The nines the demo never shows
A perfect 2014 Waymo drive still took years to become a real service. Tesla now lists seven metros and a Nevada permit. The safety report, the open federal analysis, and the community tracker still do not agree the remaining nines are marched.
Andrej Karpathy took a Waymo ride around Palo Alto in 2014, before he ran Tesla’s Autopilot. It was a perfect drive. He thought the problem was close. It was not. A decade later Alphabet’s robotaxi runs about 500,000 paid rides a week across ten cities — and Karpathy, in October 2025, still called the deployments “pretty minimal.” The cars, in his telling, live in the future and had to be pulled back into an uneconomical present.
It's a march of nines. And every single nine is a constant amount of work. So when you get a demo and something works 90% of the time, that's just the first nine.
Andrej Karpathy, Dwarkesh Patel, October 2025
At Tesla from 2017 to 2022 — five years leading Autopilot — Karpathy thinks the team cleared “maybe three nines or two nines.” There are still more nines to go. “This is not even near done,” he said post-tenure. Demos hit the first nine quickly because they are handpicked best-case runs. Product deployment needs every subsequent nine, and each one is a long fixed-cost slog. Karpathy: “I'm very unimpressed by demos.”
Two bets on how to march
Waymo is the existence proof for commercial driverless service: redundant LiDAR, radar, cameras, and HD maps; about 15 million rides in 2025; safety drivers removed after roughly 30,000 city miles to critical disengagement — about 37 times the ~809 miles community trackers now observe on Tesla FSD v14.2. Tesla is the counter-bet: cameras only, no LiDAR, no radar, a fleet-learning flywheel on ordinary streets. Karpathy still rates Tesla’s approach as the more scalable strategy, with the caveat that he is not fully independent.
Peter Diamandis, on a late-August Moonshots tape, put a cost gap on the same two bets. Cybercab, he said, is about $30,000. A Waymo fifth-generation Jaguar is about $300,000 — $200,000 for the car and $100,000 for the hardware — and sixth-generation hardware cost had fallen “from 115,000 to 20,000.” Alexander Wissner-Gross, on the same hour, counted 13 cameras, four LiDAR, six radars. Those are speaker-attributed hardware claims from one episode. They are not a fleet count and they are not a safety result.
The July Q2 Update widened the map without closing the scale gap. Tesla now lists seven US metros. Austin, Dallas, Houston, Miami, Orlando, and Tampa are “Ramping Unsupervised.” Phoenix and Las Vegas are “Preparations Underway.” The Bay Area is still FSD (Supervised) under a California TCP chauffeur permit. Cybercab started production at Gigafactory Texas. The April snapshot that still sits on the wiki — roughly 4–8 unsupervised Model Ys in Austin against Waymo’s ~100 vehicles and ~90 square miles there — has not been replaced by a third-party fleet count. Paid-ride scale is still Waymo’s.
A company number, and the ones that do not agree
Tesla’s public safety numbers do not settle the argument. Phil Koopman’s analysis of Tesla’s safety report finds a flawed baseline (old cars without modern ADAS), a five-second crash-attribution window against NHTSA’s 30 seconds, under-reporting of severe crashes, and selective categorization. “We can conclude nothing useful about the practical safety of FSD from this report,” Koopman wrote. IIHS classifies FSD as convenience, not safety improvement. Community-tracker “city miles to critical disengagement” fell from ~4,109 at v14.1’s October 2025 peak to ~809 after v14.2. Musk publicly claims FSD is “10× safer than humans.” The wiki treats that as unsupported at best.
On the 22 July earnings call, Ashok Elluswamy offered a different number: 380,000 unsupervised robotaxi miles across six cities, “zero notable incidents,” an “impeccable” safety record, and cameras-only “validation” against LiDAR, radar, and HD maps. “Notable” is undefined. The figure is first-party and unaudited. It does not repair Koopman’s critique of the published safety report, and the August pass fetched no updated community-tracker table for v15. Leave both numbers on the table.
Regulators and “driverless”
NHTSA’s Engineering Analysis EA26002, opened in March 2026, covers 3,203,754 vehicles for FSD failures in reduced visibility — nine crashes, one fatality, one injury. The official concern is degradation detection: the system did not see that the cameras were impaired, or did not alert, until immediately before impact. Tesla began a detection update the day after a fatal crash was reported; the company later said that update may have touched three of the nine incidents. On 2 July the agency issued Information Request EA26002-01. Tesla’s response was due 12 August. As of 24 August there is no fetched close-resume, recall, or mitigation order. A parallel probe, PE25012, compiled traffic-violation evidence. The recall question stays open.
California’s CPUC still says Tesla “is not operating an autonomous vehicle service.” Nevada drew a different line. On 20 August the Nevada Transportation Authority approved Tesla Robotaxi, LLC as an Autonomous Vehicle Network Company for paid rides in “fully autonomous” vehicles in Clark County — up to 5,000 cars in the first twelve months, a 120-day clock to commence operations, inspections and insurance still required, airport service a separate authorization. Approval is not a ride. Tesla’s 22 July Update still had Las Vegas in preparations. China’s market regulator, the same week, treated the consumer stack as supervised Level 2: a 2,740,642-vehicle recall (SAMR S2026M0039I) because existing driver-attention monitoring was not enough to prompt when a gaze left the road. The remedy is an over-the-air cabin camera. That is a watching-driver finding, not a vision-versus-fusion result.
The Q2 print adds the P&L the robotaxi map does not. Active FSD subscriptions were 1.48 million at quarter-end, up 56 percent year on year — the only fetched software-monetization count that already sits in automotive ancillary and deferred revenue. CFO Vaibhav Taneja said full-year capex would run above $25 billion, for two to three years, assigned to robotaxi fleet, AI compute, and Optimus. No fetched robotaxi revenue, rides, or paid-mile dollar came with the package. Elluswamy’s miles are an operations claim. They are not a line item.
Karpathy also claimed, without full corroboration, that Waymo cars with no one in the front seat may still have “very elaborate teleoperation centers” and “more human in the loop than you might expect.” Whether remote-supervised robotaxi qualifies as driverless is an open question on the wiki. So is whether vision-only perception is sufficient for Level 4 autonomy. Seven metros, a Nevada permit, and a company mile-count sit next to an open federal camera-degradation analysis. They do not close either question.