The autonomous vehicle industry confronts an uncomfortable paradox. The scenarios that matter most for safety—tornadoes, darting wildlife, wildfire smoke choking visibility—happen so rarely that testing for them in the real world would demand an absurd calculus: billions of miles, decades of driving, fleets that dwarf what any company could practically deploy.
Waymo thinks it's found a workaround.
On February 6, the Alphabet-backed robotaxi operator introduced what it calls the "Waymo World Model," a generative AI simulator capable of conjuring virtually any driving scenario engineers can imagine. Built atop Google DeepMind's Genie 3 foundation model and tailored for the peculiarities of autonomous driving, the system generates both camera and lidar data simultaneously—a technical feat that matters more than it might sound. Engineers can control it three ways: through driving actions, scene layouts, or simply by typing what they want in plain language. Need to see how your software handles an elephant ambling onto a Phoenix freeway at dusk? Type it in.
"Nearly 200 million fully autonomous miles" on public roads, Waymo noted in its announcement. An impressive tally, certainly. But buried in the same release: "billions of miles in virtual worlds." That ratio tells you everything about where modern AV development is heading.
Everyone's Building a Dream Machine
Waymo isn't pioneering this alone, though its pedigree—and Google's deep pockets—give it certain advantages. The sector-wide lurch toward AI-powered simulation suggests the industry has collectively decided that real-world testing, while essential, can't scale fast enough to meet deployment timelines investors are demanding.
Consider UK-based Wayve, which released GAIA-2 in March 2025. The system's a controllable multi-view latent-diffusion world model—jargon that boils down to this: it generates high-resolution, spatiotemporally consistent multi-camera videos that account for ego dynamics, other agents, environmental factors, and road semantics. By October 2023, its predecessor had already scaled to 9 billion parameters. The company explicitly designed GAIA-2 for simulating rare scenarios at scale, the same problem Waymo's tackling.
Canadian startup Waabi positions its Waabi World simulator as "powered by generative AI," featuring auto-built digital twins and automatic scenario creation. Its Copilot4D foundation model handles 3D spatial data plus temporal dynamics—essential for capturing how a real driving situation unfolds across time, not just space.
Then there's NVIDIA, which launched its Cosmos platform featuring world foundation models alongside something called "Cosmos Transfer." The capability converts structured inputs—segmentation maps, depth data, lidar, trajectories—into photorealistic video. Early adopters include Foretellix, Wayve (using it for evaluation), and Uber. NVIDIA's Omniverse blueprints now offer sensor RTX APIs for AV synthetic data generation. Even CARLA, the open-source simulator with a 150,000-developer base, integrated Cosmos Transfer in its latest release.
The money follows the momentum. Research & Markets projects the autonomous vehicle simulation market will grow from $1.4 billion in 2025 to $7.3 billion by 2034—a compound annual growth rate north of 20 percent.
The Math That Makes It Necessary
Here's the brutal arithmetic: A widely cited RAND Corporation analysis calculated that proving autonomous vehicle reliability through on-road testing alone would require billions of miles. Not millions. Billions. That's impractical at scale and impossibly slow for companies trying to iterate software on anything resembling a useful timeline.
Waymo's World Model tackles this through three forms of controllability, each serving a distinct validation purpose. Driving-action control lets engineers specify exact maneuvers to test decision-making logic. Scene-layout control enables precise placement of objects, vehicles, environmental conditions—the kind of surgical scenario design that's impossible to orchestrate reliably on public roads. Language control, perhaps the most striking feature, allows engineers to describe scenarios in plain text and watch the simulator generate them. It's not quite science fiction, but it's close.
The system also creates a flywheel effect. Ordinary dashcam footage from real-world driving logs becomes raw material for infinite synthetic variations, stress-testing planning algorithms against scenarios that are grounded in reality but tweaked to explore edge cases.
This arrives as Waymo scales aggressively in the physical world—more than 200,000 paid rides weekly across Los Angeles, San Francisco, and Phoenix, including freeway operations that launched in November 2025. Regulatory approvals keep expanding: California authorizations now cover broader Bay Area and Southern California territory. San Diego launches are planned for mid-2026. London pilots are scheduled for supervised operations, though "supervised" does a lot of work in that sentence.
A December 2024 Swiss Re study analyzing 25.3 million fully driverless Waymo miles found 88 percent fewer property damage claims and 92 percent fewer bodily injury claims compared to human-driven vehicles equipped with advanced safety features. Later peer-reviewed analysis extended this across 56.7 million miles, though independent researchers continue debating methodology and statistical significance.
Reality Intrudes
Not everyone's scaling, though. GM halted funding for Cruise's robotaxi development in December 2024, effectively abandoning the ride-hail business after years of heavy investment. Cruise absorbed roughly 50 percent layoffs and got folded back into the parent company—a humbling reversal for what was once considered a serious Waymo competitor.
Mercedes-Benz reportedly paused its Level 3 DRIVE PILOT rollout in 2026, pivoting instead to Level 2+ advanced driver assistance systems. Translation: the car will help you drive, but you're still the driver. Luminar Technologies, a major lidar supplier whose sensors were supposed to be essential to the AV revolution, filed Chapter 11 bankruptcy in December 2025.
Meanwhile in China, Baidu's Apollo Go claims over 11 million cumulative rides across approximately 20 cities, with 1.4 million rides in Q1 2025 alone—a 75 percent year-over-year increase. Figures vary by source and should be triangulated carefully, but the directionality is unmistakable: capital and operational momentum are concentrating in a handful of leaders pursuing fundamentally different strategies in dramatically different regulatory environments.
The Long Game
BCG's January 2026 analysis projects a global robotaxi fleet of 700,000 to 3 million vehicles by 2035. That's an evolutionary timeline, not a revolutionary one. Each new city requires $15 million to $30 million in entry costs and four to six years to scale across a metro area, according to the firm. Consumer willingness lags in the U.S. and Europe compared to China, and safety transparency remains critical for building trust—a lesson Cruise learned the hard way after a pedestrian-dragging incident that triggered its regulatory crisis.
Waymo co-CEO Tekedra Mawakana told TechCrunch in October 2025 that the company aims for 1 million weekly trips by the end of 2026, emphasizing that "it's imperative that we scale" while maintaining safety transparency. DeepMind CEO Demis Hassabis has framed world models like Genie 3 as cornerstones of agentic AI, though he acknowledges cost and consistency horizons as near-term constraints. Read: the technology works, but it's expensive and occasionally generates scenarios that violate physics in subtle ways.
The Bet
The question isn't whether generative world models will become standard infrastructure for autonomous vehicle development. They already are, as the industry activity makes clear.
The real question is whether simulated billions of miles can substitute for the decades of real-world validation that public trust and regulatory confidence have traditionally required. It's not an academic debate. Waymo's World Model represents a fundamental bet: that the companies mastering this convergence of generative AI and robotics will compress timelines enough to matter, and that regulators and riders alike will accept simulation as sufficient proof of safety.
Perhaps they're right. Or perhaps—and this is the unsettling possibility the industry doesn't talk about much—we're about to discover which edge cases the simulators missed. The road not traveled, after all, has a way of surprising you.
