Elon Enters the Chip Race, the S&P 500 Repricing, and Human Drivers Will Become Illegal | EP #242
The panel debuts a new deep-dive format built around fewer, bigger themes: Elon Musk's newly announced Terafab (a joint Tesla/xAI/SpaceX push to build 50x the world's current AI compute output), the collapse of human driving as Waymo and eVTOL data pile up, AI-driven workforce disruption (PwC's 'adapt or leave' memo, employee token-usage tracking), Chamath Palihapitiya's thesis that AI could compress S&P 500 valuations by up to 90%, a new NASA/SpaceX/China moon race alongside panspermia evidence from asteroid Ryugu, a mystery trillion-parameter model that turned out to be Xiaomi's, an AI agent that designed a working CPU in 12 hours, and a successful whole pig-brain cryopreservation from Nectome. The episode closes with an AMA covering compute location strategy, whether humanoid robots are over-engineered, AI as a war negotiator, and disease-curing timelines under a more Bayesian FDA. Overall tone is bullish-with-caveats: audacious builds are read as abundance-generating rather than monopolistic, with recurring worry that chip supply, not technology or demand, is the real bottleneck.
Terafab: Elon Musk's terawatt-scale AI chip factoryCompute▶ 0:00
Elon Musk announced a joint Tesla/xAI/SpaceX objective to build one terawatt of AI compute per year (50x current global output of 20 gigawatts), via a fully vertically-integrated Austin fab making both edge inference chips (robots/cars) and rad-hardened space chips, aiming eventually for petawatt-scale compute sourced from lunar mass drivers.
Dyson swarm and disassembling the moon for computeSpace▶ 9:49
Alex Wissner-Gross calculates that scaling lunar-mined compute to a petawatt or exawatt would consume roughly 3/100,000% to 3% of the moon's mass via electromagnetic mass drivers, framing Terafab as a possible cornerstone that unifies Tesla, xAI, and SpaceX into one industrial entity feeding a Dyson swarm.
Terafab economics, capex, and the SpaceX IPOEconomy▶ 11:33
The panel models Terafab's capex ($25B to start, $150B-$500B for full buildout) against comparables like TSMC ($1.7T) and Nvidia ($4.5T), concluding standard valuation models undershoot by an order of magnitude at 50x world chip output; prediction markets are already pricing a bigger SpaceX IPO ($2T+) partly on Terafab's SpaceX-majority ownership split.
Autonomous vehicles, robotaxi economics, and the death of car ownershipTransportation▶ 25:56
Waymo has logged 170 million autonomous miles with 92% fewer serious crashes across 3,000 vehicles in 10 cities; Uber invested $1.25B in Rivian for 50,000 robo-taxis. The panel argues removing human drivers could cut ride costs roughly 10x, collapsing the ~100-million-car-a-year industry since cars currently sit idle 94% of the time.
Joby Aviation is in FAA conforming-aircraft testing after Golden Gate Bridge demo flights and an Uber Air partnership; Archer Aviation's Midnight aircraft achieved 100% FAA means-of-compliance acceptance. Panel expects piloted eVTOL service within 18 months, autonomous within ~2 years, limited initially by FAA rather than technology.
When will human driving become illegalTransportation▶ 32:34
Panel debates whether human driving gets formally banned (starting in city centers, spreading like anti-smoking/anti-drunk-driving campaigns) versus simply redefined into an AI-supervised, always-safe 'human on the loop' hybrid model. Consensus: safety data and public pressure will move fast, but chip shortages, not technology or demand, will be the real limiting factor.
Real estate and urban redesign in an autonomous-vehicle worldEconomy▶ 38:44
With vehicles no longer needing to park, garages become obsolete (Peter's household already converted one to a bedroom), 60% of Los Angeles land area currently used for parking could be reclaimed as parks, remote/scenic real estate becomes far more accessible and valuable, and Alex floats a 'request for startups' around autonomous Winnebago-offices that synchronize your calendar with your vehicle's location.
AI-driven workforce disruption and the token-vs-payroll shiftEconomy▶ 47:06
Goldman estimates AI could automate 25% of US work hours; PwC told partners to adopt AI or leave; G42 posted a job listing seemingly for an AI agent. Jensen Huang says a $500K engineer should consume at least 250,000 tokens or he'd be 'deeply alarmed.' Dave Blundin's companies already target an 80% token-cost/20% salary-cost mix, with all prompt histories captured for AI-driven performance review.
