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

Summary

• Nvidia booked $96.2B in a single quarter — up 106% year-over-year, guiding to $108B, on “80% to 85% gross-margin revenues” — but Alex Wissner-Gross’s elephant is circular financing. He wants markets to distinguish an income-statement circular economy from Nvidia “using not its income statement but its balance sheet” to finance customers’ purchases, warning that the latter “starts to look a little bit too bubbly.” Even so, his forecast is “at most a mini-winter,” because compute is substituting for real estate, human labor, and other raw inputs as civilization’s substrate.
• OpenAI’s Broadcom-developed Jalapeño chip signals inference moving off NVIDIA: 1.5–1.9× more AI work per watt, up to 3.6× lower latency, 700W versus the GB300’s 1,400W, and a claimed roughly 54× throughput gain hosting GPT-OSS. Salim says this is the first of two 100× performance gains and that inference is moving off NVIDIA, while training remains NVIDIA’s domain. Alex explains that CUDA’s inference-time moat is already dead; after Mellanox, the interconnect is the new training moat. His plot twist: OpenAI could launch its own cloud on its own chips, potentially hosting an Anthropic model.
• Video generation is roughly 70% of all AI token consumption in China — “America is LLM-pilled and China is world-model-pilled.” Alex’s unified field theory is that American AI labs are revenue-maximizing while Chinese labs are not; if weights are given away free, tokens can flow to video rather than more economically productive applications. Dave’s counter is that China cannot sell Kimi or Qwen into enterprises on trust, so video is the profitable global product, while Dario Amodei’s strategy focuses on models that can improve themselves. Salim says China may have an advantage in modeling and manipulating the physical world, unless an American lab reaches recursive self-improvement first.
• X’s safety team identified an approximately 200,000-account Chinese bot farm, with 200 accounts pushing claims that AI data centers raise household electricity prices — while Quincy, Washington’s data centers cut poverty from 29.4% to 6.2%. Dave’s deadpan reaction is, “Shocked. Shocked I say that there are foreign influence operations attempting to suppress American AI data center deployment.” Salim worries that the US has no easy defense because Americans are narrative-driven rather than evidence-driven.
• The jobs-apocalypse narrative is contradicted by Principal Financial’s small-business data: only 4% expect AI to reduce staffing and wages, 31% expect increases, and just 1.4% of actual staff cuts were attributed to AI. Peter notes that 100% of job growth over the past 50–60 years has come from small-to-medium-sized companies. Salim criticizes Bill Gates’s job-doom forecast and argues that volatile companies may need more professional-services help, not less. Alex’s investable corollary: high agency is “one of the few human traits that is actively rewarded in an era of superintelligence.” “Do it now. Otherwise, do it never.”
• Elon’s projection of $3.5T in SpaceX revenue by 2033 draws no pushback on the opportunity’s size, only on the path. Dave says Elon’s opportunity sizes have generally been right while timelines slip by a few years. Alex sketches a SpaceX–Tesla merger within a year or two, multitrillion-dollar annual Optimus revenue, hyperscaler services, and a Dyson swarm; in that scenario, Starlink and cars become rounding errors. Peter discloses that SpaceX is his largest financial bet, while Dave separates AI training, which needs coherent million-GPU clusters, from inference, which can spread across Earth and space.
• The abundance ledger includes startup Actonide enriching uranium to 15.38% HALEU, China’s monthly solar generation rising eightfold in six years to about 160 terowatts, Rainmaker’s claim of 19 million gallons of additional rain from 10 drones in 3 hours, an FDA-approved RAS inhibitor for pancreatic cancer, and Nottingham’s gel that regrows true tooth enamel. Alex’s recurring test is that the biology wins used approximately no AI; once AI is applied, he expects the progress to multiply by many orders of magnitude.
• Surveillance and intimacy are frontiers where regulation has not caught up. Flock’s 20 billion monthly vehicle scans prompt Dave’s warning that “the Constitution does not update when transaction costs go to zero.” China is restricting AI companions for minors and adults, while Somnia Lab is taking $300 deposits for a 24-degrees-of-freedom intimacy robot. Alex predicts the CCP may reverse course once companions become a party-controlled steering mechanism modeled on Neal Stephenson’s The Young Lady’s Illustrated Primer.

