²→10³ constraint C E-3, C-4 Rapid multi-state driverless expansion C-6 The three-order gap closes on a currency that changes at each order C §2.3 F3 C-7 Robotaxi charging is grid-supportive under depot charging; grid-stressing only under public-fast-charge reliance C §5.3 Observed evening-peak fleet load C-8 Zero-occupancy-mile pricing above ~$0.08/mi is a sufficient policy remedy for the cruising trap C T-4 F9 C-9 Manufacturing capacity does not bind before ~2030 on any trajectory consistent with observed regulatory throughput C §5.4 Line-capacity-limited deployment before 2030 C-10 Fleet financing at 10⁵–10⁶ scale requires a securitization market gated by the same missing actuarial history as insurance C §5.2, O-1 F10 C-11 Autonomy removes a counter-cyclical labor absorber; effect compounds with AI displacement elsewhere C E-7, §6.2 F7 C-12 Losses geographically concentrated in exactly the metros where deployment is first viable, by construction C §6.3 — C-13 The urban land dividend is real but arrives no earlier than the late 2030s C §7.4 Large-scale parking conversion before 2032 O-1 $\beta {\sigma }_Z$

— the magnitude of correlated fleet liability risk — is unestimated for any deployed fleet O §4.2 F5 O-2 Real-world urban kWh/mi for the Cybercab duty cycle is undisclosed; 0.18 is an assumption the whole cost stack rests on O §5.1 Published fleet telemetry O-3 Ride-hail price elasticity at fares below $1.00/mi is outside the range of any existing estimate O §4.3, §6.1 Sustained sub-$1.00 pricing data O-4 Fleet-duty pack degradation curves under Texas/Florida ambient with frequent DC fast charge are not publicly characterized O §5.1, T-7 F11 O-5 The interaction between congestion feedback and fleet-sizing optimization has no closed-form treatment here O §4.3 —

Ledger integrity note. Of 11 T-tagged claims, all follow arithmetically from stated premises; the premises themselves are E- or O-tagged, and O-2 in particular is load-bearing for T-6, T-7, and by extension the entire §5.1 result. If real-world consumption is 0.15 kWh/mi rather than 0.18, the cycle requirement in T-7 falls from ~3,890 to ~3,240 — still demanding, but materially less so. T-7's strength is bounded by O-2's weakness, and this should be stated wherever T-7 is cited.

The model is a cartoon. It demonstrates internal consistency and identifies the structure of the problem. It does not resolve the real economics.

Specifically:

The congestion fixed point is asserted, not solved. §4.3 sets up the feedback between fleet size, VMT, and speed, and then reasons about it qualitatively with a single elasticity. A proper treatment requires a spatial equilibrium model with mode choice, and the sign of some second-order terms — particularly the substitution of robotaxi trips for transit trips versus for private car trips — is not obvious. Transit substitution would make the congestion result substantially worse; private-car substitution could make it better.

The insurance derivation assumes a single common factor. Real correlated risk in an AV fleet is probably multi-factor and geographically structured: a perception failure mode may be common to one road type in one climate rather than to the entire fleet. This makes $\beta {\sigma }_Z$

smaller than the single-factor model implies, and the direction of the error favors the optimistic case.

We have no model of competitive dynamics. Two or more operators at scale in one metro changes fare formation, utilization, and deadhead simultaneously, and the current single-operator framing is a simplification that will fail exactly when the market becomes interesting.

The labor absorber claim is the least defended and the most consequential. It rests on an empirical regularity established for the entry of rideshare driving, extrapolated to its exit. Whether the absorption mechanism disappears symmetrically when the sink closes is an assumption, not a finding.

The evidence base has a specific fragility. Fleet counts come from independent trackers, not disclosure. Cost estimates come from analysts, not audited segment reporting. Neither is adversarial to the thesis, but both are the kind of number that gets revised. Any claim in the Ledger tagged E on the basis of a tracker figure should be read as E-with-a-caveat.

The real question stays open. Whether autonomous mobility becomes a modest efficiency improvement in one service category or a reorganization of urban form depends almost entirely on whether the price falls below the ownership threshold within the window in which the vehicle fleet is turning over anyway. That is a race between a cost curve and a durable-goods clock, and this paper has characterized both without being able to call the outcome.

The driver bundled four functions; autonomy removes one and reprices three. $0.20/mile is a battery-cycle-life claim requiring ~3,900 equivalent full cycles, not a manufacturing claim. [T-7] Autonomous insurance cost per vehicle has a positive floor that fleet size does not erase, because correlated risk is not diversifiable. [T-5] The remote-operator ratio, not fleet size, is the informative progress metric. [T-3, C-3] Zero-occupancy cruising is the private optimum wherever parking is priced normally; the remedy is a specific, small, actionable price. [T-4, C-8] The ride-hail channel is a transfer of ~0.2% of GDP with a small surplus attached — not a productivity revolution. [T-8] Every trillion-dollar thesis is an ownership-channel thesis, and the ownership channel is rate-limited to ~5.5%/year by fleet turnover regardless of technology. [T-10] Autonomy may remove a counter-cyclical labor absorber at the moment other automation increases demand for one. [C-11] The binding constraint changes at each order of magnitude of scale; single-constraint arguments will systematically mispredict. [C-6]