# Waymo's 17-year detour was the moat: Dolgov on the foundation model behind 20 million rides

URL: https://www.thedeepfeed.ai/posts/2026-05-01-waymo-foundation-model-20m-rides/
Category: Products
Published: 2026-05-08
Author: the-deep-feed
Tags: waymo, autonomous-vehicles, foundation-models, robotics, ai-ascent, sequoia
Kind: deep

> Dmitri Dolgov's AI Ascent 2026 fireside laid out the multimodal world-action-language model, the driver-simulator-critic stack, and why the 13x safety number is the bill that competitors haven't paid yet.

## TL;DR

- Dolgov disclosed the cumulative scoreboard on stage: **over 20 million** fully autonomous Waymo rides, with **10 million** of those in the last **seven months** alone — a doubling that took five years before, and now takes a quarter of a year.
- The technical core is the **Waymo foundation model** — described as a multimodal *world-action-language* model that powers three distinct pillars: the driver, the simulator, and the critic. End-to-end is the substrate; *structured materialized intermediate representations* sit on top.
- Waymo's March 2026 safety report covers **170 million fully autonomous miles** and shows **92% fewer crashes** with serious or fatal injuries vs. human benchmarks — the **13x** number Dolgov quoted converts to *one prevented serious injury every eight days*.
- The phase transition is operational, not just technical: **eight years** to launch service in four cities sequentially, then **four cities in one day** on Feb 24, 2026 (Dallas, Houston, San Antonio, Orlando). Fleet now does **~4 million autonomous miles per week** at **~500K rides/week**, on a path to **1M/week** by year-end.
- The competitive scoreboard is now binary on Level 4 in the U.S.: Waymo runs the only at-scale rider-only service after Cruise's shutdown; Tesla's "robotaxi" still uses safety monitors; Zoox is geofenced and free; Wayve is pre-commercial. Apollo Go matches volume in China but on different roads, different rules.

![Waymo's 17-year detour visualized as the long road that became the moat.](/post-images/2026-05-01-waymo-foundation-model-20m-rides/hero-20m-rides.jpg)

The room at AI Ascent had been told to raise hands twice. First: how many of you have ridden in a Waymo? Most of the room. Then: how many of you *love* the experience? Same hands, plus a few more. The interviewer let the show of hands speak (twenty seconds of social proof for what used to be an R&D demo) and then asked Dmitri Dolgov to explain how his company had outlasted everyone who had tried to take a shortcut to the same destination.

The answer, condensed: **Waymo** built the foundation model that the rest of the autonomous-vehicle industry now wishes it had data and compute and patience to build, and the 17-year grind through HD maps, lidar, structured representations, and a rider-only service in Phoenix is precisely the asset that makes the model possible. The detour was the moat.

# The numbers Dolgov put on the board

![Waymo's 20M cumulative rides curve, with the 10M-in-7-months acceleration as the steep right-side jump.](/post-images/2026-05-01-waymo-foundation-model-20m-rides/rides-doubling-curve.jpg)

Two minutes of the 25-minute fireside carried more signal than the rest of it combined. The interviewer was reading from a deck the audience could not see. Dolgov filled in the rest:

> "It took us eight years from the day when we started our fully autonomous operations to the day when we had our service, our driver, providing rides to the public in four cities. Earlier this year, just a few weeks ago, we launched four cities in one day. We've given over 20 million fully autonomous rides. 10 of those million happened in the last seven months."

Each clause is a checkable claim. Each one checks out.

The four-cities-in-one-day refers to **Feb 24, 2026**, when Waymo opened public service in Dallas, Houston, San Antonio, and Orlando — bringing the public-service total to 10 metros and marking the first multi-city public launch in the company's history ([Waymo blog](https://waymo.com/blog/2026/02/dallas-houston-san-antonio-orlando), [TechCrunch](https://techcrunch.com/2026/02/24/waymo-robotaxis-are-now-operating-in-10-us-cities)). The 20-million-rides milestone was [announced Dec 17, 2025](https://cleantechnica.com/2025/12/17/waymo-reaches-20-million-passenger-trips/), four months before AI Ascent. The 10-million-in-seven-months math reverses cleanly against the May 2025 disclosure of [10 million cumulative trips](https://cnbc.com/2025/05/20/waymo-ceo-tekedra-mawakana-10-million.html), giving an annualized doubling rate that no other Level 4 operator outside China is currently capable of.

