DSTL - Audio in. Alpha out.

dstl

TERMINAL

LIBRARY

Inside Waymo: The Architecture of Autonomous Driving

Join

Inside Waymo: The Architecture of Autonomous Driving

Mar 26, 2026

Stripe

1:02:15

7K Views

THESIS

Waymo has crossed the threshold from scientific research to global deployment, with core driving technology now sufficient for full autonomy and the limiting factor shifting to specialization and validation.

ASSET CLASS

SECULAR

CONVICTION

HIGH

TIME HORIZON

10 to 15 years

PREMISE

Full autonomous driving requires a fundamentally different technical approach than incremental driver-assist systems

The hardest parts of building a fully autonomous, rider-only system are qualitatively different from driver-assist systems. While deceptively easy to get started with end-to-end approaches using off-the-shelf VLMs, achieving the superhuman safety bar requires an integrated ecosystem: a foundation model specialized into three off-board teachers (the driver, the simulator, and the critic), then distilled into smaller on-vehicle models. Pure end-to-end pixel-to-trajectory systems work in nominal cases but are orders of magnitude away from required safety levels. The system must augment learned representations with structured intermediate representations to enable efficient simulation, real-time safety validation, and reward function specification.

MECHANISM

Convergence of AI breakthroughs, sensor cost curves, and operational learning enables simultaneous scaling

Three forces are now aligned for rapid expansion. First, foundation model advances allow the Waymo Driver to generalize across weather conditions, cities, vehicle platforms, and sensor configurations with zero-shot or few-shot learning. Second, sensor technologies are following predictable cost decline curves—imaging radars and LiDARs are becoming dramatically cheaper while maintaining capability. Third, fifteen years of iterative operational learning have compressed deployment timelines from eight years for the first four public cities to four cities launched in a single day. The sixth-generation hardware achieves comparable cost to fancy ADAS systems while supporting full autonomy, eliminating the hardware cost barrier to mass deployment.

OUTCOME

Autonomous vehicles will dominate urban transportation within fifteen years

Waymo currently operates 3,000 vehicles completing 500,000 rides and 4 million fully autonomous miles weekly across 11 US cities. International expansion to London and Tokyo begins this year. The convergence will occur not through driver-assist systems graduating to full autonomy, but through Level 4/5 systems expanding geographically while hardware simplifies and costs decline. Personal ownership of Waymo-equipped vehicles will extend coverage to low-density areas where ride-hailing economics fail. Second-order effects include elimination of parking infrastructure, reduction of phantom traffic jams through consistent driving behavior, and fundamental restructuring of urban land use.

NECESSARY CONDITION

Regulatory frameworks must remain permissive to innovation (avoiding the 'European' model) and open source development must remain unencumbered by downstream liability.

I've had some moments where a car does something, and you look at a log, and I've been surprised. It does things that I didn't think it was capable of doing.

31:45

RISK

Steel Man Counter-Thesis

Waymo's apparent technological lead may paradoxically become a strategic liability. The company has optimized for a world where autonomous driving requires a comprehensive sensor suite, custom vehicles, extensive validation infrastructure, and depot-based operations—a high-fixed-cost model that only makes economic sense at massive scale. But the rapid democratization of AI capabilities Dolgov describes means competitors can now achieve 'good enough' autonomy with radically simpler architectures. If regulators, facing public pressure for autonomous vehicle access, adopt safety standards benchmarked to human drivers (who cause 40,000 US deaths annually) rather than Waymo's superhuman targets, then camera-only competitors could achieve regulatory approval with 10x lower capital requirements. Moreover, Waymo's ride-hailing model faces structural challenges: vehicles must be positioned for demand, creating utilization inefficiencies; depot infrastructure scales linearly with geography; and the service competes with human drivers whose labor costs may remain competitive in many markets. The personally-owned autonomous vehicle, which Dolgov acknowledges as a product request, would obviate many of these challenges—but would also eliminate the ride-hailing revenue model entirely. Waymo may have built the perfect solution to the wrong problem: optimizing for the 'number of nines' when the market-winning strategy is optimizing for cost per mile.

