Robotaxi Report #6: Cybercab specs, pauking, and fireworks

This newsletter couldn't happen without our long-time partner, Voltera. Here's a quick intro in Voltera's own words:

Voltera builds and operates high-throughput infrastructure networks for AV ride-hail fleets, and our zero-capex next-generation hubs combine charging, servicing, staging, maintenance, data offload, and operations.

You may already be aware of our recent merger with Revel – the combined company will be, by far, the market leader for AV hubs measured by capabilities, current sites, sites in development, and technology.

In case you missed it, here’s a video explaining the merger.

Important: this newsletter does not go through any editorial overview of Voltera's team, nor can they influence any of the reporting — the words in this newsletter are to be blamed on me alone.

Let's me start with the tweet that 2.3 million people nodded along to, from Ethan McKanna (link):

$4.99 plus tax.

(Waymo did refund the fee and added free ride credit later, not sure if due to the post virality there or not… and I hope not)

Funny thing, yes, but it also points at something genuinely new.

When a human Uber driver waits at the curb, there's a social contract (and a rating) keeping things moving.
A driverless car has neither. It just sits there, still active, but not parked, holding public curb space and doing nothing but waiting.

Our friend Reilly Brennan (👋 ) from Trucks VC coined a name for this in-between state: pauking.

It’s the gray zone between parking and pausing.
I recommend reading his full take on this on WITI: (link).

And pauking doesn’t come from just those waiting-for-client moments either. It can be what a robotaxi does when idle, without work.

The economics is what makes it “worth it” for the robotaxi operator side.

UCLA's Adam Millard-Ball modeled this in 2019: cruising slowly costs an AV roughly $0.50 an hour, which is cheaper than parking almost anywhere. His simulation found that as few as 2,000 AVs could drag downtown San Francisco below 2 mph.

And it isn't theoretical anymore. In 2024, Waymo fleet racked up exactly 589 parking tickets and $65,065 in fines in San Francisco, alongside an additional 75 citations in Los Angeles. But there are no pauking tickets.

A company called Meter Feeder in Pittsburgh claims that cities are missing out on $20 million in parking revenue — and potentially half a billion dollars in five years, since robotaxis don’t pay for either. (link)

The solution it offers (I mean of course the quote/claim comes from someone with an a-ha solution, although it seems directionally correct) — Meter Feeder integrated its software into Mapless AI robotaxi technology, and when the vehicles are shifted into park, they send Meter Feeder their location and license plate.

The app then automatically facilitates a 15-minute charge with the Pittsburgh Parking Authority, continuously charging Mapless AI automatically if the car is parked for longer than 15 minutes.

CEO Jim Gibbs told Axios that he sees a lot of inbound interest from the cities. You can see the obvious incentive for it.

As for larger players like Waymo, I would expect direct deals with cities rather a software integration layer like this. Who knows, maybe we’ll even end up with a standardized solution across robotaxi players?

Want to see pauking in the wild? Upshift's Ezra Goldman thinks he found where some of SF's Waymos go to kill time and … pauk?… when there’s a game at the stadium nearby.

It’s a formerly event parking site, and some comments talk of more such activity while the Giants were playing. Whether Waymo has a deal to use this vacant lot at all, we’re not sure. If they did, we’d probably see just parking, not the roaming-around. The neighbors are understandably not very happy of the extra traffic on game days. (video)

So what’s the solution?

What I’d call roaming, driving around in regular speeds or circling in a cul-de-sac, certainly isn’t one.

Pauking doesn’t help.

A depot run, even when there’s more of ‘em, between the rides wouldn’t help.

Temporarily parking in a designated, agreed-to manner, ticketed automatically by the cities? Perhaps.

I already see startups that are grabbing designated parking areas to turn them into staging areas just like this to help along.

Anything else? Let me know your thoughts on comments or feedback poll below.

One thing is for certain: as robotaxi scale goes 🚀 as we speak, and these issues aren’t solved yet, we will see a lot more of these cases pop up.

That validation just inched closer: the two-seat robotaxi's first full spec sheet surfaced in a 2026 EPA certification filing. (link)

The 418-mile figure sits well above the ~300 Tesla has cited early on, though final ratings usually land lower.

It earned its federal Certificate of Conformity on May 26, one of the last boxes to tick before it can legally carry paying passengers.

Since then, Tesla has road-tested production Cybercabs in Austin, posting a 26-second no-wheel, no-pedal cabin clip last week. (video)

And Tesla has been producing what I think should be hundreds of Cybercabs now based on drone views, with a lot of them already staged across parking lots in different cities (seen 80+ in Dallas lot for example). Feels like we’ll soon see a *click* and the fleet awakens.

On the 2025 Tesla Impact Report, launched today (our highlights here), Tesla said Cybercab, with full autonomy and maximized ride-hailing, achieves 2x fewer CHG emissions per mile… compared to Model 3 and Model Y! Yes, they compared it with already some of the most efficient cars on the roads there.

Accessibility of the Cybercab got a showcase, too.