Three more numbers, dropped later in the conversation:

- **4 million autonomous miles per week**: confirmed in a Stripe-hosted interview with Dolgov in late March 2026, where he also pegged the company at **~500,000 paid rides per week** across 10 cities ([Sherwood](https://sherwood.news/tech/waymos-now-serving-more-than-500-000-paid-robotaxi-rides-every-week/)).
- **1 million paid rides per week target** by end of 2026: co-CEO Tekedra Mawakana's [Bloomberg target from Feb 2026](https://www.bloomberg.com/news/articles/2026-02-11/waymo-co-ceo-outlines-path-to-1-million-weekly-trips-in-2026).
- **13x safer than a human driver** on serious-injury collisions: derived from Waymo's [March 19, 2026 safety report](https://waymo.com/blog/shorts/waymo-safety-impact-update-170m/) covering 170 million fully autonomous miles, which shows 92% fewer crashes with serious or fatal injuries (the reciprocal of `1 − 0.92` is 12.5; Waymo rounds to 13).

Dolgov took the 13x number a step further on stage: at current scale, that gap means Waymo prevents *a serious injury every eight days*. The frame is not "robotaxi is fun." The frame is "every fortnight a person who would have been hospitalized walks home."

# The AV scoreboard, Q2 2026

The shape of the field around Waymo has changed twice in 18 months. Cruise was shut down by GM in [December 2024](https://techcrunch.com/2024/12/11/gm-is-giving-up-on-cruise-robotaxis-pivots-to-personal-autonomous-vehicles/), redirecting its budget toward driver-assist on personal vehicles. Tesla launched its Austin "robotaxi" in [June 2025](https://techcrunch.com/2025/06/22/tesla-launches-robotaxi-rides-in-austin-with-big-promises-and-unanswered-questions) with safety monitors in the front passenger seat, and its Bay Area expansion is a [chauffeured service with humans behind the wheel](https://www.reuters.com/business/autos-transportation/tesla-roll-out-human-driven-chauffeur-service-bay-area-california-regulator-says-2025-07-25/). Zoox finally [opened to public riders in Las Vegas on Sept 10, 2025](https://techcrunch.com/2025/09/10/zoox-opens-its-las-vegas-robotaxi-service-to-the-public/) — free, geofenced to a strip of approved pickup points, no public ride-count disclosed.

The result is a U.S. market where one company runs the rider-only, paid, scaling service; everyone else is at an earlier stage of the same curve.

| Operator | Cumulative rides | Weekly rides (latest) | Cities (public) | Status |
|---|---|---|---|---|
| **Waymo** | 20M+ (Dec 2025) | ~500K (Mar 2026) | 10 US metros + Tokyo/London prep | Rider-only, paid, scaling |
| **Apollo Go** (Baidu) | 17M (Nov 2025); 20M+ (Feb 2026) | 250K–300K | 22+ cities, mostly China + Dubai | Rider-only in select zones |
| **Tesla Robotaxi** | Not disclosed | Not disclosed | Austin (rider-only zone), SF Bay (chauffeured) | Safety monitor / human driver |
| **Zoox** | Not disclosed | Not disclosed | Las Vegas (free, geofenced); SF testing | Public preview, no fares |
| **Pony.ai** | 680+ vehicles; Gen-7 commercial in Guangzhou, Shenzhen, Beijing | Not disclosed | 4 Tier-1 China cities | Driverless commercial; HK dual-listed |
| **WeRide** | 8 countries permitted; UAE driverless permit | Not disclosed | UAE, China, Singapore | Mixed; revenue +144% YoY Q3 2025 |
| **Cruise** | Discontinued | n/a | n/a | Shut down Dec 2024 |
| **Wayve** | Pre-commercial | n/a | n/a | $1.2B Series D Feb 2026, no public service |

Sources: Waymo, [CNBC](https://cnbc.com/2025/05/20/waymo-ceo-tekedra-mawakana-10-million.html), [carnewschina](https://carnewschina.com/2025/11/13/baidus-apollo-go-robotaxi-leads-global-autonomous-driving-with-17m-orders-targets-profit-this-year/), [Pony.ai IR](https://ir.pony.ai/node/7556/pdf), [WeRide IR](https://ir.weride.ai/node/7986/pdf), [TechCrunch (Wayve)](https://techcrunch.com/2026/02/24/self-driving-tech-startup-wayve-raises-1-8b-from-nvidia-uber-and-three-automakers/).