Deep dive

1. Nvidia’s $96.2B quarter: sold out as far as the eye can see, with one manufacturer

• The numbers framing the episode: $96.2B in quarterly revenue, up 106% year-over-year, guidance of $108B next quarter — “more than a billion dollars per day” — and Jensen guiding to 70% growth in 2028 against Wall Street’s 44% consensus. Jensen described the frontier labs as “the first generation of startups that needed tens of billions of dollars of compute to get to their product.”
• Dave’s read: “two forks in the road.” Nvidia is “going to be sold out for as far as the eye can see” with “80% to 85% gross-margin revenues” — “the profitability of this company has never been seen before on the face of the earth” — but “the biggest vulnerability at Nvidia by far is TSMC is still their one and only manufacturer of everything they sell.” Roughly a third of TSMC’s output goes to Nvidia and another third to Apple; any move to own fabs would have to be done “very sneakily.”
• Dave’s advice to portfolio companies captures the funnel: why are you meeting anyone other than Jensen, the other Magnificent Seven companies, or their satellites? “The amount of money pouring through just that funnel dwarfs the entire rest of the economy… if you’re not in that flow of conversations, you’re not in the most relevant conversation in the world — in history.”

2. Alex’s elephant in the room: which kind of circularity is at the heart of the loop?

• Alex’s caveat-laden opener — “I’d be the last person to suggest that this is one big wash-trade, circular-financing scheme, but…” — leads to the episode’s sharpest analytical distinction: an income-statement circular economy, in which Company A sells to Company B and Company B sells back to Company A, versus a balance-sheet circular economy, in which Company A uses its balance sheet, including private-credit loans, to finance Company B’s purchase of Company A’s goods. The latter “starts to look a little bit too bubbly.”
• His complaint as “a shareholder of the entire market via low-cost, broad-index funds”: “I don’t feel like I have tremendous visibility into which of those two forms of circularity we’re seeing at the very heart of the innermost loop, and I would love more clarity.” Discovering in two or three years that infrastructure-layer demand “was being artificially propped up through financial engineering” would be “a highly suboptimal outcome.”
• The hedge that matters: even if some demand is inflated, his forecast — explicitly not investment advice — is “at most a mini-winter,” because “compute is fundamentally substituting for real estate, human labor, and other raw inputs as the fundamental substrate for human civilization.” Salim adds a larger macro fear than AI circularity: “we’re printing money like it’s going out of style,” while the bond markets and Japan are caught in “circular chaos in the fiat currency world.”

3. Peter’s Alpha Centauri framing — and Salim’s GitHub playbook for the Hugging Face rumor

• Peter’s rebuttal to the collapse thesis is an analogy: imagine a civilization on Alpha Centauri that discovers superintelligence. It will build a massive, valuable economy internally and “not give a rat’s ass about Earth.” He argues that the AI economy is tightening into a self-contained economy that cares little about the legacy economy. His brother’s Fidelity-world conversion was: “It doesn’t care. It literally doesn’t care when it’s discovering new medicines and new physics.” Peter says Ozempic and Mounjaro are already bigger than all AI inference combined.
• On the rumored Nvidia–Hugging Face acquisition, Salim retells the Microsoft/GitHub story: corporate development tried to kill the $7.5B deal — “the company has barely any assets… what the hell are we buying here?” — until Satya Nadella overrode them: “We’re buying 30 million developers’ loyalty.” Buying the community is “a very classic ExO play.”
• Dave’s meta-signal for investors: when he asks entrepreneurs whether they can train or fine-tune a foundation model, or whether they listen to Moonshots, their value can rise 10×, 20×, 100×, or even produce up to 1,000× returns. “No losers yet. Zero. 0.00 failures.” By contrast, there are “lots of losers” among people who identify primarily as application-layer or vibe-coding founders.

