Anatomy of Zoox — A Steering-Wheel-Free Robotaxi Company Reaches Paid Service

A name borrowed from algae living in coral

Zoox was founded in the United States in 2014. It had two founders whose backgrounds barely overlapped.

Jesse Levinson had spent nearly a decade at Stanford researching autonomous driving. He wrote the navigation algorithms for "Junior," a modified Volkswagen Passat that placed second in the 2007 DARPA Urban Challenge, and his doctoral advisor was Sebastian Thrun, who later led Google's self-driving program. Tim Kentley-Klay was a designer from Melbourne, Australia, whose background was in advertising direction and running an animation studio — no automotive or robotics experience at all. He incorporated Zoox in Melbourne in 2013, then moved to Silicon Valley to find engineers and venture capital.

The company's name comes from zooxanthellae, the algae that live symbiotically with coral. According to the company, zooxanthellae generate energy through photosynthesis and maintain a mutually beneficial relationship with coral — and Zoox wanted to build a mobility service that had that same kind of relationship with cities.

One decision made in that first year shaped the next twelve. Instead of bolting sensors onto an existing car, Zoox decided to build a car from scratch. The company repeated that same idea in a blog post on July 30, 2026: "Since founding the company in 2014, we have strongly believed that the best way to solve autonomous driving is with a purpose-built robotaxi." The first prototype was built in 2015 inside a converted firehouse on a college campus, and its first autonomous drive took place in a private parking lot.

The face-to-face four-seat layout used today traces back to a single photo from 2014 — Levinson, Kentley-Klay, industrial design lead Nahuel Battaglia, and an early employee sitting in four lawn chairs arranged to face each other.

A board that fired its founder, then Amazon

The money came in fast. Zoox raised $200 million at a $1 billion valuation in June 2016, then $500 million at a $3.2 billion valuation in July 2018, with the Series B led by Atlassian co-CEO Mike Cannon-Brookes. According to Form D filings with the SEC, the amounts Zoox actually received were $103.05 million and $50 million in 2016, $462.39 million in 2018, and $200 million in October 2019.

Then, just a month after the Series B closed, on August 22, 2018, the board fired founding CEO Kentley-Klay. He tweeted that the board had "terminated me without warning, cause, or the chance to respond." The company has never given any official explanation. Reporting at the time cited investor concerns about his personality and leadership style rather than any fraud or misconduct — but all of it relied on anonymous sources. Board member and former Autodesk CEO Carl Bass stepped in as executive chairman, and Levinson was promoted from CTO to president. Kentley-Klay's name disappears from the list of related parties in the October 2019 Form D.

A permanent CEO was announced on January 14, 2019: Aicha Evans, who had spent 12 years at Intel and risen to chief strategy officer, overseeing the company's communications and devices group and leading a team of 7,000 people spread across multiple continents. She took the role on February 26. At the time, Zoox said it had more than 700 employees and had raised over $750 million cumulatively.

On March 20, 2019, Tesla sued Zoox and four former Tesla employees for trade secret theft — though the case involved warehouse, logistics, and inbound/outbound procedure documents, not self-driving technology. It settled in April 2020, with Zoox stating: "We acknowledge that some new employees who came from Tesla were in possession of Tesla logistics-related documents at the time they joined, and we regret the actions of those employees." Settlement terms were not disclosed.

Amazon announced its acquisition of Zoox on June 26, 2020. Amazon has never disclosed the price. The Financial Times reported it exceeded $1.2 billion, and outlets that reviewed the deal documents put the structure at $1.3 billion in cash plus more than $100 million in Amazon stock for employee retention. Amazon's 2020 annual report described the deal only in aggregate — "$1.2 billion net of cash acquired, of which $1.1 billion was allocated to in-process research and development intangible assets" — without naming Zoox anywhere. In fact, the word "Zoox" appears nowhere in the body text of any Amazon 10-K or 10-Q; the company's name only shows up in an earnings press release attachment (EX-99.1). It was a down round from the $3.2 billion valuation, and by early 2020 Zoox was burning through nearly $30 million a month.