The chart explains why **Sequoia, Dragoneer, and DST** led a [$16 billion round at a $126 billion post-money in February](https://waymo.com/blog/2026/02/waymo-raises-usd16-billion-investment-round/). Capital follows the only operator with a working flywheel.

# Dolgov's "foundation model," in plain terms

The intellectual core of the talk is also the part where the interviewer asked the right question and stepped out of the way. The exchange:

> "A lot of people are talking about world models. You have had all the components of world models for many years. How do you think about a world model and what is Waymo's version of a world model?"

![Waymo's driver-simulator-critic stack as the EMMA-shaped substrate mapping perception, sim, and reality.](/post-images/2026-05-01-waymo-foundation-model-20m-rides/world-model-stack.jpg)

Dolgov's answer was unusually specific for a fireside. The Waymo foundation model is, in his words, **a multimodal world-action-language model**. Each adjective is doing work:

- **Multimodal**: it ingests cameras, lidars, and radars together, not as a fused set of post-perception tokens but as native inputs. This is the part Tesla cannot copy without changing its sensor stack. Cameras-only is a fine starting point; it is not a foundation that can promote into a *world* model with depth and dynamics, which is why [Tesla's vision-only stack runs into edge cases lidar would catch](https://electrek.co/2025/03/23/everyones-missing-the-point-of-the-tesla-vision-vs-lidar-wile-e-coyote-video/).
- **World-action**: it doesn't just predict the world, it predicts the world *conditional on the agent's action*. This is the controllability property that simulator architectures have spent a decade trying to bolt on. Dolgov: "We don't only have to ... the world model has to be controllable but also we need to have a deep understanding of what it means to be a good agent in that world."
- **Language**: the model is aligned with a VLM so it can pull in the general world knowledge of a frontier language model. This is the EMMA-shaped substrate. Waymo's [EMMA paper](https://arxiv.org/abs/2410.23262), built on Gemini and published in [TMLR in July 2025](https://openreview.net/forum?id=kH3t5lmOU8), demonstrated that mapping raw camera input to planner trajectories via an LLM backbone produced first-place results on nuScenes and Waymo Open Motion Dataset motion-planning benchmarks while [keeping the system natively explainable in natural language](https://waymo.com/blog/2024/10/introducing-emma).

The architecture deploys this single foundation model into three runtime roles:

1. **The driver**: what plans the trajectory the car follows.
2. **The simulator**: what generates synthetic scenarios for training and evaluation. As of [February 2026 this is the Waymo World Model, built on Google DeepMind's Genie 3](https://waymo.com/blog/2026/02/the-waymo-world-model-a-new-frontier-for-autonomous-driving-simulation), trained to produce camera *and* lidar from learned priors so the simulator output stays sensor-faithful. The blog post makes the architectural claim explicit: simulation is one of three pillars of Waymo's [demonstrably safe AI](https://waymo.com/blog/2025/12/demonstrably-safe-ai-for-autonomous-driving) ecosystem.
3. **The critic**: what scores the driver's behavior in closed-loop training and runtime validation.

The driver-simulator-critic triad is RL infrastructure with a self-driving body. It is also the same architectural insight that runs through this conference: a base model that can be re-deployed as multiple specialists with different reward functions and different decoding strategies. [Karpathy's Software 3.0 framing](/posts/2026-05-01-karpathy-software-3-agentic-engineering/) calls this the "spiky generalist" pattern. [Jim Fan's robotics talk](/posts/2026-05-01-jim-fan-robotics-llm-playbook/) ports the same playbook into humanoid manipulation. Dolgov's contribution is to demonstrate that it works in production at safety-critical scale.

# Where the "shortcut" debate ends

The most quotable architectural claim in the talk is about a debate that has consumed AV Twitter for two years. The pre-2024 framing: end-to-end (Tesla, Wayve, comma.ai) vs. modular (everyone else). Dolgov's reframing:

> "There's a false dichotomy there that's, you know, end-to-end or something else. Really, it — my mind has always been the question of, you know, it's end-to-end *and then what else*?"

