Inside Simulation City, Waymo's latest virtual world in which it trains, tests, and validates its “Waymo driver” software in preparation for the open road
A light gray cube with a thin blue top glides down a darkened highway, beset on all sides by dozens of green cubes.
Context & Ripple Effects
Simulation City is the public successor to the simulation effort Waymo has been building for years: back in 2017, its Carcraft VR system was already running 25,000 virtual cars through 8 million miles a day alongside the physical test base at Castle, which journalists were shown around that October. This piece documents how far that virtual layer has come — a dedicated world where the Waymo Driver is trained, tested, and validated before it ever touches the open road.
What makes it worth watching is where the trajectory points: by early 2026 Waymo says it is generating realistic digital worlds with DeepMind's Genie 3 specifically to train on edge-case scenarios, meaning the hand-built environments profiled here are a stepping stone toward generated ones. Simulation has quietly become the core validation infrastructure of Waymo's autonomy stack.
First-order effects
- Waymo can validate new versions of the Waymo Driver against rare road situations inside Simulation City before deploying them to its fleet, shrinking its reliance on physical proving grounds like Castle.
- The same virtual pipeline feeds Waymo's supervised operations on the ground — with roughly 25,000 trained humans assisting about 3,000 robotaxis, software validated in sim determines what those supervisors actually face.
Second-order effects
- Rivals without comparable simulation depth are forced onto a different footing: Tesla's Austin robotaxi service already trails Waymo with far fewer vehicles and safety drivers aboard, so the gap widens along a simulation axis, not just a vehicle-count one.
- Edge-case generation becomes a competitive input — whoever can synthesize more realistic hazard scenarios (as Waymo plans with Genie 3) sets the bar for what 'validated' means across the industry.
Third-order effects
- If simulated validation continues substituting for accumulated real-world miles, regulatory acceptance of virtual evidence could become the decisive gate for scaling autonomous fleets — shifting industry structure toward companies that own both the driving stack and the world-generation tools.
- Validation itself turns into a product of AI research: the boundary between the lab (Google Brain collaborations, DeepMind models) and the road collapses into one continuous pipeline.
The trend: Autonomous-driving validation is migrating from real-world mileage toward synthetically generated worlds, making simulation capability the core moat of the autonomy stack.