RISK 01

Regulatory and Political Risk as Scaling Bottleneck

THESIS

The thesis assumes Waymo's core technology is mature and the challenge is now 'accelerated global scaling.' However, this framing treats regulatory approval as a procedural formality rather than a fundamental constraint. Different jurisdictions have vastly different liability frameworks, insurance requirements, and political constituencies (taxi unions, labor groups, privacy advocates) that can impose indefinite delays or operational restrictions. The London and Tokyo expansions require not just technical adaptation but navigation of entirely different legal regimes. A single high-profile fatality in a new market could trigger regulatory reversals across multiple jurisdictions simultaneously, as happened with Uber's autonomous testing program after the 2018 Arizona fatality.

DEFENSE

Dolgov frames international expansion as primarily a technical specialization and validation problem, stating 'the core technology generalizes really well, but there's still work that you have to do.' The regulatory and political dimensions of scaling are entirely absent from his analysis. He does not address how Waymo would respond to regulatory setbacks, liability frameworks, or the possibility that different markets may impose fundamentally incompatible requirements.

RISK 02

Unit Economics and Capital Intensity at Scale

THESIS

The interview reveals that Waymo operates approximately 3,000 vehicles doing 500,000 rides per week, implying roughly 167 rides per vehicle per week or 24 rides per day per vehicle. While Dolgov mentions Gen 6 hardware costs are 'a fraction' of previous generations and comparable to 'a fancy ADAS system,' the full operational cost structure remains opaque. The infrastructure described—depots, manual cleaning, human charging attendants, fleet management systems, remote monitoring capabilities—represents substantial fixed costs that must be amortized. At current scale, Waymo is almost certainly unprofitable, and the path to profitability requires either dramatic cost reduction or massive scale increases. Competitors with simpler sensor stacks (camera-only approaches) may achieve worse safety metrics but substantially better unit economics, potentially capturing market share before Waymo can scale.

DEFENSE

Dolgov provides no financial metrics, margin targets, or timeline to profitability. When discussing operational infrastructure, he emphasizes 'increasing efficiency and automation' but acknowledges cleaning and charging remain manual processes. The economic viability of the ride-hailing model at scale is simply not addressed, nor is there any comparison to competitor cost structures or discussion of pricing power.

RISK 03

Competitive Moat Durability Against Fast Followers

THESIS

Dolgov explicitly states that building a functional autonomous driving demo is 'very easy to get started' with off-the-shelf VLMs—you can 'just take a VLM, fine-tune it, and it will drive pretty darn well in the nominal case.' This admission suggests the barrier to entry for achieving reasonable autonomous driving performance has collapsed. The competitive moat therefore rests entirely on the 'number of nines' required for commercial deployment. However, if safety requirements are set by regulators rather than consumers, and if regulators adopt a 'good enough' standard (better than human drivers but not dramatically so), then Waymo's extensive safety infrastructure may represent over-engineering rather than competitive advantage. A competitor achieving 'only' 3-4 nines of safety might gain regulatory approval and scale faster, using real-world data to improve while Waymo waits for perfect validation.

DEFENSE

Dolgov defends the moat by arguing there is a 'qualitative jump' between driver-assist systems and full autonomy, stating 'the hardest parts of building a fully autonomous, rider-only system are very different from what you do for a driver-assist system.' He also emphasizes that Waymo's integrated ecosystem of driver, simulator, and critic models—all derived from a shared foundation model—creates compounding advantages that cannot be easily replicated. The defense is plausible but assumes regulators will maintain high safety bars and competitors cannot shortcut the validation process.

ASYMMETRIC SKEW

Downside scenarios involve regulatory roadblocks that impose multi-year delays on international expansion, competitor breakthroughs that achieve 'good enough' autonomy at fraction of cost, or a single catastrophic incident that triggers industry-wide regulatory retrenchment. Upside scenarios involve continued scaling at current trajectory with improving unit economics as Gen 6 hardware deploys and automation increases. The asymmetry appears moderately negative: the technology risks have diminished substantially, but execution risks around regulation, economics, and competition remain substantial and are largely unaddressed in the interview. The 'path dependency' risk—having built expensive infrastructure that becomes competitively irrelevant—represents a fat-tail downside that could manifest over a 3-5 year horizon.