He continued, in the most direct words a Waymo executive has used in public on the subject:

> "It turns out that the basic vanilla end-to-end system is insufficient... there's a massive difference between using end-to-end versus purely relying on it."

What Waymo built on top of vanilla end-to-end, per Dolgov: **structured materialized intermediate representations** that enable closed-loop evaluation, closed-loop training, rich reward functions for reinforcement learning, and runtime validation of the agent in the physical world. The argument is not that pure pixel-to-trajectory neural nets cannot drive a car. The argument is that they cannot be *demonstrated* to drive at superhuman safety — and without demonstrability, there is no rider-only service, no insurer underwriting, no four-cities-in-one-day. The intermediate representations are the apparatus that lets Waymo write [56.7-million-mile peer-reviewed crash-rate papers in *Traffic Injury Prevention*](https://waymo.com/research/comparison-of-waymo-rider-only-crash-rates-by-crash-type-to-human-benchmarks) and [25.3-million-mile auto-liability comparisons with Swiss Re](https://waymo.com/research/do-autonomous-vehicles-outperform-latest-generation-human-driven-vehicles-25-million-miles).

This is the gap that nobody else is paying for. A snapshot of where the industry sits architecturally in Q2 2026:

| Company | End-to-end model | Structured intermediates | Lidar | Validated by | Peer-reviewed safety reports |
|---|---|---|---|---|---|
| **Waymo** | Foundation model (driver/sim/critic) | Yes (explicit, materialized) | Yes (4) + 6 radar + 13 cameras (Gen 6) | Swiss Re, NHTSA, *Traffic Injury Prevention* | Multiple, ongoing |
| **Tesla** | FSD v13/v14 (vision-only) | No (publicly stated) | No | Internal disengagement data | None peer-reviewed |
| **Wayve** | AV2.0 / GAIA world model | Limited (research stage) | Optional, automaker-dependent | Pre-commercial | Research papers, no L4 deployment |
| **Apollo Go** | Modular; integrating Apollo Foundation Model | Yes | Yes | Internal + Chinese regulator | Not peer-reviewed in English |
| **Zoox** | Modular | Yes | Yes | Internal + NHTSA filings | Limited |
| **Pony.ai** | Gen-7 unified model | Yes (HD maps + intermediates) | Yes | Tier-1 city regulators | None peer-reviewed |
| **Cruise** | Discontinued | n/a | n/a | n/a | n/a |

Sources: company technical disclosures, [Wayve research](https://wayve.ai/thinking/the-path-to-autonomy/), [Waymo 6th-gen Driver](https://waymo.com/blog/2024/08/meet-the-6th-generation-waymo-driver), [Pony.ai prospectus](https://www.hkexnews.hk/listedco/listconews/sehk/2025/1028/2025102800087.pdf).

The pattern: every operator that survived the AV winter is converging on a stack that looks more like Waymo's, not less. Apollo Go's Apollo Foundation Model. Pony's Gen-7 unified architecture. Wayve's GAIA-2 world model paired with its [$1.2 billion Series D from NVIDIA, Uber, Microsoft, Nissan, Mercedes-Benz, and Stellantis](https://techcrunch.com/2026/02/24/self-driving-tech-startup-wayve-raises-1-8b-from-nvidia-uber-and-three-automakers/) which explicitly funds end-to-end *with* structured supervision. Even Tesla's leadership has [migrated FSD v14 toward a more explicitly multi-modal stack with an end-to-end planner head](https://www.businessinsider.com/tesla-vs-waymo-what-ceos-eautonomous-automotive-leaders-say-2025-8), though the sensor question remains unresolved.

The argument Tesla bulls have run since 2020, that Waymo's HD maps and lidar were a temporary crutch that would be retired by superior neural nets, has not aged well. What Waymo has done instead is keep the lidar, keep the maps, *and* train a foundation model on the resulting data. Lidar didn't get replaced. It got tokenized.

# Why the safety bus story matters more than it sounds

Mid-talk Dolgov told a story that sounded like a feel-good aside but is the load-bearing claim of the entire architecture pitch:

> "The Waymo driver was at an intersection. There was a bus that crossed and stopped partially blocking the intersection. Our light turned green. The Waymo driver started to proceed and as it's proceeding, it detects a pedestrian on the other side of the bus. And, you know, you can't see through the bus... it's not through lidars, not radars, not cameras."

