Research brief · Autonomy · October 2026

Robotaxi:
autonomy, when?

Waymo now runs about half a million paid robotaxi rides a week, measurably safer than humans. Tesla has pushed unsupervised robotaxis from Austin into Dallas, Houston and more. Cruise is dead. We read the forecasts and the safety data to answer the real question: when does autonomy actually scale, and how far is that from a car that drives anywhere?

By Christos Papadimitriou · House of 47 Research
Last updated:

See the road
500kWaymo paid rides per week, 2025
88-92%Fewer crashes than humans, Swiss Re
6MGlobal fleet forecast by 2035, Goldman
$10BBurned before Cruise shut down
A self-driving car with sensors on a city street
The road · 2024 → 2035

Real, but geofenced

The robotaxi is no longer a demo. It is a paid service, running millions of miles, with a better safety record than human drivers. The catch is the map: it works inside the cities it has been taught, and leaves the slide deck one zone at a time.

2024-25 · Now

It works

500k/wk

Waymo runs ~500,000 paid rides a week across about 10 cities, over 56.7M autonomous miles, with 80 to 92% fewer crashes than humans. The tech is real, inside mapped zones.

2024 · The reckoning

Cruise dies

$10B

GM pulls funding from Cruise after more than $10B spent and under $500M revenue. Good technology, broken unit economics. Capital alone does not scale autonomy.

2030 · Scaling

Fleet service

~1M

Goldman projects a global fleet near 1 million and US robotaxis at about 8% of rideshare. Still city by city, still geofenced, still a service you hail rather than own.

2035 · The market

$415B market

$415B

A ~$415B market and a ~6M fleet. But S&P sees near-zero go-anywhere Level 5 even then. Autonomy arrives as a service, not as a car that drives everywhere.

The ramp, forecast and real

Global robotaxi fleet (thousands)

Cross-check · Goldman, Waymo, Swiss Re, S&P, RethinkX

What everyone agrees on

  • All
    Waymo works

    56.7M autonomous miles, 80 to 92% fewer crashes than humans (Swiss Re). The driving problem is solved, inside geofenced cities.

  • All
    It scales slowly, city by city

    HD mapping, local operations and regulation per state. This is not a software update you flip on everywhere at once.

  • All
    Robotaxi, not self-driving-your-car

    The model is a fleet service in mapped zones, not a go-anywhere Level 5 car you buy and sleep in on the motorway.

  • Most
    Economics kills faster than tech

    Cruise had working cars and still died: over $10B spent, under $500M revenue. The hard part is the unit economics, not the driving.

  • Most
    Level 5 is far

    S&P sees essentially no go-anywhere, no-geofence autonomy before 2035. Supervised and mapped stays the norm for years.

  • Most
    Two bets on the table

    Waymo's lidar-plus-maps is expensive but proven. Tesla's camera-only is cheaper and now runs unsupervised in several US cities, but still unproven at Waymo's mileage and safety record.

Where they split

How big by 2035?

From a niche in a few cities to a mass-market replacement for driving.

NicheMass market

US robotaxi share of rideshare, 2030

What fraction of US ride-hailing is driverless by 2030.

0%20%

When go-anywhere Level 5?

A car that drives anywhere, no geofence, no safety driver.

Soon2035 and beyond

Goldman Sachs 2025 (fleet 7k→1M by 2030→6M by 2035, ~$415B market) · Waymo + Swiss Re 2024 (25.3M miles, 88-92% fewer crashes) · S&P Global (near-zero Level 5 before 2035) · RethinkX 2017 (95% of US miles autonomous by 2030, the bull case that missed) · Cruise shutdown 2024 · Tesla 2026 (unsupervised in Austin, Dallas, Houston, expanding).

Probability of outcome · House of 47 synthesis

The odds

Synthesis estimates across the sources above, not market forecasts. The field moves monthly, so read these as of late 2026.

Autonomy scorecard · where it actually stands

What has to be true

Two lines are already green. The rest, the ones that decide whether autonomy goes everywhere, are still years out.

ConstraintVerdictThe number

Figures from Waymo + Swiss Re, Goldman Sachs, S&P Global and the Cruise shutdown. The binding constraints are unit economics, mapping and go-anywhere capability.

House of 47 read · the autonomy playbook

How autonomy ships

The robotaxi is the most public test of real-world AI autonomy, and its lesson is blunt: autonomy ships narrow, supervised and slow before it ships wide. The same curve governs putting AI agents to work inside a business, which is the work of theagency47.

Start geofencedbound the scope

Waymo won by mastering one mapped city at a time. For a business AI agent, the move is the same: narrow and define the scope before you widen it.

Keep a human in the loopsupervised first

Every robotaxi started with a safety driver. Autonomy earns trust on data, not promises. Supervised autonomy is the on-ramp, not a failure.

The last 10% is the businessdemo vs production

A demo that works once is easy. Reliable, unsupervised operation across the edge cases is years of work, and it is where the value sits.

Own the operationnot the pilot

Cruise had great technology and died on economics. The winners run the system reliably at scale. Own the operation, not the proof of concept.

The parallel

Lidar-plus-maps versus camera-only is the same reliability-versus-cost tradeoff you face with any AI agent. The reliable, well-mapped option costs more up front and wins on trust.

Why it matters here

Deploying agents in a business follows the robotaxi curve exactly: geofence, supervise, harden, scale. Skipping a step is how you become Cruise.

The car that drives itself is here. The car that drives anywhere is not. Autonomy arrives as an operation, one mapped city at a time.

Sources · figures verified against primary reports

Sources

Changelog · what changed in this brief, newest first

Changelog