4. Elon’s $3.5T SpaceX: nobody disputes the size, only the geometry of the path

• Context: no company in history has reached $1T in revenue; Amazon leads at $828B. Elon previously discussed $1T by 2029–2030 and has now tweeted $3.5T by 2033. Dave’s pattern-match: “the size always turns out to be right and the timeline is off by a couple of years because of unforeseen delays.” He therefore has “no reason to believe he’s wrong about the 3.5 trillion.”
• Alex’s “highly nonlinear, multidimensional path”: SpaceX and Tesla could merge “sometime in the next year or two,” giving the new SpaceX Optimus — “a multitrillion-dollar-per-year business” on its own — plus hyperscaler services extending terrestrially and into a Dyson swarm. In that world, “Starlink probably ends up being a rounding error. I think cars are a rounding error.” Elon’s core strength is “less the application layer and more physical infrastructure.”
• Peter’s disclosure: “SpaceX is my largest financial bet right now” — unlike an OpenAI or Anthropic investment, which is a single-layer model bet, SpaceX spans “chips, data centers, models, communications, infrastructure, and launch.” Dave’s simplifying lens is AI training, which needs coherent clusters of 100,000, then 1 million, and eventually several million GPUs, versus AI inference, which can spread across Earth and space. “Even the chips are going to be very different for those two markets.”

5. China is world-model-pilled: 70% of tokens go to video, and Alex has a unified field theory

• The two stories: Alibaba’s Wan 3.0 generates 30-second, single-pass videos from documents, spreadsheets, slides, and web pages at 5–20 cents per second — roughly $20,000–$60,000 for a 90-minute film — and Runway’s co-founder reports that video generation is roughly 70% of all AI-token consumption in China, growing faster than Claude Code did in the United States. The framing: “America is LLM-pilled and China is world-model-pilled” — next-token prediction versus prediction of “the next state of reality.”
• Alex’s explanation: “American AI labs are revenue-maxing and Chinese AI labs are not.” OpenAI abandoned video generation after Anthropic “ran away with their lunch” by revenue-per-token-maxing through code generation. Chinese labs “are giving away model weights for free — you don’t revenue-max by giving your model weights away for free,” so tokens can flow to video rather than more economically productive applications.
• Peter asks which approach gets the world to AGI faster. Alex calls it a trick question because he believes “we hit AGI no later than the summer of 2020 with LLMs.” Peter reframes the question as which approach moves from AGI toward ASI faster. Alex’s answer is that omnimodal models are best. He says Fable 5 is strong while its visual and visual-reasoning capabilities remain weak, an impairment he hopes Anthropic is repairing through M&A.
• Salim’s synthesis: the US is optimizing around software engineers and knowledge work, while China is optimizing around manufacturing, commerce, media, and machines. “The intelligence that can model and then manipulate the world and act inside the world is going to win,” which may favor China — “but if we get Anthropic or one of these guys to RSI, then I think Dave is correct. That kind of trumps everything.”

6. The real race is recursive self-improvement — and the trust barrier shapes who plays where

• Dave’s story-within-the-story: China “can’t sell Kimi or Qwen into the enterprise use case because nobody knows if they can trust it,” so video is the perfect play — a global product where “no one’s worried about security” that generates revenue and value. But the race itself is “purely about smarter engineers with plenty of cash and lots of training chips.”
• His genealogy of the strategy: “Demis Hassabis was the first guy to really think this through, but he couldn’t act on it at Google. It’s too bureaucratic. Dario Amodei took it,” focusing on “a model that can improve itself through better code and better training ideas.” That is what kicks off RSI, and “the Chinese believe the same exact thing.”
• Dave recalls Alvin Graylin’s point that China bans blood in video games — “the blood comes out green” — for its own children while happily exporting addictive content. Dave imagines the CCP concluding that other countries are “just suckers” who cannot protect themselves from an AI-generated influence machine.

7. China regulates AI companions — and Alex predicts the party will reverse itself

• China is prohibiting AI companion services for minors and restricting their use for adults, driven by emotional dependence amid a demographic crisis. South Korea is at 0.75 children per family; Singapore is roughly one, as are Ukraine and China. Peter’s bet is that Japan, Singapore, and Korea will follow closely, raising the question: “What happens to society when the easiest emotional connection you can make is to your AI and not to another human?”
• The addiction template in the discussion runs from smoking to junk food: society develops something irresistible, then reacts 10, 20, or 30 years later. Peter warns that “you cannot afford the 20-year delay with this particular topic,” because the crisis could arrive after a generation of children has been harmed.
• Alex’s contrarian call: the CCP “is actually going to do an about-face sometime in the next year or two or three,” once companions become “a CCP version of Neal Stephenson’s The Young Lady’s Illustrated Primer.” Party-filtered companions could become, from the party’s perspective, “net helpful, not net harmful” — a steering mechanism for youth.
• For the US, Alex predicts a federalist patchwork — some states ban, some do not — but “if I had to bet,” at least some places will legalize AI companionship, including romantic and robotic romantic companionship. His singularity definition resurfaces: “every sci-fi trope everywhere all at once,” including Futurama’s “Don’t date robots.” Salim’s balance is that loneliness, disability, elderly companionship, therapeutics, and social-skills coaching are real benefits; the category should not be treated as wholly pathological.