What had happened: the lidar return was bouncing under the bus and picking up sparse signal from a pedestrian's feet. The model, trained on enough such cases to recognize the pattern, inferred a pedestrian, predicted future motion, and reacted defensively. The pedestrian then stepped out from behind the bus.

This is the kind of inference a vision-only camera stack cannot make, regardless of model scale. It is also the kind of behavior that emerges from a *world-action* model trained on a dataset most of whose miles include lidar. The story matters because it's the cleanest public example of why Dolgov's "end-to-end and what else" formulation is not academic. The "and what else" includes a sensor that produces information no camera can recover, processed by a model that has learned to read sparse foot-shaped lidar returns as people. None of that emerges from a 100-mile San Francisco demo. It emerges from 170 million miles of fleet data with the same sensor suite.

Which loops back to why Cruise died and Tesla is still using safety monitors. Cruise's [October 2023 incident](https://www.npr.org/2024/12/11/g-s1-37700/gm-to-retreat-from-robotaxis-and-stop-funding-its-cruise-autonomous-vehicle-unit) was not a foundation-model failure; it was an edge-case behavior failure that the system had not been validated against. Tesla's robotaxi in Austin still has a human in the front passenger seat with a kill switch six months after launch because the closed-loop validation Waymo runs on its 170M-mile dataset is the closed-loop validation Tesla [does not currently disclose running](https://www.businesswire.com/news/home/20240829768166/en/Wayve-and-Uber-Partner-to-Accelerate-the-Future-of-Automated-Driving) at comparable scale. The bill the foundation model claim implies (peer-reviewed crash rates, sensor diversity, 170M miles of operational data) is the bill nobody else has paid yet.

# The phase transition is operational, not just architectural

![Waymo's Feb 24, 2026 four-cities-in-one-day launch — Dallas, Houston, San Antonio, Orlando.](/post-images/2026-05-01-waymo-foundation-model-20m-rides/four-cities-launch.jpg)

The interviewer's framing of Waymo's history ("16 years to 100 million miles, six months to 200") is the right shape of the story but understates what changed. The bottleneck shifted. Until 2024, the bottleneck was the driver: was the system capable enough to operate without a safety driver in this neighborhood, on this kind of road, in this weather. After 2024, with the foundation model generalizing across cities, the bottleneck moved to operations and trust: fleet ops, charging, insurance, regulator conversations, ride-hail demand-supply matching, customer support.

Dolgov: "more often than not today we're seeing that the driver is generalizing incredibly well and it's just a matter of high-fidelity rigorous evaluation and validation before we deploy."

Which is why the [sixth-generation Waymo Driver matters more as a cost story than as a sensor story](https://waymo.com/blog/2024/08/meet-the-6th-generation-waymo-driver). Gen 6 (13 cameras, 4 lidars, 6 radars) is described by Waymo as *significantly cheaper* than Gen 5 while delivering more resolution, range, and compute. It rolls onto the [Hyundai IONIQ 5 platform](https://waymo.com/blog/2024/10/waymo-and-hyundai-enter-partnership) and went into [fully autonomous operations in February 2026](https://blog.waymo.com/blog/2026/02/ro-on-6th-gen-waymo-driver). The unit economics don't close at six-figure-per-vehicle hardware. They close at Gen-6 prices on a Hyundai-assembled platform with Tier-1 supply chains.

Internationally, [London is targeted for Q4 2026](https://waymo.com/blog/2025/10/hello-london-your-waymo-ride-is-arriving) (testing began in April 2026) and Tokyo is in pre-deployment ([Tokyo media event March 2026](https://waymo.com/blog/shorts/waymo-recently-hosted-an-event-in-tokyo-to-share-more-about-our-autonomous/)). Both are right-hand-drive markets, which is the test of whether the foundation model generalizes or whether it has memorized US road semantics. If it generalizes, Apollo Go's domestic-only moat becomes a liability rather than a hedge. If it doesn't, a 17-year detour suddenly looks shorter than it should.

