Applied Intuition and Foretellix sell coverage-driven verification — enumerate the scenario space, measure what the stack has been tested against, and quantify the gap — while dRISK assembles edge-case risk graphs from crash and near-miss data. This industrializes validation but inherits its own hard problem: proving the scenario library itself is complete.

Aurora publishes a structured Safety Case Framework it must satisfy before each driver-out expansion, and Waymo publishes peer-reviewed comparisons of its crash rates against human benchmarks across tens of millions of rider-only miles. Transparency builds the regulatory and actuarial record, but a safety case is an argument, not a guarantee — and it is graded pass/fail by the next collision.

Waabi built its bet around Waabi World, a closed-loop neural simulator intended to expose the driver to more dangerous scenarios than reality ever safely could, and Wayve trains generative world models to test its end-to-end driver. The approach attacks the long tail directly, but validating a learned driver inside a learned simulator stacks two black boxes — exactly what skeptical regulators struggle to audit.

Zoox designed a steering-wheel-free vehicle around the ride itself and is readying series production in California, while Waymo is diversifying from bespoke Jaguar I-Pace retrofits to cheaper Hyundai Ioniq 5 and Zeekr RT platforms as it scales toward 10,000 vehicles. Purpose-built hardware cuts long-run cost per mile but front-loads enormous capital before the volume that justifies it exists.

Pony.ai is targeting a total Gen-7 vehicle cost below roughly $33,000 by 2027 — base vehicle, battery, and full autonomy hardware included — and WeRide rides the same domestic supply chain into Middle East deployments with Uber. If robotaxi hardware reaches ADAS-adjacent prices, the economics flip; the open question is whether geopolitics lets those cost curves reach Western markets.

Nuro licenses its driver into a Lucid-built, Uber-distributed robotaxi; Mobileye supplies Lyft's Dallas launch with Marubeni owning and financing the fleet; May Mobility deploys onto established ride-hail networks. Splitting vehicle capital, autonomy, and demand across specialists lowers each party's exposure — but also splits the margin three ways and leaves no one owning the customer.

Daimler purpose-built the fifth-generation Freightliner Cascadia with redundant safety-critical systems for Torc, Volvo did the same with the VNL Autonomous for Waabi, and Aurora's next-generation hardware is integrated with platforms including the International LT. Factory-built redundancy is the only path to scale, but it chains every startup's timeline to an OEM's product cycle.

Gatik runs fixed, repeatable middle-mile routes for Walmart, Kroger, and Tyson — known roads, short hauls, no sleeper-cab economics needed — while Kodiak went driver-out first on private roads in the Permian Basin, logging 750+ driverless hours for Atlas Energy. Shrinking the operational domain buys revenue and evidence now, at the cost of deferring the open-highway problem that carries the big prize.

Aurora launched Dallas–Houston, then Fort Worth–El Paso, then validated a ~1,000-mile Fort Worth–Phoenix lane, and extended driverless operation into rain, fog, and heavy wind — treating the envelope itself as the product roadmap. The method compounds credibly, but every added condition multiplies validation burden, and winter-state freight remains outside anyone's driver-out envelope.

Wayve trains a single end-to-end AI driver intended to work across vehicles, brands, and unseen cities — the pitch that pulled Mercedes-Benz, Nissan, and Stellantis into its $1.2B Series D — and Waabi applies the same philosophy to trucking. The approach scales with data rather than surveying, but interpretability and validation of a monolithic neural driver remain the open flank.

Mobileye harvests road data from millions of ADAS-equipped production cars into its REM map layer and pairs it with a formal safety model, letting new geographies come pre-mapped by the existing fleet. It converts an installed base into a mapping advantage, though the leap from crowd-mapped ADAS to driverless-grade assurance in each new city is still nontrivial.

Waymo added Dallas, Houston, San Antonio, and Orlando in quick succession and is laying groundwork in 20+ more cities including Tokyo and London, while Zoox tests in six-plus U.S. metros ahead of service. This is generalization by process rather than by model — repeatable, but each launch still burns months of local validation and fleet capital.

Zoox petitioned NHTSA to operate up to 2,500 steering-wheel-free robotaxis commercially, and NHTSA expanded its Automated Vehicle Exemption Program to domestically built vehicles in 2025. Exemptions unblock individual programs but cap volumes far below commercial scale — a bridge, not a framework.

NHTSA proposed rules in March 2026 to strip human-driver equipment requirements — including the brake pedal — from vehicles designed for automated driving, part of a broader AV framework to modernize federal safety standards. Durable, but rulemaking runs on multi-year clocks and can reverse with administrations, as the FAA's stalled Part 108 shows.

Texas, Arizona, and Nevada host nearly all U.S. driver-out operations — Aurora's freight lanes, Zoox's Las Vegas service, Waymo's fastest expansions — because state law grants operating authority the federal government cannot yet. It gets services live and generates the safety record, but concentrates the industry in Sun Belt conditions and leaves national scale hostage to fifty legislatures.

Skydio's dock-based drone-as-first-responder deployments — now across 1,000+ public-safety agencies, including multi-million-dollar programs in Los Angeles and Orlando — accumulate BVLOS waivers site by site, using autonomous docks to remove the on-scene pilot. It works today and builds the operational record regulators want, but each site is a bespoke approval that Part 108 was supposed to make unnecessary.

Zipline has flown more than two million commercial deliveries and is compounding U.S. volume double-digit percent week-over-week, while Wing runs multi-city residential delivery for major retailers. Millions of incident-free autonomous flights are the strongest safety dataset in the sector — the wager is that demonstrated reliability at volume forces the regulatory unlock.

Auterion supplies a common operating system and swarm-capable autonomy layer adopted across drone makers and defense customers, and Percepto pairs its industrial drone-in-a-box with some of the earliest nationwide-scale BVLOS approvals for site monitoring. Standardization gives regulators one auditable stack instead of hundreds — but only captures value if the ecosystem consolidates around it.

Saronic designs its unmanned vessels from the keel up for autonomy and high-rate manufacturing — its 180-foot Marauder went from design to on-water trials in under a year, and its Louisiana shipyard expansion targets up to 20 per year. Vertical integration attacks cost and reliability together, but the model is unproven past prototypes and priced, at $9.25B, as if the production ramp is certain.

Ocean Infinity operates its Armada fleet with lean crews and shore-based remote operations, and Sea Machines sells autonomy and remote-command kits that bolt onto existing workboats. Proven hulls sidestep the mechanical-endurance gamble and satisfy today's rules by keeping humans in the loop — at the cost of never fully removing the crew economics autonomy promises.

The Navy's MUSV Marketplace is running at-sea demonstrations with seven vendors through October 2026, effectively writing the reliability and autonomy test regime the commercial world lacks, and Blue Water Autonomy is building full-size autonomous ships against those requirements. Defense demand funds the endurance engineering — but it also means the certification standard that emerges may be a military one, with commercial rules still years behind.