2026-08-29

NVIDIA's $96.2B Quarter, China's 200,000 Fake Accounts, & OpenAI's New Chip | EP #284

The panel (Emad Mostaque absent this week) opens on Nvidia's record $96.2B quarter and whether its financing of its own customers is quietly inflating a bubble, then pivots to Elon Musk's $3.5T SpaceX revenue bet and a possible SpaceX-Tesla merger. China dominates the middle third: Alibaba's Wan 3.0 and the claim that video generation is 70% of Chinese AI token use, a new crackdown on AI companions amid a demographic collapse, a Flock Safety surveillance backlash (Darth Vader testifies at a San Diego city council), and an X-confirmed 200,000-account Chinese bot farm pushing anti-data-center narratives. A hard-data detour pushes back on AI-jobs doom (Goldman Sachs vs. a 130,000-business survey), followed by OpenAI's new Jalapeno inference chip challenging Nvidia, Apple's belated AI catch-up with the M5 Ultra Mac Studio, and closing stories on uranium-enrichment startup Actinide, Rain Maker's cloud-seeding drones, a pancreatic-cancer drug approval, tooth-enamel regeneration, lunar-economy estimates, Tesla's Cyber Cab rollout, and a robot intimacy company.

▶ Watch on YouTube

Topics

Nvidia's $96.2B quarter and circular-financing bubble worries Compute ▶ 0:05:08
Nvidia posted $96.2B in quarterly revenue (up 106% YoY, 80-85% gross margins) with Jensen Huang guiding 70% growth through 2028; Alex Wissner-Gross flags concern that markets can't tell how much of that demand is Nvidia financing its own customers.
Elon Musk's $3.5T SpaceX bet and a possible SpaceX-Tesla merger Space ▶ 0:06:09
Musk tweeted SpaceX could hit $3.5T in revenue by 2033 (vs. Amazon's current $828B, the highest ever); Alex Wissner-Gross sketches a path via a SpaceX-Tesla merger combining Optimus robot revenue with orbital data-center and Dyson-swarm revenue.
China's world-model bet: Wan 3.0 and 70% of AI video tokens AI ▶ 0:25:00
Alibaba's Wan 3.0 generates 30-second videos from documents for 5-20 cents/second; a Runway co-founder claims video generation is ~70% of Chinese AI token use, framed as America betting on language models ('LLM-pilled') while China bets on world models ('world-model-pilled').
China tightens AI companion rules amid a demographic collapse AI ▶ 0:35:02
China is restricting AI companion apps for minors and adults, citing loneliness and collapsing birth rates (China, South Korea, Singapore, Ukraine all near or below 1 child per family); the panel debates whether the US should regulate similarly and whether the CCP will eventually embrace AI companions as a control tool.
Flock Safety's license-plate surveillance network sparks backlash AI ▶ 0:45:07
Flock Safety's AI license-plate cameras (6,000+ communities, 20B scans/month) are dramatized via a Darth Vader parody testifying in San Diego; the panel debates civil-liberties costs, cites cases of police misusing the system to stalk exes, and invokes the Ben Franklin liberty-vs-safety quote.
Job-apocalypse skepticism: Goldman Sachs vs. hard SMB data Economy ▶ 1:07:45
A Washington Post piece argues AI displaces tasks, not jobs; Goldman Sachs warns professional-services firms face a dangerous skill gap between AI-fluent juniors and AI-resistant partners; Salim Ismail counters with a Principal Financial Group survey of 130,000 small businesses showing minimal AI-driven layoffs.
A 200,000-account Chinese bot farm targets US data-center sentiment Geopolitics ▶ 1:21:22
X's Safety team found a ~200,000-account bot farm, with 200 accounts pushing false claims that AI data centers raise household electricity prices, contrasted with Quincy, Washington, where data-center tax revenue cut poverty from 29.4% to 6.2%.
OpenAI's Jalapeno chip challenges Nvidia's inference dominance Compute ▶ 1:27:06
OpenAI's Broadcom-built Jalapeno inference chip delivers 1.5-1.9x more work per watt and up to 3.6x lower latency than Nvidia's GB200/GB300 at half the power; the panel discusses whether Nvidia's CUDA moat is dying for inference and speculates OpenAI could become its own hyperscaler.
Apple's belated AI catch-up: M5 Ultra Mac Studio and M6 chip Compute ▶ 1:33:17
Apple's Mac Studio pairs the M5 Ultra with up to 512GB unified memory (enabling large local models) and unveiled the 2nm M6 chip; Salim sees it as a big win for HIPAA-bound and law firms, while Dave Blundin blasts Apple for having essentially no AI strategy.
Energy and water abundance: uranium enrichment and cloud-seeding drones Energy ▶ 1:41:06
Startup Actinide enriched uranium to 15.38% purity (5x traditional levels) for SMR fuel; China's solar generation grew 8x in six years versus flat US nuclear output; Rain Maker's drones generated an estimated 19M gallons of rain over Alaska via cloud seeding, raising the prospect of programmable weather.
Health breakthroughs: pancreatic cancer drug and tooth-enamel regeneration Health ▶ 1:56:44
The FDA approved durasenib, nearly doubling median survival for metastatic pancreatic cancer; separately, University of Nottingham researchers created a biomimetic gel that regrows true tooth enamel in lab tests, with Alex noting neither breakthrough used AI-driven drug design yet.
Lunar economy estimates and Elon's Star Mine vision Space ▶ 2:05:25
Deloitte estimates a $566B lunar economy by 2050, which the panel calls a drastic underestimate versus Musk's Star Mine vision of manufacturing AI satellites on the Moon, projecting up to 100 terawatts/year of orbital compute worth roughly $2 quadrillion/year in gross revenue.
Tesla's Cyber Cab rollout accelerates the robotaxi race Transportation ▶ 2:12:47
Tesla registered 79 Model Ys for Cyber Cab service in Texas and plans 2,500 vehicles in Las Vegas within a year, aiming for 1M robotaxis long-term; the panel compares Tesla's camera-only 8-sensor approach favorably against Waymo's LiDAR-heavy 13-camera design.

