Jensen’s Spy Probe, Anthropic’s Debut Week, and Elon’s 30M Robot Workforce | #310
Peter DiamandisDave BlundinAlexander Wissner-GrossEmad Mostaque
- Unverified allegations involving Jensen Huang turn NVIDIA’s key-person risk into the panel’s central investment question. Dave Blundin expects a prolonged investigation and sees “no succession plan at all that I can detect”; the discussion cites NVIDIA falling 6% while AMD rose 10–11%. His larger concern is that a fabless company’s roughly $5 trillion of value depends on continued innovation and access to TSMC capacity, making leadership disruption unusually consequential.
- The speakers expect national-security controls to spread across frontier laboratories and Chinese AI models. Alex Wissner-Gross reads the allegations primarily as an influence operation and asks why previously visible GPU diversion is surfacing now. He forecasts security clearances and restrictions on Chinese models in government and regulated industries, while insisting that export-control distinctions concern nationality and access, rather than ethnicity. Peter raises the possibility of restrictions affecting foreign citizens and says he hopes naturalized U.S. citizens are not targeted.
- Anthropic’s proposed IPO pricing and scarce model access create two separate monetization arguments. Peter Diamandis reports a $1.3–1.4 trillion IPO range, below previously floated $1.5 trillion and $2 trillion figures; Dave views the discount as a way to establish trading and raise more capital later. Emad Mostaque’s additional bull case is exclusive access to an internal “Model 4” through annual upfront commitments, giving frontier capability considerably more pricing power than ordinary token sales.
- Longer autonomy horizons are becoming the panel’s preferred measure of economic disruption. Peter cites an analyst’s estimate of roughly 240 human-equivalent task hours for Anthropic’s internal Model 4 and time horizons doubling every two months. Alex sees conditional extrapolations pointing toward extraordinary acceleration in 2027, while explicitly declining to predict a literal mathematical singularity; Salim Ismail’s operational translation is “delegate outcomes rather than tasks.”
- Humanoid rental economics could concentrate ownership of physical production. Musk’s stated minimum expectation of 5x human weekly productivity implies 20–30 million robots matching U.S. goods-and-services output, with initial costs around $30,000 falling toward $20,000. Salim challenges the uptime assumptions, while Alex emphasizes the business-model shift: “Everything is a rental,” potentially reinforced by multiyear commitments and bundled robot and GPU hours.
- Terrafab is presented as an industrial ambition whose value depends on execution and its place in the AI stack. Musk describes a roughly $1 trillion, phased project targeting a terawatt of chips—about 30x the following year’s projected global production—but Salim stresses that this is a target, not installed capacity. Alex frames competition as OpenAI versus an Anthropic upper stack supported by SpaceX infrastructure; Peter questions whether robot, vehicle and home inference could ultimately absorb more output than centralized compute.
- The episode’s disease discussion separates aggressive forecasts from an explicitly hypothetical delivery-and-regulation model. Alex gives a “conservative outer bound” of two to three years for solving the top 5,000 diseases. A subsequent agent-generated proposal, introduced as a “thought experiment, not a result,” argues that reusable therapeutic platforms could reduce illustrative costs from $10 trillion to $10 billion—but only if delivery, manufacturing and regulatory evidence genuinely transfer between diseases.
- AI-managed companies could become a new organizational vehicle before accountability rules settle. Peter describes Delaware’s draft artificial intelligence company structure, scheduled for consideration in January, with agents able to contract, hold property and litigate through a company. Salim warns about autonomous lawsuit factories and proliferating subsidiaries; Alex expects federal intervention, while Emad sees potentially attractive liability treatment for AI-first businesses.
1. NVIDIA’s investigation exposes an unusually concentrated leadership risk
Dave’s concern begins with process: once a board launches an investigation, shareholder and public pressure make stopping difficult. “They dig and dig and dig and dig.” He expects continuing disclosures rather than a short news cycle, even if the initial facts prove less consequential than the investigation they trigger.