Amazon's acquisition announcement didn't mention logistics or delivery once — only ride-hailing. Levinson put it more bluntly in an August 2026 interview: "They liked the purpose-built robotaxi. If we hadn't had that vehicle, they wouldn't have bought us." After the deal, Evans and Levinson continued running the company, and Zoox remained an independent Amazon subsidiary. As of August 2026, Evans is still CEO and Levinson co-founder and CTO. Company materials put headcount above 3,000.

What it actually means to have no steering wheel

According to company specs, the Zoox robotaxi is 3,630mm long, 1,936mm tall, has an 8.4-meter turning circle, a designed top speed of 75 mph (about 121 km/h) in either direction, and up to 16 hours of driving range per charge. Under federal regulatory filings, its gross vehicle weight rating (GVWR) is 3,000kg, and it's classified as a passenger car seating up to four. At its 2020 unveiling, the battery was listed at 133 kWh, but starting in 2025 the company dropped kWh from its materials entirely, describing capacity only in hours. That figure went unupdated even after Panasonic Energy announced in November 2025 that it would begin supplying 2170 cylindrical cells starting in early 2026.

75 mph is a design spec, not an operating condition. At the time the federal exemption took effect, the speed cap in Zoox's declared Operational Design Domain (ODD) was 45 mph.

Removing the steering wheel isn't just dropping one component — it meant redesigning the entire safety architecture.

Airbags are the clearest example. With no dashboard, there's nowhere to deploy a frontal airbag outward from. According to the body of Zoox's federal petition, the system has four types. A "horseshoe" airbag deploys from the ceiling in a U-shape on each direction of travel, preventing occupant ejection in side impacts and rollovers while also serving as the reaction surface the frontal airbags push against. A frontal airbag drops from the ceiling above each seat; a side airbag built into each seat protects the pelvis; and an airbag inside the seat cushion lifts the front of the cushion during a high-speed frontal collision to keep occupants from sliding forward. That last one only deploys in forward-facing seats. The vehicle manual and occupant manual attached to the same petition list a fifth type — a rear airbag that deploys behind the headrest on each direction of travel. Since the company's own documents disagree on the count, the total number can't be stated with certainty.

Crumple zones were another problem — the car is too short to leave room to fold. Zoox's solution was to pack the drive module and motor into a tight space and let that structure itself absorb and dissipate crash energy before it reaches the passenger compartment. The company ran thousands of virtual crash tests before building a physical prototype in 2018, and conducted its first physical crash test in 2019. It says it applied a higher standard to the face-to-face seats than federal rules require for rear seats. That said, there is no third-party crash rating — from IIHS, NCAP, or anyone else — for this structure. Every piece of validation to date has come from the company itself.

Redundancy borrows from aviation. There are two batteries, a primary and secondary steering system, and simultaneous use of three cellular modems — AT&T, Verizon, and T-Mobile. Zoox describes this as "fail-operational" rather than "fail-safe" — designed to keep operating safely through a failure rather than simply stopping. Per the petition, if all three modems lose connection for an extended period, the vehicle automatically pulls over to the shoulder.

The company won't say how many sensors it has

There are five sensor types: high-resolution cameras, lidar, radar, long-wave infrared cameras, and microphones. Four "sensor pods" mounted at the roof's corners form the core, with radar embedded around the body and microphones inside the cabin. The four pods actually use only two distinct part numbers and are designed to be updated separately from the rest of the vehicle. Because there's no hood or trunk, the corner pods have an unobstructed line of sight down to the ground, which is the basis for its close-range vision.