ALPHA

NOISE

The Consensus

The market believes autonomous driving remains a fundamentally unsolved technical problem, with full Level 4/5 autonomy still years away from widespread deployment. Consensus holds that the gap between Level 2/3 driver-assist systems and full autonomy is incremental—a matter of degree rather than kind—and that companies working from driver-assist upward have a viable path to full autonomy. The market also believes that custom-designed autonomous vehicles are necessary for commercial viability and that significant city-specific engineering work creates scaling barriers.

Market logic holds that autonomous driving requires solving an essentially unbounded long-tail of edge cases, that each new city requires substantial mapping and localization work, that sensor fusion complexity creates irreducible system brittleness, and that the path from demo capability to commercial safety standards spans many years. The consensus also believes hardware costs remain prohibitively high for mass deployment and that the jump from retrofitted vehicles to purpose-built autonomous cars is essential for unit economics.

SIGNAL

The Variant

Dolgov asserts the core technology problem has been solved. He explicitly states: 'I don't see today any limitations or any gaps in the core technology. The driving is good enough now.' The remaining work is specialization, validation, and scaling—engineering execution rather than fundamental research. More critically, he rejects the incremental view: driver-assist and full autonomy are 'fundamentally two different problems' requiring a 'qualitative jump,' not continuous improvement. Companies cannot work their way up from ADAS to L4. The technology generalizes across cities and even vehicle platforms far better than expected, enabling rapid geographic expansion without proportional engineering investment.

Dolgov's causal model centers on a foundation model architecture that creates compounding returns: one foundation model specializes into three off-board teachers (driver, simulator, critic), which then distill into deployable models. This architecture means improvements at the foundation layer propagate automatically through the entire system. He claims VLMs now provide zero-shot or few-shot generalization to new cities by inheriting general world knowledge, dramatically reducing city-specific work. The sixth-generation sensor stack costs 'a fraction' of previous generations—comparable to 'a fancy ADAS system'—suggesting hardware cost barriers are collapsing faster than appreciated. The retrofit vehicle strategy was deliberate risk management, not a limitation; the purpose-built Ojai platform arrives when the software risk is retired.

SOURCE OF THE EDGE

Dolgov's claimed edge rests on operational reality and architectural insight unavailable to outside observers. The operational component is credible: Waymo runs 500,000+ rides weekly across 11 cities, generating proprietary data on edge cases, rider behavior, and fleet operations that no competitor can replicate. His architectural insight—that foundation models can be specialized into teacher models and distilled for deployment—is more difficult to assess from outside, but his description of emergent capabilities (the pedestrian-behind-bus detection) suggests genuine empirical discovery rather than post-hoc narrative construction. The claim that driver-assist cannot scale to full autonomy is structural industry knowledge from 20 years of observation, not marketing positioning—and notably runs counter to narratives that would benefit Google commercially against Tesla. However, his assertion that 'core technology' is solved warrants skepticism: insiders always underestimate remaining work, and the gap between 99.9% and 99.999% reliability may be larger than current operations reveal. The edge is real but possibly overstated on the timeline for full problem resolution.

CONVICTION DETECTED

• 'I don't see today any limitations or any gaps in the core technology' • 'The driving is good enough now' • 'I think you have to tackle… it is a qualitative jump' • 'Eventually, it will, absolutely. There's no doubt in my mind' • 'There's no silver bullets' • 'The core technology generalizes really well'

HEDGE DETECTED

• 'I'm not going to give it a date today' • 'We still have work to do' • 'There is a lot of work to do in specialization and in validation' • 'I don't want to say you can't make the jump, but it is a qualitative jump' • 'I don't want to speculate too much on the psychology thing' • 'It remains to be seen, I think' The ratio reveals genuine operational confidence rather than performative certainty. Dolgov hedges on timelines and commercial specifics but speaks with conviction on technical architecture and strategic direction. This pattern is consistent with an engineer who knows the system deeply but respects implementation complexity. He never hedges on whether the technology works—only on when deployment milestones will occur. This suggests high credibility on technical claims but appropriate humility on business execution, and indicates his core thesis (technology solved, scaling underway) deserves significant weight.