Chamath's terminal value collapse thesis and S&P 500 repricingEconomy▶ 1:02:08
Chamath Palihapitiya argues AI destroys the assumption that competitive moats persist, potentially compressing the S&P 500's average 22x free-cash-flow multiple to 7x (-2/3 value) or even 2x (-90%). Dave and Alex push back that capital doesn't vanish, it reallocates to infrastructure and agile management teams; Salim notes non-S&P mid-caps are already trading near 7x, and expects AI automation to triple free cash flow at many of those firms even as multiples compress.
New space race: NASA/SpaceX Artemis, China's 2030 moon push, and panspermiaSpace▶ 1:14:32
NASA is elevating SpaceX's Starship within the Artemis program (Boeing's Artemis 2 circumlunar flight targeted April 1) while China has confirmed intent to land on the moon by 2030. Separately, all five DNA/RNA nucleobases (A, G, C, T, U) were found on samples from asteroid Ryugu, and the NASA Administrator reportedly put >90% odds on imminently finding microbial life evidence on Mars, both bolstering panspermia.
Mystery AI model Hunter Alpha and the collapse of model moatsAI▶ 1:26:24
A free, anonymous trillion-parameter model with a 1-million-token context window ('Hunter Alpha') appeared on OpenRouter and processed 160B+ tokens before being revealed as Xiaomi's, sending its stock up 5.8%. Panel debates whether brand/capitalization still function as moats when distillation keeps compressing frontier capability into cheap, unattributed models, concluding that proprietary data (not model weights) is the more durable moat.
AI designs a CPU from concept to tapeout in 12 hoursCompute▶ 1:35:41
Vector's 'Design Conductor' AI agent autonomously designed a 1.5GHz Linux-capable RISC-V CPU from concept to tapeout in 12 hours versus a traditional ~90-day cycle, illustrating recursive self-improvement moving beyond software into chip design itself, with implications for cheap, use-case-specific (potentially 10x more efficient) custom silicon.
Data center politics: Genesis Project funding, Ohio's NIMBY ban, and Nvidia's China reversalGeopolitics▶ 1:40:28
The DOE's ~$300M Genesis Project funds AI-driven national challenges across manufacturing, biotech, and energy as a step toward US industrial policy. Meanwhile rural Ohio residents are pushing a constitutional amendment to ban data centers over 25 megawatts, and Nvidia won approval to resume selling H200 chips to Beijing after acknowledging its export ban had failed to slow China's chip development.
Whole pig-brain cryopreservation breakthrough at NectomeLongevity▶ 1:48:16
Nectome published results showing a full pig brain (organ scale comparable to humans) was chemically preserved with minimal cellular damage, marking real competition in the preservation/cryonics space alongside 21st Century Medicine's vitrification approach; Alex reiterates his standing call for listeners to sign up for cryonics.
Predictions made
openPeter Diamandis: Elon Musk's Terafab reaches its full one-terawatt-per-year AI compute target.
openSalim Ismail: A new non-photolithography approach to semiconductor manufacturing will be discovered within the next year or two using current LLM AI, but won't reach lithography-scale manufacturing volume for five to seven years.
“Something will get discovered in the next year or two using current LLM AI running on GPUs that will then dictate a very different non-lithography future. But it'll probably be five or six or seven years before those start getting manufactured at the same scale as lithography.”
Your call:
openPeter Diamandis: eVTOL flying car services will begin operating in the US within about 18 months, becoming prominent in Los Angeles by 2028.
“We're going to start to see them operating in the US in the next uh 18 months. Should be here in LA in 2028 in a big way.”
Your call:
openSalim Ismail: Full self-driving autonomy will be achieved within about two years, before his 14-year-old son is old enough to get a driver's license.
“I've got my 2 years for us to get to full autonomy before my son gets his driver's license. He's 14 right now.”
Your call:
openDave Blundin: eVTOLs will move from single-pilot operation to fully autonomous flight within about two years, once the AI command-and-control work already underway for self-driving cars extends to aircraft.
“there'll be a couple year very short period of time, in my opinion, 2 years or so.”
Your call:
openDave Blundin: A public tipping point (comparable to anti-drunk-driving or anti-smoking campaigns), including a TV ad campaign, will turn public opinion sharply against human driving on public roads within three to five years, though a chip shortage will be the main factor delaying full illegality.
“there'll be a there'll be a TV ad campaign probably 3, 4, 5 years from now with lots of ugly images in it... But, I think the thing that would make it later is purely the shortage of chips.”
