AVs vs. Subways

Plus, Waymo wades into sensor discussion fray, drone delivery war heats up, and Pony.ai is on a roll

This is the Ride AI newsletter: The most comprehensive weekly digest of news and intelligence at the intersection of technology and transportation.

Top Story

PC: Waymo

There’s no AI shortcut to full self-driving, says Waymo, laying out ten lessons from its first 200 million fully autonomous miles. The company emphasised that even the most advanced AI models can’t replace redundant sensors, human safety review, and years of real-world validation. Waymo has gradually consolidated dozens of specialized models into a handful of larger ones over its 15+ years on the road, but says moving to a single, camera-only model asks AI to do more than it can safely handle today.

Srikanth Thirumalai, Waymo’s vice president of onboard software, said that the industry’s growing bet on end-to-end systems, ones that read raw sensor data and output steering commands directly without hand-built maps or redundant hardware carries risks Waymo isn’t willing to take. “Even the best AI models with trillions of parameters still hallucinate,” he said. “There is no click reboot or reload or refresh [in] physical AI. You have to deal with the consequences of it.” This seems to reflect the broader industry consensus, as even Motional, which took its robotaxis off the road to rebuild itself around an end-to-end system, still integrates a critic into its AV stack.

Waymo also weighed in on the industry’s biggest sensor debate, saying its own testing found cars see significantly better when cameras work alongside lidar and radar than with cameras alone. “The AI can only make sense of what it sees, and if you just can’t see it, the AI can’t do much,” Thirumalai said, a direct jab at Tesla’s bet that cameras alone can eventually solve full autonomy. He said the same holds for high-definition maps, which some rivals consider unnecessary overhead.

I certainly think it’s interesting that this piece lands just days before Tesla is set to reveal its production Cybercab robotaxi in Austin on September 3. When the Cybercab starts scaling on public roads, it will be the clearest test to date of whether a camera-only, end-to-end approach can catch up to Waymo’s head start built on multi-modal sensing and HDmaps. However, there might be a bit more to this debate… more on this below.

Domestic News

Waymo and Zoox safety drivers have suffered more than two dozen injuries from hard braking since 2024, according to OSHA filings, with whiplash and other injuries from sudden stops sidelining some workers for months. The filings only cover companies large enough to trigger mandatory reporting, so the real number across the industry’s other safety drivers may be higher.

PC: Gatik

Gatik has raised a $200 million Series D, money it plans to use scaling its driverless middle-mile trucks past the Fortune 50 routes it already runs for retailers and grocers including Walmart and PepsiCo. Gatik is one of the companies currently running autonomous trucks in the US. The company says it’s completed more than 85,000 fully driverless orders with a 99% on-time delivery rate, running dozens of trucks with nobody behind the wheel across Texas, Arizona, Arkansas, and Canada, and has a rather spiffy live dashboard of running freight on its website homepage, I might add. I should also note that the types of trucks Gatik is running are not the same as the 18-wheelers used in long-haul trucking.

Amazon Prime Air plans to expand sixfold to nearly 500 US cities by the end of 2026, now that drone delivery is suddenly a hot topic after Alphabet’s Wing and Uber Eats’ drone partner Zipline announced their expansions. Uber is targeting 1 million daily Eats deliveries by drone by 2029, while Wing plans to reach more than 270 Walmart stores and 40 million people by 2027. Let’s just hope they figure out the drop-off situation…

Public Transit vs. AVs

I saw a thread last week that argued that BART (Bay Area Rapid Transit, or more plainly the subway) is a worse deal than driving your own Tesla, citing operating costs of $1.15 to $1.18 a passenger mile for BART against 70 to 85 cents for a Tesla Model Y driven solo. Given that I’m generally a proponent for public transit and micromobility, less so of more cars, but at the same time a proponent of AVs, I dug in to see if this actually holds up.

Both figures check out. Federal transit data puts BART’s operating expenses at $883 million against 751 million passenger miles in fiscal year 2024, or $1.18 a mile, matching the top of the range cited in the thread. AAA’s 2025 driving cost study puts the average EV, insurance, depreciation, maintenance, and electricity included, at 71.2 cents a mile, inside the range too.

The BART numbers have moved since fiscal 2024, though. The agency’s fiscal 2025 financial report, released in December, shows ridership grew 6.3% year over year to 52.7 million trips, or 51% of 2019 levels. Monthly ridership has kept climbing since. May 2026 posted the highest average weekday ridership since the pandemic, 202,650 riders a day, up 12% from a year earlier.

That was a lot of numbers just now, all to say that at face value, the post is technically true.

Now, others have done a better job of talking about the throughput efficiencies of public transit vs personal vehicles, but that’s not the comparison I am interested in here. The comparison I’m interested in is the energy usage between BART and autonomous vehicles, and there are currently two high profile variations of AVs: multi-modal sensors, and camera-only. The two approaches vary significantly in hardware and implementation, so how do they measure up against BART?