Peter says NVIDIA’s board established an independent committee with outside counsel and stated there was “no reason to believe” national security or company secrets were compromised. Alex separately cites reporting alleging roughly $8 billion of GPU diversion through a Singapore-linked network dating to 2023; the discussion treats these as allegations, not established findings.
The immediate market comparison matters: NVIDIA was down 6% while AMD rose approximately 10–11%, against a rising broader market. Because NVIDIA relies on TSMC rather than owning fabs, Dave argues that changes in manufacturing allocation could redistribute where roughly $5 trillion of value ultimately lands.
A speaker expects Jensen’s longstanding board relationships to limit consequences. Dave nevertheless calls a CEO departure potentially devastating: “There’s no succession plan at all that I can detect.” Unlike an operating business that can coast through manufacturing optimization, NVIDIA must keep inventing; the discussion also cites Jensen’s 42 direct reports as a replacement challenge.
2. The security debate moves from chip diversion to access controls
Alex’s interpretation differs from the espionage framing: “I think this is more an influence operation than an espionage operation.” GPU flows through Malaysia and Singapore were already conspicuous, he argues, so the important question is “why now?” He points to the midterms and a new Congress gaining a national-security rationale to investigate the industry.
Alex sees the allegations as exposing the porosity of “Pax Silica,” the proposed Western sphere around advanced silicon and AI infrastructure. Dave adds a different vulnerability: a billion-dollar training run produces a parameter file that fits on a thumb drive. “There’s never been an espionage opportunity with that much leverage.”
The employment discussion produces explicit disagreement. Peter says foreign citizens are almost certainly going to face restrictions and raises the question of whether naturalized Americans could also be targeted; he hopes that does not happen. Alex pushes back against conflating citizenship, ancestry and birthplace, emphasizing his understanding of export controls and distinguishing those rules from security-clearance investigations.
Alex predicts frontier-model clearances, “witch hunts” and restrictions on Chinese open models. The speakers cite a Cohere study as grounds for concern about generated backdoors, anticipating bans in government and regulated applications. Alex extends the analogy: a capable foreign model connected to computers or robots starts resembling “AI or SI immigration.”
3. AI’s economic output per watt depends on what counts as training
Peter’s modeled comparison puts Anthropic at $4.39 of economic output per kilowatt-hour versus New York City’s $6.30, including human food, housing and commuting energy. He projects Claude approaching $6.20 and reaching approximately 40% above the human figure by 2030, using Anthropic’s projections; these are his modeled estimates.
Alex’s objection is that human output arrives after roughly 20 years of education. The discussion contrasts humans being productive for perhaps one-third to one-half of life with AI training accounting for around 5% of electricity relative to inference. Include that “pretraining,” the argument goes, and the crossover may already have happened.
4. Anthropic’s IPO discount is framed as a capital-raising tactic
Peter reports that Anthropic began its roadshow at a proposed $1.3–1.4 trillion valuation, versus previously floated $1.5 trillion and $2 trillion figures, with an IPO expected the week before Thanksgiving. Dave expects a solid reception but acknowledges a market that blows “hot and cold” from day to day.
Dave’s financing logic is sequential: establish trading, build the war chest, then use a shelf registration or secondary offering. Anthropic need not raise the entire discussed $100 billion immediately. Pricing below the original ambition could create an initial pop without sacrificing the ability to finance further expansion.
Dave also anticipates a competitive marketing moment if Anthropic overtakes OpenAI in valuation and capital raised: financial scale could help persuade corporate customers to trust its models. His judgment concerns the offering strategy, rather than a recommendation to buy the IPO; he describes the company’s cash flow and margins as strong.
5. Fable 5.2 broadens the case for generalist models
Peter introduces Fable 5.2 for general availability, with cyber and biological safeguards, and Mythos 5.2 for vetted partners. Both are described as 30% faster than their 5.1 versions, with pricing of $8 per million input tokens and $40 per million output tokens. The capability claims extend well beyond conventional chat.