Predictions made

open Elon Musk (reported): SpaceX will reach $3.5 trillion in annual revenue by 2033.
· due: 2033 · ▶ watch
“he may well reach 3.5 trillion in revenue in 2033”
Your call:
open Alex Wissner-Gross: SpaceX and Tesla will merge within the next year or two, combining Optimus robot revenue with orbital data-center and Dyson-swarm hyperscaler revenue.
· due: 2027-2028 · ▶ watch
“SpaceX and Tesla merge sometime in the next sometime in the next year or two”
Your call:
open Alex Wissner-Gross: The Chinese Communist Party will reverse its current crackdown and embrace AI companions within a few years as a tool for ideological control over youth.
· due: 2027-2029 · ▶ watch
“once the CCP has decided that it has a way to steer the country's youth via AI companions it will decide no actually we like these AI companions after all”
Your call:
open Alex Wissner-Gross: At least some US states will legalize AI companionship, including romantic and robotic companionship.
· due: unspecified · ▶ watch
“at least in some places in the US AI companionship including romantic companionship including robotic romantic companionship will be legal”
Your call:
open Alex Wissner-Gross: OpenAI could launch its own cloud/hyperscaler offering hosted on its own chips, potentially even hosting rival models like Anthropic's.
· due: unspecified, next 1-2 years · ▶ watch
“an open AI compute cloud hosts an anthropic model and OpenAI and Anthropic win at the same time”
Your call:
open Elon Musk (reported by Peter Diamandis): Lunar-manufactured orbital AI compute could scale to 100 terawatts/year, generating roughly $2 quadrillion per year in gross revenue.
· due: unspecified · ▶ watch
“the estimate if you're building a 100 terowatts per year of orbital compute the economy would be two quadrillion dollar per year of gross revenue”
Your call:
open Salim Ismail: His son will never get a driver's license because robotaxis will fully roll out before he'd need one.
· due: ~2027 · ▶ watch
“My kid will never get a driver's license”
Your call:

Numbers that matter

Worth digging into

🕳️ Is Nvidia financing its own demand?
Alex Wissner-Gross flags concern that a meaningful share of AI infrastructure demand may be backstopped by Nvidia's own financing of customers and data centers, which would make current growth numbers look more bubble-like than they appear.
🕳️ OpenAI becoming its own hyperscaler
Jalapeno's startling ~54x throughput claim on GPT-OSS raises the question of whether OpenAI could pivot to selling compute on its own chips, potentially even hosting competitor models like Anthropic's.
🕳️ The truth behind the Chinese data-center bot farm
The panel admits genuine uncertainty about whether this reflects deliberate CCP strategy to slow US AI dominance or unrelated inauthentic activity, and explicitly calls for evidence over politics.
🕳️ Post-WWII nuclear governance stagnation
Alex Wissner-Gross argues the Atomic Energy Commission's formation after the Manhattan Project locked in outdated centrifuge-based enrichment technology for 80 years, only now being unwound by startups like Actinide.
🕳️ Lunar manufacturing's real killer app
Alex Wissner-Gross argues Deloitte's $566B lunar economy estimate radically undercounts the opportunity because it doesn't model what industries get unlocked once launch costs drop an order of magnitude, with chip and data-center fabrication on the Moon as the actual critical path over tourism or NASA missions.