What it is: The autonomy programme the panel uses as its worked example of a cost curve collapsing inside one hardware generation.
Why it is in this archive
Gate: cited in ≥2 episodes — 33 mentions in Moonshots #283, 6 in Moonshots #284. Comfortably earned.
Facts
⚠️ On-air figures. Not independently verified — flagged as such in the timeline too.
| Autonomy hardware | "$115k → $20k" in one generation (−82.6%) |
| Silicon | custom 5nm chip |
| Vehicle | OEM'd from a Chinese manufacturer |
Upstream — the bill of materials
Method: Reverse bill of materials. ⚠️ All on-air from Moonshots #283 — none of it filing-verified.
| input | supplier | note |
|---|---|---|
| Autonomy compute | custom 5nm silicon | in-housed — the cost collapse came from here |
| Vehicle platform | OEM'd, Chinese manufacturer | the buy-don't-build half of the decision |
| Lidar / sensing | not mapped | historically the dominant cost line |
The whole story is in one row: hardware $115k → $20k (−82.6%) in a single generation. That is a BOM being attacked deliberately — in-house the expensive part, outsource the commodity part.
Downstream — who buys it
| customer / segment | share | note |
|---|---|---|
| Rider-hailing, per-city | n/a | it sells rides, not vehicles — so the BOM is a cost line, not a revenue line |
Position
Waymo is the inverted case and that is why it earns a page. NVIDIA sells the parts; Waymo buys them and its whole strategy is BOM attack. Two directions on the same method.
⭐ And it is the cleanest available worked example of buy-the-cheap-stack vs build-your-own — the
same decision he faces on studio hardware. Register idea 283-05 pairs it with a decision he actually
made this year → the ideas register.
Where it came up
- Moonshots #283
1:38:36— the cost collapse, and the buy-vs-build turn - Moonshots #284 — carried forward in the hardware-sourcing thread
- Moonshots #unnumbered
0:06:50,0:21:05— Astro Teller: "Waymo is at the top of the list" of what came out of X; "I'm talking about the cost of building Waymo. I guarantee that the value of Waymo is so huge compared to not only the cost of that team, but also all the other things that didn't work". No figure given; the guest's own claim.
Why it matters here, beyond cars
It is the cleanest available case of buy the cheap stack vs build your own — the same decision he
faces on studio hardware. Register idea 283-05 pairs it with a decision he actually made this year.
⚠️ Unverified
Every number above is from the tape. Run verify-first before any of it reaches a script.
EP307 (Moonshots #307, aired 2026-10-31): the scale today, a chart against it, and a union in Cambridge
- Scale (Peter,
1:23:50): "Whimo has 4,000 robo taxis on the road, 500,000 paid rides per week… 15 cities, fully driverless." ✅ ~4,000 cars, >500k paid rides a week, 15 US metros (Reuters, 20 Oct). ⚠️ The China-built Zeekr vans "facing tariffs" comes from X posts only. - The chart: the Cybercab-vs-Waymo cents-per-mile figures are read as ARK's; ARK's own note is a 2031 projection. See Tesla and Cathie Wood. Dave: "I wouldn't write off Whimo/Google anytime soon", because the winner is whoever builds a chip fab first (
1:24:34). - Boston: Alex says Waymo "just opened up a very large repair facility or maintenance facility" north of Boston (
1:29:56). ⚠️ It leased 65,000 sq ft at 180 Charles St, Malden, use undisclosed. - Cambridge: Teamsters Local 25, with Boston Carmen's Union Local 589 and the App Drivers Union, rallied at City Hall on 22 Oct (Cambridge Day), told on tape as "on Thursday" (
1:33:08). The panel's history lesson has one error (the Red Flag Act's flag went in 1878, not 1896) and two misses (Fort Lauderdale was 2015; "5,000 less dead kids" and "half the court cases" fail). See the EP307 anatomy §10.