The exact sensor count is not disclosed. Even in the federal petition, the sensor layout diagram is redacted as "Confidential Appendix 1." Outside reporting gives conflicting numbers — 64 sensors according to EE Times in 2020, 48 according to Forbes in 2025, and roughly 40 according to TechCrunch in 2026 — likely reflecting differences in vehicle generation and counting methodology (whether microphones or interior cameras are included). Detection range figures also don't match across company materials: a July 2025 vehicle brochure says "150m or more," while the website's safety FAQ says "up to 200m." No evidence was found that the company designs its own lidar. Senior VP of hardware Nelson Federeiro has said, "We have several manufacturing partners working on our behalf, and more than 130 suppliers worldwide."

The most detailed software disclosure came in a "Safety Case Framework" document published August 12, 2026. The autonomous system runs on five interlocking functions — localization, perception, prediction, planning, and control — with an independent collision-checking layer sitting on top. That layer uses its own perception and its own algorithms to judge whether the trajectory the main system has planned is safe, and to intervene if it isn't. The petition calls this layer the Collision Avoidance System (CAS), responsible for both longitudinal and lateral intervention and for deciding when to end an intervention. The safety metric is CIF — miles driven between collision, injury, or fatality events. The human-driver baseline was built from NHTSA's CRSS and FARS data plus Federal Highway Administration SHRP2 and mileage estimates.

The prediction model's generations are publicly named. Zoox moved from UAP (Unified Active Prediction), a graph neural network approach, to QTP (Query-centric Trajectory Prediction) after concluding UAP couldn't handle unexpected behaviors like jaywalking or illegal U-turns. QTP pulls behavior directly from data rather than assuming human reactions or destinations.

Simulation runs on more than 10 million miles of real-world driving data. Synthetic simulation searches for conditions that could produce a collision; log-based simulation replays actual driving logs through updated software, using machine learning to prioritize rare, safety-critical events for replay. Results are weighted by real-world exposure to build risk estimates. The company also transfers scenes captured in San Francisco into an Austin-like environment to test cross-city transfer.

However, cumulative simulation mileage has never been disclosed — a contrast with Waymo, which cites 20 billion miles. The federal petition redacts everything from the number of scenario variants run per software release to cumulative training miles and driverless miles; a line item for "cumulative simulation miles" doesn't appear at all.

There's one more wrinkle. At an AWS re:Invent session (AMZ304) in December 2025, Zoox said it was training a multimodal foundation model combining camera, lidar, and radar on Amazon SageMaker HyperPod. According to AWS presentation materials, this involved more than 64 GPUs running at 95% utilization, with vehicle data collection reaching up to 4TB per hour. This work has never been mentioned through any of Zoox's own channels. There's a gap between the stack the company describes in its public documents and the stack its own engineers describe at conferences.

Four years of self-certification, then an exemption

The longest-running problem in Zoox's history wasn't technical — it was paperwork. The U.S. Federal Motor Vehicle Safety Standards (FMVSS) are written assuming a car with a steering wheel and pedals. A vehicle without one has to choose between two paths: self-certify that it meets the standards, or apply for an exemption.

On July 11, 2022, Zoox announced it had chosen the first path, calling itself "the first company to self-certify a purpose-built vehicle that does not rely on traditional driving controls like a steering wheel — without seeking a regulatory change or exemption."

Here's how the next four years played out.

Date	Event
2022-07-11	Zoox announces FMVSS self-certification
2022-09-06	NHTSA issues a special order demanding evidence for the certification
2023-03-03	Audit query AQ23001 opens, covering 64 Zoox self-certified vehicles
2024-12	Following a field inspection, NHTSA issues a report citing "multiple apparent instances of noncompliance with applicable FMVSS"
2025-06-05	Zoox applies for a display exemption
2025-08-04	Exemption granted; AQ23001 closed. Condition: delete/redact all prior claims of FMVSS compliance
2026-07-30	Part 555 commercial exemption granted — paid rides allowed
2026-08-10	First paid rides begin in Las Vegas

NHTSA's closeout report explained why it opened the inquiry: Zoox was unusual in the industry for choosing self-certification instead of applying for an available exemption, and public information suggested it might not meet certain standards. It added: "As a result of discussions with the agency during the investigation, Zoox elected to apply for an FMVSS exemption."