Tesla took a Cybercab to the National Federation of the Blind convention in Austin on July 3, showing features such as Braille lettering on the door releases and emergency controls, room for service animals and canes, and wheelchair-height seating. For the ~2.2 million visually impaired Americans, a steering-wheel-free ride that takes care of you well is quite a strong pitch.

At CVPR 2026, Tesla AI VP Ashok Elluswamy showed a world map of where FSD Supervised is approved, versus pending, and confirmed 1.3 million Teslas now run it worldwide.

For those not tracking it closely, the latest versions of FSD (Supervised) are the same stack as Tesla uses for its driverless (and its safety-driver version) robotaxis in the US. The responsibility is, of course, different — with the FSD (Supervised), the driver is required to do exactly that: supervise and intervene when possible.

In my opinion, most of the talk around Tesla’s FSD system is clouded by the large variation of it. First of all, we’ve seen years of iterations through the poorer versions of the software as it was being built (yet shipped in Teslas), across different Tesla Hardware levels — both of which have been able to create a sense of an unsafe system in a lot of eyes.

It’s still not perfect. But with the newer Teslas (using AI4 (HW4) hardware), on the latest FSD (S) software (FSD v14.3.4), we’re seeing significantly fewer issues and significantly more long streaks of fully self-driving. And by long, I mean ones like David Moss’s, who has driven, as of today, 12,713.4 miles without a single intervention.

You might remember us covering his 3,760-mile Canada coast-to-coast interventionless FSD(S) trip recently. Yep, he is still going.

Tesla is confident enough now to show the “Streak” right on the vehicle screen, creating a sort of gamified sense of how long people can keep it, with confetti flown across the screen on 100 and 1,000 milestones and whatnot.

Anyway, back to the main thing:

Approved countries with FSD(S) are as of today: the US, Canada, Mexico, China (partial), South Korea, Australia, New Zealand, and the European countries of the Netherlands, Lithuania, Estonia, Belgium, and Denmark.

Pending: a near wall-to-wall sweep of Europe plus Saudi Arabia, the UAE, Turkey, Colombia, Chile, Japan, Taiwan, India, Malaysia, Thailand, and, notably, 🇪🇹 Ethiopia.

The last bit surprised me a bit, but perhaps it’s due to Ethiopia becoming the first country ever to ban new gas and diesel car imports? Must have supportive legislation then, too.

Not everyone in Europe is about, however: Czechia has spelled out why it won't recognize the RDW sign-off. “EU-wide” stays a patchwork until the bloc actually votes.

The testing keeps piling up. Tesla logged 275,471 km of FSD Supervised on Spanish roads with zero serious incidents, its 30-car fleet the country's most active tester. (link)

Finland, meanwhile, says it may approve FSD before the EU's October vote, and Tesla's now even hiring an FSD test operator in Dubai.

Tesla now has a Level 4 footprint, on paper. It self-certified its robotaxis to SAE L4 the day Texas's SB 2807 took effect (a self-attestation, not a regulator's finding), then asked Nevada to run up to 5,000 robotaxis around Las Vegas in year one.

The catch: its actual fully driverless fleet is, reportedly, only ~42 to 59 cars, even after doubling the Austin zone to ~245 sq mi. A 5,000-car ask from a ~50-car fleet only pencils out only if the mass-produced Cybercab gets “unlocked”.

Tesla did just take Robotaxi off its home turf, though: Tesla launched in Miami, its first market beyond Texas and California, running Model Ys straight into Waymo territory. (link)

The geofence zone is currently just a slice west of downtown, West Miami, the Fontainebleau area and the airport, stretching from Doral down toward Coral Gables, against Waymo's roughly 60-sq-mi footprint that's been open to the public there since April.

There have been some videos already reviewing the service there, but an interesting one came from David Moss as his robotaxi went through a somewhat flooded road, carefully but surely (video). That’s a touchy subject currently in the robotaxi world, with the mishaps of Waymos in the floods recently causing that recall…

Tesla started pushing FSD V14 Lite to Hardware 3 vehicles on June 29, distilling the newer HW4 driving behavior onto older cameras and compute (park-to-park, speed profiles, the works). Musk notes HW3 has only ~15% of HW4's memory bandwidth and it will not be able to reach the Unsupervised version of the FSD.

Early reviews are glowing; one seven-year-old Model 3 owner said it “feels like it grew a new brain overnight.”

Tesla also pushed back on a fatal-crash headline. After reports tied a deadly Katy, Texas crash to Autopilot, Ashok Elluswamy and Elon Musk said the logs show the driver floored the accelerator to 100% and hit 73 mph on a residential street, overriding the system. We fought back at the NYT et al headlines a bit on this one: (link).
By now, the data review has led to the driver being charged with manslaughter as he overrode FSD by flooring the pedal.

Prosecutors also reportedly found Google searches from his phone, like: “Tesla fsd not aggressive enough 2026,” “FSD is not aggressive enough for city driving,” and “Tesla fsd too timid.”