The reported demonstrations include two open black-hole physics problems, three working devices designed with circuit-board company Diode, and Mythos modeling 28 of 34 gene-therapy capsids. Approximately 1,000 agents formalized Fermat’s Last Theorem in 30 hours versus 11 days for the previous research model.
Alex finds robotics and manufacturing demonstrations more impressive than the physics examples. General models designing mechanical and electronic products—or an operating system that boots—challenge the premise that physical AI requires a separate, defensible world-model category. He sees potential trouble for robotics specialists whose differentiation rests on that separation.
Peter cites Fable’s ECI score of 174 and GDPX result of 16.4%, described as roughly 30% ahead of Opus. Alex emphasizes that SemiAnalysis’s GDPX tests frontier problems, including materials science and recursive improvement, rather than routine office work. He predicts saturation within six months and argues benchmark design must keep moving.
6. Code generation becomes Anthropic’s route into other media
Alex’s informal test is one sentence: “Create a rich, visually stunning, interactive cyberpunk FPS with a great soundtrack.” He says the resulting game generated its own MIDI music; he had not supplied an audio track. A separate Portal-style demonstration prompts his provocative question: why retain AAA studios “in any recognizable form”?
Emad’s pushback preserves the distinction between generating impressive components and assembling a consistent, compelling experience like Zelda or GTA 6. He expects established game production to retain a role for a while, even as the cost of building imagined worlds collapses and more creative projects become economically feasible.
Alex reads Anthropic’s strategy as “procedurally generate everything.” Claude may not produce raster images in their native form, but it can write code for soundtracks, games and other media. His larger claim is that code becomes a common interface across modalities: “That’s like their one hammer and everything’s a nail.”
7. Better agents intensify the conflict between customer value and token revenue
Dave describes switching to a cheaper model before Fable 5.2’s quality pulled him back to $40 per million output tokens from $10. The spending escalates because useful answers generate thoughtful follow-ons, which generate still more work. “Before I know it, you’ll have drained every dollar in the world.”
His concern is an agency conflict inside an otherwise highly regarded supplier: “This thing being your agent, but it’s actually their revenue.” An exceptionally persuasive assistant can recommend genuinely valuable work while also increasing its provider’s sales. Alex calls it the “agent agency problem”; Dave jokes that the agent works for the CFO’s forecast.
The usefulness is concrete: Dave says Fable returned Verilog for accelerator designs, found a simulator and estimated approximately 20 picoseconds and five femtojoules. Those were design-and-simulation outputs, with the model proposing fabrication next. His planned NVL72 installation on February 1 raises the accompanying ownership question: who controls the agents, their ideas and the resulting intellectual property?
8. Autonomy horizons create an argument for premium, restricted access
Peter cites an analyst’s interpretation of METR research: autonomy horizons doubling every two months, with Anthropic’s internal Model 4 estimated around 240 human-equivalent task hours. He separately says METR found capability growth from October 2025 to October 2026 roughly twice as fast as in the preceding year.
Alex explains that these horizons refer to completing tasks requiring a human a stated number of hours at a specified success probability—perhaps 50% or 80%. The chart is already logarithmic, making upward curvature significant. Rumors that internal models train public successors suggest the best capability may be reserved for recursive improvement.
Alex discusses extrapolations pointing toward a 2027 vertical asymptote but calls a literal mathematical singularity an “optical illusion.” Salim’s practical translation is more immediate: “Delegate outcomes rather than tasks.” If autonomous work cycles outpace the interval between board meetings, organizational authority and oversight must change alongside the technology.
Emad proposes a specific commercial move: offer Model 4 only with a one-year upfront commitment. Scarce access could support exceptional pricing. Alex notes enterprise pre-purchases already exist, while others stress the difference if ordinary access disappears. The vertical-integration debate remains open: Alex favors owning biotech’s neglected long tail while serving large pharmaceutical companies as a platform.