In other words, the 2022 self-certification wasn't validated — it ended with Zoox switching to an exemption and scrubbing its compliance claims. And the exemption didn't even cover every problem vehicle. The closeout report states: "This exemption covers all Zoox vehicles that were the subject of this investigation, except for certain 2023 model-year vehicles that Zoox has excluded from public road operation." Zoox added a note to its 2022 announcement post reading "Amended August 6, 2025," and in an August 15, 2025 post explained that its vehicles would be labeled "demonstration vehicles" for the time being, meaning they "do not meet all FMVSS requirements related to manual driving controls and driver equipment." A July 2026 post described those same four years as "thoughtful conversations with regulators about outdated FMVSS requirements."

The commercial exemption announced July 30, 2026 and effective July 31 covers partial relief from eight standards: windshield defrosting (103), windshield wipers and washers (104), turn signal and headlamp switching controls (108), rearview mirrors (111), foot-operated braking controls (135), sun visors (201), the requirement that windshields use AS1-grade glass (205), and airbag warning labels on sun visors (208). None of these are full exemptions — each covers only specific provisions. Occupant crash protection and braking/lighting performance are not exempted, and glass safety performance testing requirements remain in place. The exemption runs for two years, until July 31, 2028, and caps commercial deployment at 2,500 vehicles per 12 months.

The structure of the exemption's conditions matters. NHTSA attached a standing oversight mechanism called "Operational Authorization" to this exemption — a device that sets conditions by operating environment and can be modified in future approvals, with the first operational authorization issued simultaneously with the exemption decision. It effectively pulls the oversight that the now-withdrawn AV STEP program was meant to provide into the Part 555 framework instead. Separately from the annual production cap of 2,500, the number of vehicles that can operate simultaneously is set by this operational authorization.

NHTSA was explicit that its safety-equivalence finding applies only to the specific items Zoox sought exemption for, and that it reached no conclusion on compliance with any other standards. Yet the same notice, in the very next paragraph, states that "in determining whether granting the exemption serves the public interest, the agency considered the safety and maturity of Zoox's automated driving system." In short: driving performance was excluded from the standards review but factored into the public-interest determination.

July 2028 isn't necessarily a cliff edge. Federal regulation 555.8(e) states that if a renewal application is filed at least 60 days before expiration, the exemption stays in effect until the agency approves or denies it. The real risk isn't expiration itself — it's the possibility of denial, and the fact that operational authorization can be scaled back or modified at any time. In the meantime, one standing program disappeared: NHTSA withdrew its proposed AV STEP rule on June 26, 2026. Industry groups said it was too burdensome for an autonomous vehicle program; safety advocates said it lacked mandatory oversight.

Five recalls — all software, all voluntary

Because Zoox doesn't sell vehicles but owns and operates them directly, there's no owner-notification process for recalls. All five recalls to date have involved software defects, and all were self-reported.

Recall	Filed	Vehicles	Defect
25E-019	2025-03-13	258	Excessive hard braking in response to crosswalk cyclists or fast-approaching vulnerable road users from behind, creating rear-end collision risk
25E-029	2025-05-01	270	Above 40 mph, mispredicts whether vehicles at side entrances will stop
25E-037	2025-05-22	270	Fails to continuously detect a road user who has fallen immediately next to a stopped or slow-moving vehicle
25E-090	2025-12-22	332	Unnecessary lane incursions near intersections (62 instances observed, no collisions)
26E-044	2026-07-16	105	Fails to detect and respond to dense smoke, particularly at active emergency scenes

The first three each stem from a real incident. 25E-019 traces back to a modified test vehicle with a safety operator aboard, which braked hard and was then rear-ended twice by a motorcycle — an event that led to NHTSA's Preliminary Evaluation PE24-015 in May 2024. 25E-029 involved a driverless Zoox vehicle colliding with a passenger car in Las Vegas on April 8, 2025. The company identified the conditions two days after the crash, confirmed the cause and a fix three days later, and rolled it out fleet-wide within eight days. Driverless operations were fully suspended in the meantime.