Let’s start with the vehicles themselves. I know the Ojai just launched, but Waymo’s fleet is still mostly the Jaguar I-PACE, EPA-rated at 76 MPGe combined, or about 0.44 kWh a mile. Tesla’s robotaxis run on the Model Y, and the Long Range AWD version is EPA-rated at 117 MPGe, or about 0.29 kWh a mile. The Robotaxi hardware variation shouldn’t add very much energy usage to this configuration. Before any self-driving hardware enters the picture, Tesla’s vehicle is already roughly 35% more efficient than Waymo’s, so that’s our baseline.

Moving onto the autonomy related hardware. A University of Michigan and Ford study modeled a large, Waymo-style rooftop sensor array specifically and found it added about 5% to energy use, mostly aerodynamic drag, using a 200-watt computer as its baseline. Waymo’s actual compute load runs well past that baseline today. A 2023 MIT study on AV computing emissions used 840 watts as its working assumption for a self-driving computer, and a more recent review of the literature puts the Waymo Driver specifically at around 1 kilowatt. Tesla’s cameras, by contrast, sit flush in the body, and the company’s own AI4 computer draws about 80 to 100 watts, roughly a tenth of Waymo’s. If we use the higher, more current compute figure, in addition to the drag estimate, that’s roughly 10% to 20% overhead for Waymo’s setup and 3% to 8% for Tesla’s, putting propulsion plus hardware at 0.49 to 0.53 kWh a mile for Waymo and 0.30 to 0.31 kWh a mile for Tesla.

Then there’s how many of those miles actually carry someone. A study of 13.8 million California robotaxi trips found only 54% of Waymo’s vehicle miles in the state have a passenger onboard, and when one does, it’s carrying 1.4 people on average. We don’t yet have similar data for Tesla’s robotaxi service, and most rides there still run with a human safety monitor, so I applied Waymo’s measured deadheading rate to Tesla’s numbers too as it’s unlikely to significantly change. Dividing vehicle energy by passenger miles instead of vehicle miles gives 0.65 to 0.70 kWh per passenger mile for Waymo and 0.39 to 0.41 kWh for Tesla.

lower = better

BART’s own sustainability reporting puts systemwide electricity use at about 368,000 megawatt-hours a year, or 0.49 kWh per passenger mile across the 751 million passenger miles NTD recorded for fiscal year 2024. That puts BART at 24% to 30% ahead of my Waymo estimate, but 16% to 20% behind my Tesla estimate.

So is BART a “worse deal” than AVs? As it turns out, yes and no. A leaner, camera-only car has a lead, but only if a robotaxi service runs dense and driverless enough. BART doesn’t have that problem to solve. Running a train costs roughly the same in electricity whether it’s half full or packed, so if ridership keeps climbing, its number keeps improving. This can actually be seen with the Tokyo metro system. The difficult part is making BART more attractive for more people to ride.

PC: @GrantSlatton via X

International News

PC: Waymo

Waymo has picked Munich for its first robotaxi launch in the European Union, targeting a public, driverless service by the end of 2027. It plans to follow its usual playbook, mapping the city’s streets manually, then testing with safety drivers, then offering driverless rides to employees and invited guests before opening up to paying customers.

The company holds no permits yet in Germany, but the legal path there is friendlier than in most of Europe, since the country passed a law enabling Level 4 driving on public roads more than five years ago.

On the other hand, Transport for London said this month that it’s in no rush to publish the guidance operators need before launching, potentially ruling out paid, driverless rides there before 2027. Wayve and Uber hold licenses for 15 supervised vehicles so far, and Baidu is testing separately, but zero vehicles have been certified for driverless service and zero of the required government permits have been issued. Zagreb beat London to become Europe’s first robotaxi market back in April , through Uber, Pony.ai, and local operator Verne. Yes, yes, Brexit, I know, but still true.

China has drafted a law that would make automakers responsible for traffic violations committed while a car drives itself, shifting liability away from the driver for Level 3 and Level 4 systems while leaving standard driver-assistance features under existing rules. The draft revision to the country’s road traffic law runs 170 articles across nine chapters and adds a dedicated section covering autonomous-vehicle violations, accidents, and insurance. A related safety standard for Level 3 and Level 4 systems takes effect in July of 2027, by which point Level 2 driver-assistance is predicted to have shipped in more than 70% of new passenger vehicles sold in China.

PC: FutureLink

Pony.ai is on a roll, bringing its robotaxis to Seoul and partnering with local operator FutureLink to deploy 200 BAIC-built vehicles as a first step toward fully driverless service. The companies are targeting 2028 for a driverless commercial launch pending certification from Korean regulators, building on more than 80,000 kilometers of accident-free testing Pony.ai has already logged in the city’s Gangnam district.

Alright, that’s it from me… until next week. If you enjoy this newsletter, share it with your friend, colleague, or boss. Thank you for reading, Sophia out!