Waymo's L.A. crash tally tripled in 2025 as robotaxi data splits safety debate

Waymo reported 219 crashes in Los Angeles County in 2025, more than triple its 69 in 2024. Another 186 have been filed so far this year.

Those numbers come from a Los Angeles Times review of autonomous vehicle crash reports filed with the National Highway Traffic Safety Administration, published on 5 October. The same review found that more than 80% of this year's reported Waymo crashes in the county caused no apparent injuries. Many happened while the vehicles were stopped.

The Los Angeles figures land in the middle of a wider argument about how robotaxi safety gets measured, and about which data anyone should trust. Companies point to lower injury rates than human drivers. Researchers say the operating conditions are still narrow enough that the comparison is not like for like. California remains the main proving ground: Amazon's Zoox is already testing driverless vehicles in L.A., and Tesla's Cybercab and others are expected to follow. Matthew Raifman, who studies autonomous vehicles at UC Berkeley's Safe Transportation Research and Education Center, has compiled a dashboard of robotaxi accidents using government data. He told the Times that the state is absorbing much of the trial and error. "I wish it wasn't a robotaxi race, even though it is," he said. "I hope it's a slow stroll towards safe mobility."

The fatal cases are rare. In the ones the Times describes, they are not always Waymo's fault.

At an intersection in San Francisco in January, a collision involving at least five vehicles and an empty Waymo left one person and a domestic animal dead. The Times reports it was the first time an autonomous vehicle with no safety driver was involved in a fatal incident. In August in Dallas, a Waymo was one of the vehicles involved in an accident that killed a pedestrian. In Arizona last year, a motorcyclist was fatally injured after hitting a Waymo and being flung into the road. In each of those fatal accidents, the Times says, Waymo did not appear to be at fault.

Tesla, meanwhile, is expanding its robotaxi hours in Austin to 11 p.m. from 10 p.m., the Times reported on 5 October. Asked why it could not go later, Elon Musk said the company was working on protecting pets. "The main thing we're trying to solve is making sure that we don't run over pets when they're hard to see at night," he said in a post on X, according to the Times. "Literally trying to avoid grey kittens on grey tarmac in the dark."

What the data can and cannot show

The crash reports are a regulatory byproduct, not a designed dataset. They tell you how often a robotaxi was involved in something reportable, not whether the robotaxi was the cause. That is why the same numbers get read two ways. Some see evidence that driverless fleets are accumulating incidents as they scale. Others see evidence that most of those incidents are low-speed contact in heavy traffic.

The data problem has a technical side that gets less attention. A blog post published on 5 October by an engineer working with autonomous driving data argues that size is the easy part. Using Motional's 2019 nuScenes dataset as an illustration, the post calculates that six cameras at 1600x900 and 12Hz produce about 311MB per second, or roughly 1.1TB per hour, before lidar and radar are counted. Storage and compression are solved problems, the author writes. The hard parts are time alignment across sensors, finding the rare incidents worth learning from, and reproducing which data and label versions trained a given model. At 100km/h a car covers about 28 metres a second, so a 10ms gap between sensors shifts an object's position by roughly 28cm.

That framing matters for anyone reading crash statistics. The events that matter most are the rarest ones, and the pipelines that surface them are still being built.

Storage, power and the physical footprint

The compute behind all this analysis has its own constraints. They showed up on 5 October from several directions.

Toshiba plans to double production capacity for hard disk drives used in AI data centres within fiscal 2027, measured against its 2025 level, Nikkei Asia reported on 5 October. The Philippine plant expansion is described as Toshiba's first major investment in its HDD business in about five years. On the database side, Asana published an engineering post on 5 October describing how it shards petabyte-scale customer data across well over a hundred physical MySQL instances. The customer ID is the partition key, and a single customer's data lives on one instance. Datadog, in a post dated 5 October, described building Distributed DataFusion, an open source Rust framework that spreads one query across many machines. The company decided ClickHouse was not extensible enough for its needs, and that Spark and Ballista materialised intermediate data.

Smaller projects are chasing the same workloads. Pivot, a Rust analytics database for Delta Lake and Iceberg, is on GitHub. Its own README says it is in early development and not yet production ready. SQLite published documentation on 5 October for sqlite3_rsync, a remote-copy tool that turns a replica into a consistent snapshot of an origin database while both stay live. It costs one database per invocation and at least one endpoint on the local machine.

None of that is sport. All of it is the plumbing that sports analytics, video review and scouting platforms sit on.

Power, water and permission

Data centre siting has become a three-front fight over grid costs, water and local land use, according to Data Center Knowledge. The outlet reported on 5 October that a wave of state action in late September 2026 changed the rules. California's AB 1577, AB 2383 and SB 887 require developers to cover the full incremental cost of grid upgrades, and Massachusetts Executive Order 658 applies to facilities with peak demand above 25MW. The same outlet reported separately on 5 October that operators are turning to biomimicry and community co-design to soften the visual impact of facilities. It cited Microsoft's site in Noord-Holland, where it planted 150 native trees and more than 2,000 square metres of shrubs and grasses in 2022. Microsoft says the practice has since reached more than 20 sites in the US and Europe.

A 20-storey data centre tower proposed for downtown Kansas City, Missouri, would support 24MW of IT load on about 0.23 acres, Data Center Knowledge reported on 5 October. Only two of its 20 floors would hold computing equipment, with roughly 109,000 of its 141,000 gross square feet given to mechanical space. Developer Revitalization Unlimited put total development costs at $183 million, about $1,300 per square foot. "We are not arguing that every data center belongs downtown," said founder and CEO Steve Austin. "We are arguing this one does, and we are paying a construction premium to prove it." A related city ordinance is listed as on hold until 13 October.

When the infrastructure fails

Two environmental stories from 5 October are worth keeping in view for anyone who assumes the cloud is weightless.

Finland's Licensing and Supervisory Authority is examining more than 300 hectares of forest cleared for a Google data centre in Muhos. The regulator told Google's project company in May that an environmental assessment was needed first, Wood Central reported on 5 October. Sari Myllyoja, who runs the inquiry at the authority, said it looked as though the project company had gone ahead regardless. Google communications manager Michiel Sallaets said trees had been felled "in full compliance with the requirements of the Forest Act". Muhos municipality told the regulator on 5 October that it has not issued the project a building permit or any other permit.

In New Jersey, a 5,000-gallon diesel spill at Equinix's NY2 facility in Secaucus reached an unnamed tributary of the Hackensack River, the Times of India reported on 5 October, citing the state Department of Environmental Protection. Crews had recovered about 3,500 gallons. Secaucus Mayor Michael Gonnelli initially attributed the incident to a "computer glitch", while Equinix said two broken valves caused the release. The department said it had observed no impact on local wildlife and that no drinking water intakes were near the affected area.

Denmark's national civil registration system offers a sharper warning about data governance. The CPR administration said on 5 October that unauthorised parties used a Danish company's legitimate search access to obtain names, addresses and CPR numbers covering about 8.8 million registered people. The Copenhagen Post reported the same day that the breach took place during September and that the administration became aware of it on the Friday evening before. Minister for Research, Education and Digitalisation Christina Egelund called it "a deeply serious incident". Datatilsynet said it received notification on Sunday and could not yet assess the circumstances.

For sports technology, the practical reading is less dramatic than the headlines. Injury models, tracking systems and broadcast analytics all depend on federated data, shared infrastructure and permission to operate. The Waymo crash numbers are the most visible test of that arrangement. They will keep being published whether or not anyone agrees on what they mean.