25E-037 stems from a May 8, 2025 incident in San Francisco, where a scooter rider struck a stopped Zoox vehicle mid-turn at low speed. The rider fell and remained right next to the car when it began moving again. Injuries were minor, but the incident exposed a blind spot: the system's failure to detect a fallen person pressed close against the vehicle's body. The company's Ultra-Short-Range Detection (USRD) system, described in its federal petition, is designed to close exactly this gap — it blocks the vehicle from moving if any object, regardless of classification, is detected immediately around the sides or top of the vehicle. That said, the company has never specifically cited USRD as the corrective action for this recall.

The 2026 incident was a different kind of problem. On June 20, a driverless Zoox vehicle entered thick smoke at an active fire scene that hadn't been cordoned off with cones. The vehicle braked hard, changed direction, and stopped; a remote operator then backed it up so firefighters could place cones. No one was injured. Zoox notified regulators on June 25 and pushed a fleet-wide software update on July 15. The recall report doesn't name the city where it happened, and the company has declined to disclose it.

Notably, Zoox's own declared ODD includes "light to moderate fog, haze, or smoke." Dense smoke falls just outside that boundary — meaning this was a failure at the edge of the operating domain, not outside it entirely. The recall was also filed during the window in which NHTSA Administrator Jonathan Morrison had, in early July, told autonomous vehicle developers he saw "a clear pattern of driverless vehicles interfering with law enforcement and emergency response" and demanded solutions by the end of July. The agency hasn't disclosed which companies received that letter.

On paper, the record is clean. Across NHTSA's entire investigation database, only two formal investigations have ever targeted Zoox — AQ23-001 and PE24-015 — and both are closed. As of the database's August 21, 2026 update, there are no open investigations. But much of that cleanliness comes from the company issuing recalls before regulators act. And while Zoox published blog posts explaining each of its March and May 2025 recalls, it did not do so for the December 2025 or July 2026 recalls.

The factory can build 10,000; the permit allows 2,500

Zoox's Hayward factory opened on June 18, 2025. It spans 220,000 square feet (about 20,000 m²) — roughly three and a half football fields, by the company's own description — and has annual production capacity for more than 10,000 vehicles at full utilization.

Notably, automation is minimal. Robots are used only for tasks like applying window adhesive and moving the body along the line; humans handle everything else. There's no welding, cutting, or painting stage at all, which also means lower power consumption than a typical automaker's plant. Suppliers pre-assemble major components before shipping them in, and Zoox's quadrant-symmetric design allows for modular assembly.

End-of-line inspection is split across eight stations. At the sensor calibration bay, halogen lighting calibrates the visible-light cameras while also heating dots on a board to calibrate the infrared cameras on the same surface. At the wheel and headlight alignment stage, since there's no steering wheel, the zero-degree steering angle is set electronically. On the dynamometer, the vehicle drives itself up to 75 mph without a driver, with a laser-based system detecting lateral drift to keep it centered on the rollers. From there it moves through a leak test, a light-tunnel visual inspection, a static functional test (FST), and a buzz/squeak/rattle test (BSR). The final dynamic test (FDT) is the vehicle's first autonomous drive — and because the car is bidirectional, it runs for hours in both clockwise and counterclockwise directions.

Mass production began in July 2026. On June 24, the company unveiled a "production-intent vehicle," saying it could ramp up to 100 units a week "pending regulatory approval." But what actually changed in the June reveal was mostly cosmetic — aloe green seats, stone gray flooring, seat cushions, headrests, the touchscreen, charging pads, cup holders, the position of the color-shifting reflectors used to indicate direction of travel, and door speakers and microphones. No changes to sensors, compute, battery, or drivetrain were announced.

This is where the numbers stop lining up.