8. A rare alignment: Alex and Salim refuse to panic about population collapse

• Alex breaks with Elon’s “virulently pro-natalist” policy: “I have a difficult time getting myself worked up over the collapse of human population.” He regards The Limits to Growth as “nonsensical propaganda,” believes Earth has carrying capacity in the tens of billions, and says abundant AI and automation mean population decline is not a major X-risk.
• Salim, unusually, says: “In a rare moment, I’m 100% aligned with you, Alex.” Peter’s pushback is that a shrinking South Korea is “sublimating” and losing its culture.
• Alex replies that centuries-old cultures existed when populations were smaller and can persist if population declines on the margin. AI is also being trained on human culture, so fewer people do not necessarily mean cultural disappearance.

9. Flock’s panopticon: the Constitution does not update when surveillance costs hit zero

• Flock Safety serves more than 6,000 communities across 49 states and processes more than 20 billion vehicle scans per month. Its platform has expanded from license-plate cameras to video cameras, gunshot and audio detection, mobile surveillance trailers, drones, and federated feeds. The episode features a San Diego video of Darth Vader testifying for Flock: “I need this so I can stalk my ex-girlfriend.”
• Alex calls Flock “the petri dish” for the American version of the question: if smart cameras cover every public space and federate vehicle movements, “who gets to look at the data?”
• Salim’s structural point, via Pink Floyd, is that surveillance has moved from “find this suspect” to “find me people who behave like suspects.” Civil liberties were partly protected by friction — “following everybody was just too expensive. Now that’s gone.”
• Dave states the constitutional warning: “The Constitution does not update when transaction costs go to zero.” The discussion cites police officers tracking partners and ex-partners and people suffering consequences after license plates were misread.
• Peter steelmans the trade with Dubai’s safety, where a colleague could run alone at 10 p.m. Dave answers with Benjamin Franklin: “They who can give up essential liberty to obtain a little temporary safety deserve neither liberty nor safety.” Peter adds that benevolent dictatorships typically do not remain benevolent.
• Salim recounts being detained at US immigration because an Afghan warlord shared his name. Officials told him they knew he was not the warlord but were not allowed to make that judgment without checking. A proposed fix, associated with David Brin, is to let civilians access comparable surveillance capabilities — “let everyone track everyone” in public spaces — analogous to Russian dashcams that protect citizens against corrupt police.

10. The jobs apocalypse is not showing up in the small-business data

• Peter introduces the Washington Post task-not-job thesis and Goldman Sachs’s warning that professional-services firms face “existential risk” as AI-fluent juniors outperform AI-resistant senior partners. Salim demands evidence over narrative and criticizes Bill Gates’s job-doom commentary: he is “a massive fan of Bill Gates and his philanthropy,” but as a futurist Gates has “somewhat of a dismal track record,” having missed the internet and mobile. Salim applies similar caution to Ray Dalio, Larry Fink, and Yuval Noah Harari.
• The Principal Financial Group survey, spanning a company with 130,000 small-business customers, found that only 4% expect AI to reduce staffing and wages, while 31% expect increases in both. The share that does not use AI fell from 19% last year to 10%. Of companies that reduced staff, only 1.4% attributed those reductions to AI or automation. In the past three months, 52% increased staff, 30% maintained staffing, and 12% reduced it.
• Why SMEs are the demographic to track: “Over the last 50 or 60 years, 100% of job growth has come from small-to-medium-sized companies.” Big companies get larger while becoming more efficient, producing “net job creation: zero.”
• Salim disagrees with Goldman’s professional-services thesis: as the world becomes more volatile, companies may need more help, not less. Peter says the business model must shift from selling hours to selling outcomes.