Your call:
openAlex Wissner-Gross: Human driving will never be formally outlawed; instead, 'driving' will be redefined into progressively more abstract, always-safe human-supervised AI control (a human-machine hybrid).
“I think never. I think we'll simply redefine driving to represent higher and higher levels of abstraction.”
Your call:
openDave Blundin: AI-driven job disruption will be severe over the next two to three years, but the broader economy will recover and conditions will improve again by 2030-2032.
“there's going to be huge amounts of job disruption in the next two or three years and then it'll turn around and by 2030, 2032, things will be good again.”
Your call:
openAlex Wissner-Gross: AI has a realistic chance of solving most diseases within about five years, per the more optimistic estimates from organizations like CZI, Biohub, Arc Institute, and Anthropic.
“China has confirmed their intention to land on the moon by 2030.”
Your call:
Numbers that matter
20 gigawattsCurrent total global AI compute output, used as the baseline Elon's Terafab aims to exceed by 50x.
1 terawatt/year target (50x current output)Terafab's stated compute production objective across Tesla, xAI, and SpaceX.
$16 billion Samsung chip deal (up to ~$45 billion if fully realized)Elon's existing production commitment with Samsung, alongside building Terafab himself.
10 million tons of mass to orbit per year, 274 Starship launches per day (one every 5.3 minutes)Implied launch cadence required to hit Terafab's orbital compute buildout using next-gen Starlink-class satellites.
$25 billion initial capex; ~$150 billion (potentially up to $500 billion) for full buildoutEstimated cost to start and then fully scale the Terafab, versus TSMC's $1.7 trillion valuation.
TSMC valued at $1.7 trillion; Nvidia valued at $4.5 trillionComparables the panel says undershoot Terafab's implied value at 50x world chip output.
SpaceX IPO pricing moved from ~$1.5 trillion to $2+ trillion in prediction marketsReflects Terafab's SpaceX-majority ownership (80% SpaceX, 20% Tesla) getting priced into SpaceX's valuation.
170 million autonomous miles driven (equivalent to 200 human lifetimes), 92% fewer serious crashes, 3,000 vehicles across 10 citiesWaymo's safety and scale record cited as the case for accelerating human-driving restrictions.
$1.25 billion Uber investment in Rivian for 50,000 planned robo-taxisUber's autonomous fleet expansion plan.
Cybercab priced at roughly $30,000Elon's announced consumer price point, which Peter suggests buying in bulk to run as a personal robo-taxi fleet.
~100 million new cars made per year, sitting empty 94% of the timeBasis for the claim that autonomy collapses much of the existing car industry's market structure.
Ride costs could fall roughly 10x (estimated 10-30 cents per mile, 4-5x cheaper than owning a car)Projected effect of removing driver labor cost from ride-hailing.
60% of Los Angeles' land area is parkingAmount of urban land the panel expects to be freed up (parks, green space) once fewer parked personal vehicles are needed.
Goldman Sachs: AI could automate 25% of US work hoursCited as a possibly conservative estimate of AI's labor-market impact.
Jensen Huang: a $500,000 engineer should consume at least 250,000 tokens or he'd be 'deeply alarmed'Nvidia's framing for minimum expected AI token usage per high-value employee.
Worth digging into
🕳️ Terafab's physical feasibility
The plan implies 274 Starship launches per day and 10 million tons to orbit per year; it's unclear how realistic this launch cadence and lunar mass-driver roadmap actually is versus marketing framing.
🕳️ Chamath's terminal value collapse article
The 22x-to-2-7x free cash flow compression thesis is a provocative, checkable macro call that the panel is split on; worth verifying the original argument and tracking mid-cap multiples over time.
🕳️ Ryugu asteroid nucleobase discovery
Finding all five DNA/RNA bases on one asteroid sample is a major panspermia data point that deserves primary-source verification beyond the podcast's secondhand summary.
🕳️ Hunter Alpha mystery model episode
An anonymous trillion-parameter model beating attribution and moving a stock price on reveal is a clean case study for how fast AI moats are eroding.
🕳️ Nectome's whole pig-brain preservation paper
Alex Wissner-Gross says this is Nectome's first bioRxiv-style publication since their original Brain Preservation Foundation award-winning paper; worth reading directly given his informal advisory role.
🕳️ Vector's 'Design Conductor' AI chip designer
A claimed 12-hour concept-to-tapeout RISC-V CPU design (versus ~90 days) is an extraordinary recursive-self-improvement claim that warrants checking against the actual company/product documentation.