Item	Value
Hayward production capacity	10,000+ vehicles/year
Federal exemption cap	2,500/year (5,000 total over 2 years)
Simultaneous operating fleet	Set separately by operational authorization
Nevada permit	100 vehicles
Actual cumulative production	~100–105 vehicles

Running the factory flat out would exhaust an entire federal year's allowance in about 13 weeks. The bottleneck isn't manufacturing capacity — it's regulation and operations, and the company itself acknowledged as much by attaching the "pending regulatory approval" caveat.

Where things stand right now

As of August 24, 2026, Las Vegas is the only city where Zoox charges fares.

City	Status	Timeline
Las Vegas	Paid, publicly available	Free public launch 2025-09-10 → paid 2026-08-10
San Francisco	Free, waitlist invite-only	Launched 2025-11-18, service area expanded 2026-03-24
Austin / Miami	Early access for employees and friends only	2026-03-24
Atlanta, Los Angeles, Seattle, Washington DC, Phoenix, Dallas	Test driving / mapping	Phoenix and Dallas added 2026-03-09

Cumulative autonomous driving exceeds 3 million miles. Ridership figures diverge by source. Zoox itself reported on July 30, 2026 that it had served "500,000-plus riders, with a waitlist over 500,000," while Axios reported in early August that combined ridership across Las Vegas, San Francisco, Austin, and Miami was "approaching a million." The company's official figure and secondhand reporting differ by roughly a factor of two.

Fares are calculated as a base fare plus distance and time, locked in at booking so a returning vehicle doesn't rack up extra charges. Destination fees apply separately at Harry Reid Airport, Sphere, and T-Mobile Arena. Pricing is set to match the "comfort" tier of standard ride-hailing services. However, Zoox does not disclose its base fare or per-mile/per-minute rates. Any specific dollar figures reported in the media are estimates.

Zoox still can't charge fares in California. The company holds a driverless testing permit from the California DMV and a piloting permit (with and without a driver) from the California Public Utilities Commission (CPUC) — but lacks the deployment permit from either agency required to charge fares. The CPUC permit explicitly states: "This permit does not authorize receiving any form of compensation for transportation services provided using a testing autonomous vehicle." Waymo, by contrast, holds a deployment permit and on August 14, 2026 was approved to expand in the Bay Area and Los Angeles, plus enter Sacramento and San Diego as new markets. Zoox can't generate revenue in the state where it's headquartered.

Zoox partnered with Uber on March 11, 2026. Under the deal, Zoox will become bookable through the Uber app in Las Vegas starting summer 2026 and in Los Angeles by mid-2027, while keeping its own app running in parallel. As of August 24, though, there's no confirmation that Zoox rides have actually gone live on Uber's app in Las Vegas.

Won the regulatory fight, losing the scale fight

One thing is clear about Zoox: it's the only company in the U.S. charging fares with a vehicle that has no steering wheel or pedals. Waymo's Jaguars, Zeekrs, and Ioniq 5s are all vehicles a human could still drive, and Tesla's Cybercab hasn't been deployed yet.

Every other metric runs the other way.

Metric	Zoox	Waymo
Vehicles on the road	~100–105	~3,500 (est. July 2026)
Cumulative autonomous miles	~3 million	220+ million (as of end of March 2026)
Weekly paid rides	Undisclosed (2nd week of paid service)	~500,000
Cities with paid service	1	11 metro areas
Third-party safety validation	None	IIHS study (2026-07-23)

Zoox's entire cumulative ridership is roughly what Waymo does in one to two weeks. The gap with Chinese players is just as stark. Pony.ai reported 1,975 robotaxis as of late June 2026, targeting over 3,500 by year-end; WeRide had more than 1,800 robotaxis on the road as of late July. Baidu Apollo Go reported, in its Q2 2026 disclosure, operations across 28 cities worldwide and cumulative autonomous driving of 350 million kilometers.