The charge. A Reuters “investigation” recast Tesla's European push around “misleading” safety stats, its self-published US marketing claims that FSD is up to 10x safer than a human and could have saved 32,000 lives. Critics said the comparisons are stacked, and 10 of 11 independent reviewers called it marketing, not a safety study.

The caution. On June 12, the European Transport Safety Council asked EU ministers to hold off until Tesla answers a few questions in public: why adopt a system US regulators are still probing (red-light running; failure to warn on degraded cameras), and who investigates a failure, since Europe has no NTSB equivalent. It says it's not anti-tech, just anti-rushing. (link)

Our read & rebuttal. The system Europe approved wasn't cleared on Tesla's slides. The Dutch RDW signed off only after 18-plus months of its own testing, and says flatly it “does not rely on marketing claims or external statistics.” (our rebuttal to Reuters)

The RDW has since published a full step-by-step of that process, 3,000+ hours of its own testing, 1,000+ test runs, and monthly (not yearly) reporting from Tesla going forward. (link)

It also helps to be precise: FSD Supervised is legally a Level 2 system, the same class as Ford's BlueCruise and BMW's hands-off motorway mode, which draw none of the same fire.

Waymo launched its first membership program, invite-only for its most frequent riders.

Members get:

Waymo's now at roughly 500,000 paid rides a week across 10 US metros.

One interesting find: the new Ojai (Waymo's Zeekr RT van) is now its own pick in the app, separate from the Jaguar I-PACE, even though both seat four. So far, most reports I’ve seen say the Jag is more luxurious with better screens, but a lot less roomy (ofc) compared to the Ojai.

Splitting the fleet can mean slightly longer ETAs, since the system matches you to a vehicle type instead of pooling every car.

Waymo bought a 5,500-acre proving ground in Wittmann, Arizona, the former Apple Project Titan test site (Apple paid $125M in 2021, then killed Titan in early 2024).

It's got a 115-acre city course, a four-mile oval and a freeway course, rounding out a network alongside Waymo's Castle facility in California and the Ohio TRC.

The moat isn't only tech and test tracks, it's also the unglamorous operations layer. Element Fleet Management, the world's largest pure-play fleet manager (1.5M+ vehicles), just signed a multi-year deal to run Waymo's fleets end to end through its Element Mobility arm.

That covers vehicle lifecycle, charging (which, granted, I think it will somewhat outsource), maintenance, and day-to-day readiness, starting in San Diego, while Waymo keeps the Driver and the rider experience.

It's the enabler layer I keep telling you about: robotaxis get the headlines, but someone still has to keep them charged, clean, and on the road. (link)

And it isn't only Element. In Nashville, Lyft's Flexdrive team has fully taken over day-to-day fleet ops for Waymo, the same backbone role Element and Moove play elsewhere. It looks less like Waymo picking one partner than auditioning several fleet managers at once as it scales.

Nashville is also where that model is going public: Waymo just opened the service to all riders (no invite code), running 24/7 across an initial 60-square-mile zone, with Flexdrive running the fleet.

Waymo ran its first-ever national TV commercial. It’s the first time an AV company has pitched itself to a broad broadcast audience, not just the tech crowd. (video)

The spot leans entirely on safety, in a typical sweet emotional way: the Waymo Driver is, it says, 10x safer than a human in the cities it serves. I did like the early prototype history glimpses there, though.

Remember, Waymo first built their own little two-seater.

The claim has fresh numbers behind it: through March, Waymo logged 220M+ fully autonomous miles with 94% fewer serious-injury crashes than humans on the same roads (and 82% fewer airbag-deployment crashes), on a peer-reviewed methodology backed by IIHS, the University of Michigan and Swiss Re. (link)

There's substance under the slogan, too. With TU Delft (in Nature Communications), Waymo released a benchmark it calls the Reference Driver, modeling how a careful human reads a situation as it unfolds and picks an evasive move, not just last-second reactions.

The point is a peer-reviewed yardstick for the question behind every AV crash: would a good human have avoided it? Waymo is releasing the code under an academic license and lobbying SAE and regulators to adopt it. (link)

… or downright faulty AV systems. Here’s a section showing you some of the documented misbehaving robotaxis.

Waymo's rough Fourth of July in SF. After the fireworks, more than a dozen Waymos ran their batteries flat in gridlock and had to be towed, with pedestrians spotted shoving dead robotaxis out of the road. (link)
Worse, an empty Waymo Ojai caught fire after another Waymo just went and drove over some fireworks that was lit up (video through link above). Luckily no injuries.

Waymo says it's “evaluating ways to strengthen resilience in major traffic disruptions.”

An AV ride to remember in Dallas. A Japanese visitor hailed an Uber, got an autonomous car (an AVride Ioniq 5), and filmed it getting into a low-speed crash mid-ride (safety operator aboard, no injuries). He hailed again, got a second AV. The clip's past 2.2M views. (video)

Just recently, we wrote about a NHTSA probe into AVride operations due to 16 crashes recorded: (link). We know AVride has a full (R&D + service) fleet of ~200 Ioniq 5s now.