# What the talk admits, by what it doesn't say

Dolgov is not doing a victory lap. He is asking the room (half of it founders, half of it investors) to internalize a specific lesson about safety-critical AI products that the LLM-app generation has not yet absorbed: **how you go about the first 90% is a totally different problem from how you go about the next nines.** The post-90% regime is where everyone except Waymo currently sits. It is where Tesla still has safety monitors. It is where Cruise died. It is where Zoox is geofenced. It is where Wayve is pre-commercial despite a $1.2B round.

What the talk does not say, but the architecture implies: the foundation model is downstream of the data. The data is downstream of the fleet. The fleet is downstream of the patient capital that took 16 years to get to 100M miles. **And patient capital is downstream of being inside Alphabet.** Waymo is the rare AV company whose owner could afford the wait. Cruise wasn't, and got cut. Wayve isn't yet, and is racing. Tesla is, but chose camera-only and a different bet. Apollo Go has Baidu's domestic moat but a permitting environment that does not yet generalize.

The pitch to the founders in the room, implicit in everything Dolgov said, is that *foundation models in safety-critical domains are won by whoever owns the data flywheel that nobody can shortcut into existence*. The tax codes haven't changed. The road rules haven't changed. The 1.19 million human drivers killed on the world's roads each year haven't changed. What changed is that one company finally has the model, the sensors, the validation apparatus, and the patience to compound on all four at once.

# Where this leaves the field

![The 17-year detour as the closed-loop moat competitors haven't paid into existence yet.](/post-images/2026-05-01-waymo-foundation-model-20m-rides/detour-as-moat.jpg)

Waymo's foundation-model architecture is a deeper story than the rides number, and the rides number is the deepest story in commercial AI right now. The ride exponential is real. The 13x safety claim is documented. The four-cities-in-one-day cadence is the physical-world version of `kubectl apply -f new-region.yaml`. The world-action-language model is doing what every other foundation model in 2026 is also doing (collapsing previously separate systems into one differentiable substrate), but it's doing it in a domain where the cost of being wrong is somebody's life and the regulator's tolerance for hand-waving is zero.

The position to take: **the AV race outside China is over for any operator without a foundation model trained on multimodal sensor data validated to peer-reviewed crash-rate standards, and the list of operators who clear that bar is one long.** Apollo Go is the China-shaped second entrant. Pony.ai and WeRide are running the same playbook with different geographies and capital structures. Wayve is the only Western challenger with a credible architectural answer, but it has not deployed at scale, and the foundation-model gap to Waymo is real even before you discount for sensor stack. Tesla's robotaxi is a different product — a vision-only, lower-cost, larger-fleet thesis that may eventually scale but is currently nowhere near the operational regime Waymo is in. Cruise is gone. Zoox is on training wheels.

The interesting question is not whether Waymo has a moat. It is whether the moat is durable enough that the next wave of foundation-model improvements arrives at Waymo first because Waymo has the largest, most diverse, most validated multimodal driving dataset on Earth — and whether competitors can survive long enough to build their own. On current evidence, the answer is yes to the first and not obviously yes to the second.

Dolgov's "end-to-end and what else" framing will probably be the line from AI Ascent 2026 that ages best. Three years from now, when one or another founder is explaining why their pure-vision agent stack hit a wall the labs warned them about, somebody in the room is going to say it back to them. The detour was the moat. The shortcut was the trap. Waymo just spent a decade and change demonstrating which was which.