And just ten days after Zoox began charging fares, on August 20, Nevada's transportation authority approved commercial robotaxi permits in Clark County for 5,000 Tesla vehicles, 1,000 Waymo vehicles, and 1,000 Uber-affiliated vehicles. Zoox's own permit covers 100 vehicles. Some of the 1,000 Uber-affiliated units will go toward partnerships with Motional and Zoox, so not all of it is competing supply — but it's still clear that the window in which Zoox was effectively alone in the only market where it could charge fares lasted less than two weeks.

What about vehicle cost? Zoox has never disclosed it. For comparison: Waymo's Jaguar I-Pace is reportedly priced in the low $200,000s per unit. Waymo's imported Zeekr has a U.S. import declared value around $38,000, but after China's 127.5% EV tariff, that climbs to roughly $86,500, and adding roughly $25,000 for the sixth-generation autonomous hardware pushes the total past $100,000. Tesla claims a manufacturing cost of $30,000 for the Cybercab, and Baidu's Apollo RT6 is reportedly priced at $28,000.

There's a counterargument, though, that vehicle cost may not be the deciding factor. According to an analysis by Brad Templeton, vehicle depreciation accounts for roughly 25% of a robotaxi's per-mile operating cost. Even if a vehicle costs 25% more, total cost only rises by 7–10% — a negligible difference against current ride-hailing prices of around $2.50 a mile. If that math holds, Zoox's real constraints aren't vehicle price but the 2,500-per-year federal exemption cap, the simultaneous-fleet limit set by operational authorization, Nevada's 100-vehicle cap, the lack of a California deployment permit, and a production rate that started at roughly one vehicle a day.

Editor's view

There are two ways to read Zoox.

One is about the price of time. The 2014 decision — build the car from scratch — paid off exactly as designed, twelve years later. Zoox is now the only company charging fares with a vehicle that has no steering wheel, and while Waymo pays more than $48,000 in tariffs to import a single Zeekr, Zoox is stamping out its own vehicles from a factory in California. But that same choice also cost twelve years of delay. Because Waymo retrofitted existing cars, it could scale in the early 2020s, and the 220 million miles it accumulated in that time is now feeding third-party research like IIHS's. Zoox has only 3 million miles, which is why it's had to substitute simulation and analysis for accumulated mileage in proving out safety. When VP of safety Qi Hommes said the company had "no choice but to lean on analytical simulation and data-driven approaches rather than logging huge amounts of miles," that reads less like a chosen methodology and more like a consequence of circumstance.

The other way to read it is that regulation and scale are two different games. Zoox won the regulatory game. It spent three years going down the self-certification path, got flagged for apparent noncompliance, changed course — and ultimately became the first purpose-built robotaxi to win a commercial exemption. But the prize for that victory was a cap of 2,500 vehicles a year. Ten days later, the same county approved 7,000 vehicles for other companies. Clearing regulatory approval first and capturing the market first turned out not to be the same thing.

Three things are worth watching over the next year. First, whether Zoox secures a driverless deployment permit from California's CPUC — the current state of affairs, where it can't charge fares in its own home state, isn't sustainable indefinitely. Second, whether Hayward actually hits 100 vehicles a week — the gap between one vehicle a day at launch and that target is the company's real bottleneck, compounded by whatever simultaneous-fleet cap operational authorization ultimately sets. Third, the renewal review in July 2028. Under the rules, filing on time keeps the exemption in force during the review period — but the outcome of that review is entirely out of Zoox's hands. What the company has said it hopes for, in its own petition, is that FMVSS itself will be amended by then, making its vehicle simply legal on its own terms — which is an even bigger variable outside Zoox's control.

No one knows how much Amazon has actually put into Zoox. Amazon has never disclosed it, and the company's name doesn't even appear in the body of its 10-K filings. Andy Jassy's only mention of Zoox in his 2025 shareholder letter was a single phrase: a company "just beginning commercial service." Levinson has said scaling the technology going forward will require "tens of billions of dollars." That this is a forecast of future spending rather than a disclosure of past spending may be the most accurate summary of where Zoox stands today.