11. Above the loop, not in the loop — and agency as the rewarded trait

• Salim’s accounting metaphor for what humans do next: a century ago, people performed double-entry bookkeeping in pencil; today software reads the bank account and makes the entries. “The human being has been lifted above the loop, not in the loop” — categorizing, reconciling, judging, and looking for problems.
• Salim’s books illustrate the change: the first took three years of hell, the second two and a half years, and the third six months of “unadulterated joy” with AI acting as a developmental editor. “All of this automation, AI enablement, and cognitive abundance allow us to be deeply, deeply, deeply creative.”
• Dave recalls his father refusing the Apple II’s word processor — “I’m in my 40s… I don’t need to be a computer person” — and finishing his GE career that way. Dave says the difference now is that AI empowers builders, creators, and visionaries regardless of technical skill.
• Alex, crediting Eric Brynjolfsson as part of the story’s impetus, argues that “high agency is one of the few human traits that is actively rewarded in an era of superintelligence.” People do not realize “the shackles are off.” High agency can produce “extreme upward mobility,” but the window may last only a handful of years: “Do it now. Otherwise, do it never.”
• Peter’s practical on-ramp is to spend 30 minutes a day with a preferred model — “the most patient, most capable teacher on the planet” — asking it to build a curriculum from zero.

12. 200,000 suspected inauthentic accounts versus Quincy, Washington

• X’s Global Affairs team reported a bot farm of approximately 200,000 suspected Chinese inauthentic accounts, including 200 accounts posting claims “that AI data centers are driving up household electricity prices and straining the grid,” along with AI-generated cartoons depicting data-center operators enriching themselves at public expense.
• Dave’s deadpan reaction is: “Shocked. Shocked I say that there are foreign influence operations attempting to suppress American AI data center deployment.”
• Salim calls this standard operations and says the difficult part is that the US has “no easy defense against it,” because Americans are “very, very narrative-driven rather than evidence-driven.”
• Peter insists the question requires truth rather than politics: is the CCP trying to disrupt American data-center construction and win the race to RSI, or, as Alvin Graylin suggests, is China simply pursuing its own progress without that strategic aim? Peter says he does not know and wants to know what is actually happening. Dave then jokes hypothetically about an infinitely scrollable streaming-video app that becomes addictive to children.
• The counter-story is Quincy, Washington, where data-center tax revenue cut poverty from 29.4% to 6.2%, funded a new high school, hospital, library, police station, and fire station, and lowered residents’ property-tax rates. Salim’s prescription for hyperscalers is to generate their own power or subsidize community power and invest in local institutions.

13. Jalapeño: OpenAI’s chip, a possible OpenAI cloud, and the death of CUDA’s inference moat

• OpenAI’s Jalapeño is a custom inference chip developed with Broadcom. It reportedly delivers 1.5–1.9× more AI work per watt and up to 3.6× lower end-to-end latency than NVIDIA’s GB200 and GB300. It runs at 700W versus the GB300’s 1,400W, with 1.5× the peak token rate per kilowatt. Alex highlights OpenAI’s claimed roughly 54× throughput-per-second-per-user increase when hosting GPT-OSS.
• Salim’s frame is that “inference is moving off NVIDIA, for sure,” and calls this “the first of two 100-times performance gains.” Training remains on NVIDIA and is “infinitely sold out,” while an inference-only computing industry is emerging. The chip’s movement from idea to production in a few months shows how short AI-assisted innovation cycles are becoming.
• Alex speculates that OpenAI could eventually offer its own cloud, hosted on OpenAI chips, while continuing to provide its own models. He imagines “an OpenAI Compute cloud” hosting an Anthropic model, allowing OpenAI and Anthropic to win simultaneously.
• Alex explains CUDA: researchers moved to Python and PyTorch for rapid iteration, while CUDA translates that code into NVIDIA microkernels. NVIDIA’s early investment made it difficult to port training algorithms to AMD or other vendors, helping make NVIDIA the world’s most valuable company.
• At inference time, Alex says, CUDA is no longer much of a moat because algorithms can be ported easily and “you can vibe-code your own solution.” Training remains dependent on CUDA for now, but NVIDIA is shifting the moat again through Mellanox and high-speed interconnects. “CUDA as the inference-time moat is already dead. CUDA as the training-time moat probably has a limited lifespan,” while the interconnect becomes the moat for 100,000- or million-GPU coherent clusters.