## Sources

- [Sequoia Capital — Dmitri Dolgov fireside (AI Ascent 2026)](https://youtu.be/I_0Kuf6Aa2c)
- [Waymo blog — 20 million paid rides (Dec 17, 2025)](https://cleantechnica.com/2025/12/17/waymo-reaches-20-million-passenger-trips/)
- [CNBC — Waymo at 10M trips, doubled in five months (May 20, 2025)](https://cnbc.com/2025/05/20/waymo-ceo-tekedra-mawakana-10-million.html)
- [Waymo — Ready to Ride: Dallas, Houston, San Antonio, Orlando (Feb 24, 2026)](https://waymo.com/blog/2026/02/dallas-houston-san-antonio-orlando)
- [TechCrunch — Waymo robotaxis now in 10 US cities (Feb 24, 2026)](https://techcrunch.com/2026/02/24/waymo-robotaxis-are-now-operating-in-10-us-cities)
- [Waymo — Safety Impact Update at 170M autonomous miles (Mar 19, 2026)](https://waymo.com/blog/shorts/waymo-safety-impact-update-170m/)
- [The Verge — Waymo at 170M miles, 92% fewer serious-injury crashes](https://theverge.com/transportation/896837/waymo-170-million-miles-safety-crashes-injuries)
- [Waymo / Swiss Re — auto liability claims at 25.3M autonomous miles](https://waymo.com/research/do-autonomous-vehicles-outperform-latest-generation-human-driven-vehicles-25-million-miles)
- [EMMA: End-to-End Multimodal Model for Autonomous Driving (arXiv 2410.23262)](https://arxiv.org/abs/2410.23262)
- [Waymo — Introducing EMMA (Oct 2024)](https://waymo.com/blog/2024/10/introducing-emma)
- [Waymo — Demonstrably Safe AI for Autonomous Driving (Dec 2025)](https://waymo.com/blog/2025/12/demonstrably-safe-ai-for-autonomous-driving)
- [Waymo — The Waymo World Model (Feb 6, 2026)](https://waymo.com/blog/2026/02/the-waymo-world-model-a-new-frontier-for-autonomous-driving-simulation)
- [Waymo — sixth-generation Waymo Driver (Aug 2024)](https://waymo.com/blog/2024/08/meet-the-6th-generation-waymo-driver)
- [Waymo — beginning fully autonomous operations on 6th-gen Driver (Feb 2026)](https://blog.waymo.com/blog/2026/02/ro-on-6th-gen-waymo-driver)
- [Waymo — $16B funding round at $126B valuation (Feb 2, 2026)](https://waymo.com/blog/2026/02/waymo-raises-usd16-billion-investment-round/)
- [TechCrunch — Waymo $16B raise (Feb 2, 2026)](https://techcrunch.com/2026/02/02/waymo-raises-16-billion-round-to-scale-robotaxi-fleet-london-tokyo)
- [Bloomberg — Waymo on path to 1M weekly rides in 2026](https://www.bloomberg.com/news/articles/2026-02-11/waymo-co-ceo-outlines-path-to-1-million-weekly-trips-in-2026)
- [Sherwood News — Waymo at 500K paid rides per week (Mar 30, 2026)](https://sherwood.news/tech/waymos-now-serving-more-than-500-000-paid-robotaxi-rides-every-week/)
- [TechCrunch — GM kills Cruise robotaxi (Dec 11, 2024)](https://techcrunch.com/2024/12/11/gm-is-giving-up-on-cruise-robotaxis-pivots-to-personal-autonomous-vehicles/)
- [TechCrunch — Tesla launches robotaxi rides in Austin (Jun 22, 2025)](https://techcrunch.com/2025/06/22/tesla-launches-robotaxi-rides-in-austin-with-big-promises-and-unanswered-questions)
- [Reuters — Tesla rolls out human-driven Bay Area chauffeur (Jul 25, 2025)](https://www.reuters.com/business/autos-transportation/tesla-roll-out-human-driven-chauffeur-service-bay-area-california-regulator-says-2025-07-25/)
- [TechCrunch — Zoox opens Las Vegas robotaxi service to public (Sep 10, 2025)](https://techcrunch.com/2025/09/10/zoox-opens-its-las-vegas-robotaxi-service-to-the-public/)
- [TechCrunch — Wayve raises $1.2B from NVIDIA, Uber, automakers (Feb 24, 2026)](https://techcrunch.com/2026/02/24/self-driving-tech-startup-wayve-raises-1-8b-from-nvidia-uber-and-three-automakers/)
- [Apollo Go — 17M cumulative rides, 250K/week (Nov 2025)](https://carnewschina.com/2025/11/13/baidus-apollo-go-robotaxi-leads-global-autonomous-driving-with-17m-orders-targets-profit-this-year/)
- [Pony.ai — Q3 2025 update, Gen-7 robotaxi launch, HK dual listing](https://ir.pony.ai/node/7556/pdf)

---

Canonical: https://www.thedeepfeed.ai/posts/2026-05-01-waymo-foundation-model-20m-rides/
Site: https://www.thedeepfeed.ai
Full corpus: https://www.thedeepfeed.ai/llms-full.txt