Autonomous vehicles: the future we can't buy

01The most mature embodied AI, and the leading indicator

Autonomous driving is a decade ahead of humanoids, which makes it the single most valuable thing to read in Physical AI: it shows us, in advance, how an embodied-AI market resolves once reliability becomes life-critical, capital becomes national-scale, and regulation becomes binding. The headline is that it resolves toward consolidation around a few deep balance sheets, the opposite of what we found in defense, and for instructive reasons.

Waymo is the proof that autonomy at scale is real: ~3,500 vehicles, driverless in roughly ten cities and expanding into twenty more, with a genuine, audited safety record. Tesla, by contrast, is running a small, still partly-supervised pilot in two cities, perhaps 200 vehicles, most rides still monitored, and has conceded its earlier hardware cannot reach unsupervised autonomy. The gap between the two is the gap between a deployed system and a promised one, and it is the first thing the chart below shows.

02The framework, stress-tested

Why integration won here, and won't be undone

Defense showed recombination arriving early because the buyer wanted cheap, disposable bodies and a software moat. Autonomous vehicles show the reverse, and the master framework predicts exactly why. Three conditions all peak here at once: the reliability bar is not 99% but closer to 99.9999%, because a failure is a fatality on a public road in an unbounded domain; the capital required is national-fleet-scale, with per-city mapping and operations measured in quarters and funding rounds measured in tens of billions; and regulation is granular, per-city and per-state, a moat that compounds for whoever is already inside it. Those are precisely the conditions under which the master thesis says vertical integration and deep balance sheets win, and they have. Cruise is dead, Argo is dead, Motional retrenched; the survivors at the crown are Waymo (Alphabet), Zoox (Amazon), Tesla and Baidu. This is World A in its most extreme form, and nothing on the horizon unwinds it.

Where the growth-accessible value actually is

Here is the investor's problem, stated plainly: the robotaxi crown is not ownable. Waymo is Alphabet, Zoox is Amazon, Tesla is Tesla, Baidu is Baidu, a growth fund cannot buy the winners except through their mega-cap parents. So the discipline is to find the parts of the AV opportunity that are both real and reachable, and there are three.

Trucking: the reachable edge with the best economics

Autonomous trucking reaches driver-out economics earlier than robotaxis for structural reasons: highway driving is a far more bounded operational domain than dense urban streets, freight is high-value and time-sensitive, and the driver shortage is chronic. Kodiak is already fully driver-out on commercial freight; Aurora launched driverless in Texas and plans to remove its in-cab observer and scale to hundreds of trucks. These are public, loss-making and early, but they are the part of AV where the unit economics turn first and where a growth investor can actually take a position.

03The economics: a long road to $0.25

The bull case for robotaxis rests on a terminal cost of roughly $0.25 per mile. It is worth being honest about how far away that is. Today a Waymo ride costs the company about $1.98 per revenue mile against a fare near $2.00, gross margin close to zero, and still prices above Uber in most cities. The near-term make-or-break is roughly $0.99 per mile, about half of today's cost, which gets to a ~16-month vehicle payback. The $0.25 figure is a long-run destination that assumes cheaper hardware, higher utilisation and stripped-out mapping and operations overhead, not the present.

China complicates the picture and confirms the master thesis's bifurcation. Chinese operators run larger fleets at lower cost and claim city-level unit-economic breakeven already, but they are effectively locked out of the US on national-security grounds and expand instead through the Middle East, Europe and Asia. Two parallel markets, split by trust, exactly as the security overlay predicts.

04Why this chapter matters even though we mostly can't buy it

AV earns its place in the thesis less as a place to deploy capital and more as the sector's most reliable forward indicator, because it runs on the same stack as everything else in Physical AI. The convergence toward end-to-end neural and world-model driving, Wayve, Tesla FSD, Nuro, Waabi's shared "AI brain" for trucks and robots, is the identical foundation-model, simulation and sensor-fusion technology underpinning humanoids. Wayve explicitly frames itself as embodied AI bridging vehicles and robots. So what happens in AV first, how the end-to-end bet resolves, how fast reliability climbs the last nines, how the cost curve bends, how regulators treat probabilistic agents, is the clearest available preview of how the humanoid and industrial stories play out a few years later. Reading AV is how we calibrate the rest of the compass. And the cross-cutting picks-and-shovels it shares, NVIDIA compute and simulation, LiDAR and sensor suppliers, Uber as demand aggregator, are the same layer we favour in Infrastructure, which is the one place AV exposure and Orien's edge genuinely overlap.

05Orien's verdict

AV signposts we track (as sector indicators)

• Does robotaxi UE cross breakeven? Waymo's path to ~$0.99/mile and its stated 1M-rides/week profitability inflection is the clearest real-world read on whether embodied-AI economics actually work at scale.
• Camera-only vs sensor-rich. If Tesla's vision-only end-to-end bet reaches unsupervised parity, it validates the cheap-generalisation thesis for the whole sector, humanoids included. If it stalls, modular-and-sensor-rich stays the safe architecture.
• Trucking driver-out scale. Watch Aurora removing its observer and Kodiak scaling to hundreds of trucks, the first place AV economics turn genuinely positive, and our most direct exposure.

06What breaks this call

Tesla is right and it re-rates everything. Our "mega-cap crown, reachable edges" framing assumes the robotaxi leaders stay a closed club. If Tesla's camera-only approach cracks cheap, scalable, unsupervised autonomy, cost collapses and the whole map redraws, including for humanoids that share the vision-first architecture. We hold this as the key upside risk to a cautious stance.

Trucking driver-out slips. The reachable edge depends on driver-out economics actually arriving; observers-in-cab and regulatory caps have slipped before. If trucking timelines extend, the one place we take direct AV exposure loses its near-term thesis.

The public names are volatile and loss-making. Aurora, Kodiak, Pony and WeRide are real businesses but small, cash-burning and sentiment-driven, with China risk on the latter two. Position sizing and entry discipline matter more here than conviction.

Robotaxi economics never reach $0.25. If the terminal cost stalls above human ride-hail, the grand TAM numbers compress and the sector stays a premium niche, a reminder that the master thesis's reliability-and-cost caution applies to the most mature vertical too.

07Sources