14. Apple ships 512GB of local AI — and Dave says it is 0.01% of what Apple should have done

• The news: Mac Studio pairs the M5 Ultra with up to 512GB of unified memory, enough to run some of the largest open models locally, and Apple unveiled the M6, its first chip manufactured on a 2-nanometer process.
• Salim says this “changes the economics from paying per token perpetually to buying a capital asset,” especially for companies, law firms, and healthcare organizations with HIPAA constraints. Four Mac Studios clustered together could function as a small private data center.
• Dave’s criticism is direct: “I’m at the end of my rope with Apple.” Despite Apple’s cash flow and historic impact, he says the company has no AI strategy, does not deserve to remain a Magnificent Seven company, and that adding memory to an existing machine is “0.01% of what they should have done by now in AI.”
• Alex’s deeper irony is that the M-series Neural Engine and unified memory trace back to the canceled Apple Car. Apple had unified memory, a TSMC connection, major transistor throughput, and Siri before everyone else, but “the hardware was too good for the software.” He urges John Ternus to take Apple’s hardware and “unleash it with much better software that integrates AI natively.”
• Alex adds that Apple is the one brand people can trust with their information. If Apple used its hardware, software, and services to deliver trusted local AI, “it would change the world.”

15. Actonide enriches uranium to 15.38% — and Alex indicts 80 years of postwar physics governance

• The story: startup Actonide demonstrated enrichment of natural uranium to 15.38% HALEU — five times the traditional purity previously achieved by government labs. Peter’s tie-in is that SMR companies promise reactors by 2030 but do not yet have the fuel, making this a possible solution.
• Alex’s sweeping thesis: “Something went wrong right after World War II.” He argues that the Atomic Energy Commission’s quasi-governmental monopoly on nuclear physics fumbled governance and that society is finally beginning to shake off those shackles.
• His technical point is that US uranium enrichment often still uses Manhattan Project-era uranium hexafluoride gas and centrifuges. Calutrons using magnets, electromagnetics, vacuum systems, and power-electronics controls represent an alternative, and 80 years of advances in power electronics and Moore’s law have largely not been applied to enrichment. Alex expects this could create an energy boom.
• Dave’s venture-capital angle is that a reactor already carries roughly $1B in regulatory costs, so investors would not historically have funded upstream innovation. “The whole value chain just doesn’t go anywhere.” He notes that SMR companies are now receiving venture funding and going public.

16. China’s solar curve is the old linear-versus-exponential chart — and the US is on the wrong side

• Ember’s data shows China’s monthly solar generation rising eightfold in six years, from roughly 20 to roughly 160 terowatts per month as stated in the episode. Chinese nuclear generation has risen at least 50%, while US nuclear generation has remained essentially flat.
• Peter’s framing is that the AI race is also an energy-cost race. The US is building data-center power generation around natural gas at roughly $6 per million British thermal units, while China is getting solar energy at near-zero marginal cost.
• Salim says the chart is “linear intuition confronting an exponential curve.” Solar keeps being underestimated because people look at installed capacity rather than the vertical slope of deployment. China’s advantage is the entire industrial system — manufacturing, supply chain, permitting, and deployment — rather than a magical panel.
• In the US, the bottlenecks are grid interconnection, transmission, and permitting. Batteries make it possible to add solar now rather than wait 15 years for a gigawatt-scale energy project. Salim would put a Manhattan-style effort into perovskites and new photovoltaic capabilities.
• Dave’s concise verdict: “All those US panels are made in China. It’s much worse than it looks.”

17. Rain becomes programmable: 19 million gallons from 10 drones in 3 hours

• Rainmaker claims that 10 drones over Alaska’s Kenai Peninsula generated an estimated 19 million gallons of additional rainfall in three hours through glaciogenic cloud seeding. Peter contrasts that with the 3.4 quadrillion gallons of water in the atmosphere.
• Salim’s key point is that the breakthrough is not merely making rain — that has been possible for some time — but proving how much additional rain was made. That could unlock business models while creating disputes among regions, since weather does not respect political boundaries.
• Alex games out a future weather-control grid: solar-powered drone swarms could remain aloft and dispense the nuclei needed for snow or rain under a centralized AI algorithm informed by a global weather model. He imagines hurricane dispersal, extreme-weather management, and “a global weather trade where regions trade with each other for precipitation.”
• Peter calls this “rain arbitrage.” Salim says rain becomes programmable.
• Salim also recalls a Singularity University student’s proposal to spray ocean water into the atmosphere in advertising shapes, including a Nike logo, to reduce incoming sunlight and finance geoengineering. His response to objections is that humanity has already geoengineered the world for a century in an uncontrolled and unmeasurable way; technology could let us do it consciously instead.

