Gates Grok
Skim: Belief flips, not ops trackers. Unsupervised that sticks · cost edge with utilization · hard-city parity · disclosure that survives · densify that pays · China stays control.
First-principles questions that must flip from “maybe” to “yes.” Thresholds of belief — not invented fleet counts or crash rates.
Gate A Unsupervised that sticks
Unpublished Judgment
Can a fleet run useful unsupervised hours/day in weather and cities that matter — without monitors creeping back after incidents? Gate is not a single “we removed the safety driver” press hit; it’s utilization that stays online.
What would change my mind
Rising named-city unsupervised counts with stable ops after weather and minor incidents. Counter-signal: fleets that quietly reintroduce monitors or shrink geofences after every headline.
Gate B Cost edge converts to unit economics
Judgment Tesla’s ~$30–40k vehicle-cost thesis only matters if utilization × fare − opex − liability beats Waymo’s denser-but-costlier stack. Gate: evidence that cheap vehicles earn, not that they park in lots.
Fact Sep 14: Austin robotaxi still surface-street ODD (no highways / avoid at-grade rail); rider report ~10→~70 min detour on a short hop — same constraint on Cybercab and Model Y robotaxis.
Fact Sep 23: Waymo began public driverless freeway rides in Austin via Uber (gradual); same city now has freeway-capable Waymo vs surface-street-only Tesla ODD.
What would change my mind
Paid stranger hail at meaningful utilization on low-cost vehicles — or, bear side, lot inventory growing while unsupervised miles stay tiny. Cost without utilization falsifies the edge. Highway-capable ODD (or competitive trip times vs Waymo freeway rides) would materially reopen Gate B quality.
Gate C Hard-city / hard-regulator parity
Judgment Camera-only (or any stack) clears Austin-class / Bay-class / eventually NYC-class cities with crash/mile and insurer comfort that doesn’t require permanent special ops. Easy cities do not open this gate. Purpose-built Cybercab (no wheel/pedals/mirrors) now faces an NHTSA Special Order on FMVSS self-cert (AQ26002 escalated; Sep 30 deadline passed — no public outcome yet) — regulator parity is a live constraint, not a slide.
What would change my mind
Independent safety prints competitive with human baselines in hard cities, plus regulator language that doesn’t treat the stack as probation forever. Persistent +4%-worse-than-human style local surprises keep the gate shut.
Gate D Disclosure that survives sunlight
Judgment Soft gate markets ignore until they don’t: can Tesla (and peers) publish unsupervised rides, fleet, and crash/mile definitions that survive auditor-and-regulator sunlight? NHTSA Special Order under AQ26002 (Cybercab FMVSS self-cert; Sep 30 deadline passed — no public outcome yet) is the live sunlight test. Ambiguity is a feature until an incident — or a Special Order — makes it a bug.
What would change my mind
Unpublished Consistent unsupervised metrics in filings or serious ops disclosures. Continued darkness after a multiple shock = narrative still doing the work of disclosure.
Gate E Densify that pays (Waymo path)
Judgment Waymo’s gate: rides/week and rides/car/day in core metros compound toward a cash-printing density — not endless thin logos. Dual-app matching (Nashville Waymo×Lyft, same fleet) is a channel densify signal; it does not itself prove rides/car/day. Trajectory toward high weekly rides with improving unit economics closes the gate.
What would change my mind
Clear densification in served markets with utilization that looks like a business — including whether third-party app demand (Lyft) raises rides/car/day without thinning the core. Counter-signal: metro or app-count rising while rides/car/day stall — expansion theater.
Gate F Keep China as control, not peer
Judgment Intellectual gate for investors: refuse to import China weekly rides as US race ranking. Use China only to stress-test “what scale looks like under different permission.” Closing this gate is discipline, not a KPI.
What would change my mind
Only if US permission/liability regimes converge in a way that makes China ops transferable — they mostly don’t. Default: separate panels forever.