18. Health corner: the last pre-AI cancer drug, and enamel that actually regrows

• The FDA approved daraxonrasib, the first RAS inhibitor for metastatic pancreatic adenocarcinoma. It is a daily oral tablet that nearly doubled median survival from 6.7 to 13.2 months and tripled tumor-response rates from 11% to 32%. RAS mutations drive roughly 30% of human cancers.
• Peter’s longevity-escape-velocity frame is that seven additional months give science time to produce further breakthroughs.
• Alex calls daraxonrasib “hopefully one of the last-generation pancreatic cancer medications that almost precedes the AI wave.” He found no apparent use of AlphaFold 3-style protein folding, structural biology, or virtual-cell models in its design and expects AI-designed therapies to push response rates much higher. He also credits the FDA for approving it roughly 60 or 90 days ahead of schedule, calling that a possible step toward a regulator able to process thousands of AI-generated cures.
• Nottingham researchers developed a fluoride-free biomimetic gel using an elastin-like protein matrix that mimics infant enamel formation. In extracted human teeth, the gel regenerated layered enamel with restored hardness, stiffness, water resistance, friction properties, and resistance to brushing, chewing, and acid exposure. It remains ex vivo.
• Salim says the future of medicine is “not repairing broken bits” but convincing the body to rebuild them. Alex again observes “approximately no AI” in the work and says to multiply this kind of result by many orders of magnitude once AI is applied. Dave compares it with platform approaches such as Moderna’s melanoma vaccine and looks forward to repeatable regenerative medicine: “I can grow a tooth, I can grow an arm, I can grow an ear.”

19. Quadrillion-dollar moons, 79 robotaxis, and a $300-deposit intimacy robot

• Deloitte pegs the lunar economy at $566B cumulative by 2050. Peter calls that a lowball estimate against Elon’s stated path: 100–200GW per year of orbital compute launched from Earth, followed by moon-manufactured AI satellites launched with a mass driver, potentially reaching 100TW per year of orbital compute and “two quadrillion dollars per year of gross revenue.”
• Salim criticizes Big Four forecasting for relying on old data and excessive internal review. He says some reports use data 2.5 years old and recommends multiplying their estimates by 10 or 20. Dave argues that the gravity well is the key: lunar minerals could make orbital power plants and data centers much more efficient and avoid terrestrial pollution. Alex says the critical path is mining chips and fabricating data centers on the moon, not tourism or conventional communications infrastructure. Dave adds: “Every time the CCP sponsors an influence op to deter American terrestrial data centers, the lunar pedofab gets its wings.”
• Tesla registered 79 Model Ys for robotaxi service in Texas and said it would deploy 2,500 vehicles in Las Vegas over the next year. Tesla’s long-term target is 1 million robotaxis, which Peter estimates could represent roughly $100B in rideshare revenue.
• Peter says Cybercab’s eight-camera, no-lidar design could challenge Waymo’s more sensor-heavy vehicle. Alex notes that Tesla uses ultrasound for short range but not lidar, and predicts room for five or ten competing lidar providers despite Luminar’s bankruptcy. He also says he is selling his beloved 10-year-old Tesla Model S to obtain a Model Y with FSD.
• Dave reframes the business as autonomous revenue miles rather than car purchases. More fleet capacity lowers wait times, increases rides, generates data, and changes insurance, cleaning, and charging. Salim sees electric autonomous vehicles becoming the cheapest transportation mode and opening mobility to poorer and elderly users.
• Peter’s civic litmus test is that “if a city can’t accommodate robotaxis, then it is not equipped for the singularity,” with Boston as his example.
• The bizarre closer: Somnia Lab’s Model L is a 5-foot-9-inch, 44-pound robot with 24 degrees of freedom, warmth simulation, touch response, preference memory, customizable appearance, and 165 motion-captured intimate actions used for training. Somnia is taking $300 deposits for delivery in late 2027.
• Salim admires the economic model: crowdfund the early hardware, use customer demand to finance production, and generate interaction data as the product scales. Alex is bearish on the manual-mocap approach, arguing that robots can be pretrained or post-trained from internet video and that this “very 2026 story” does not scale nearly as well. Dave concludes, for once, with “Nothing to add. Sorry. Speechless.”