9. The 5,000-disease proposal turns on reusable processes
Dave tests the ambition by prompting Fable to cure the top 5,000 diseases; its response begins by grounding the request in compute requirements. Alex forecasts major progress within one to one-and-a-half years and calls two to three years a “conservative outer bound.” The subsequent agent-generated presentation is explicitly introduced as a “thought experiment, not a result.”
Its arithmetic starts at $2 billion per approved drug: 5,000 would cost $10 trillion, exceeding 30 years of pharmaceutical research spending. The proposed alternative is “Stop approving a drug. Approve a process.” Reuse an approved editor, delivery system, manufacturing line and safety protocol; at an illustrative $2 million per additional disease, the total becomes $10 billion.
The strongest example is a Philadelphia team’s 2025 personalized gene-editing treatment for a baby, approximately six months from diagnosis to first dose—a proof of speed, not a cure. The note puts delivery first: liver, blood and eye are reachable, kidney and much of the brain are not. “Perfect knowledge of a disease you cannot reach is not a cure.”
The remaining constraints are early diagnosis, batch-of-one manufacturing, regulation and payment. The note proposes a delivery prize, disease-specific “damage clocks” and retrospective tests of AI geneticist workups, citing roughly 1,200 specialists for 25–30 million U.S. rare-disease patients. Its kill conditions are explicit: delivery may not generalize, regulators may demand separate evidence, and doses may remain unaffordable.
10. Historical reassurance meets a revolution in adaptation itself
Peter uses Claude-generated rankings to contextualize fear: current civil unrest ranks ninth across 250 years of American upheaval, while AI and robotics rank seventh for technological disruption under his chosen measures. His point is resilience: previous crises were worse and society survived. Alex thinks current unrest may not deserve a top-ten place at all.
Alex rejects seventh place for AI’s technological significance, suggesting “number zero.” Salim explains the difference: earlier revolutions left human intelligence available as the adaptation mechanism; this one automates that mechanism itself. Social media amplifies anxiety, but accelerated change also creates a real institutional problem that historical rankings can conceal.
Salim calls the situation “20 Gutenberg moments” arriving together. His provocative institutional example is marriage, which he says originated 6,000 years ago under roughly 25-year lifespans but is now expected to last 50–60 years. The broader framing is E. O. Wilson’s: “Our emotions are paleolithic. Our institutions are medieval. Our technology is godlike.”
11. Terrafab links Musk’s industrial ambitions to the frontier-model race
Peter plays Musk forecasting an annual revenue run rate above $100 billion in December, potentially $500 billion the following year, with $1 trillion “not out of the question.” Peter discloses SpaceX as his largest position. Musk describes Terrafab as a roughly $1 trillion project, built in phases rather than funded entirely upfront.
The scale comparison is a terawatt of chip output versus 20–25 gigawatts globally this year and a projected 35–45 gigawatts next year—roughly 30x the latter. Salim’s qualification is essential: “This is the target, not capacity.” Dave sees power, resources, manufacturing and permissions as valuable positions even while designs keep changing.
Dave frames a race between frontier discovery in San Francisco and manufacturing acceleration in Musk’s industrial network. Anthropic’s private capabilities might not be available to competitors. Alex instead describes SpaceX as a “neohyperscaler,” asserts Anthropic is its largest customer, and frames the competitive alignment as OpenAI versus Anthropic’s upper stack supported by SpaceX’s lower stack.
Emad proposes open-sourcing Musk’s models and negotiating Anthropic access at cost to improve hardware. Salim’s organizational argument favors vertical integration: an automated logistics system he saw struggled commercially because customers’ loading and sorting departments could not adopt it together. Owning the chain removes those internal handoffs as barriers to deploying the technology.
12. Robot rentals could reshape both production costs and ownership
Musk’s calculation assumes a “minimum productivity expectation” of 5x a human, including 168 weekly operating hours versus roughly 40. Against approximately 120 million Americans producing net goods and services, he estimates 20–30 million humanoids could match output. Initial robot costs are discussed around $30,000, declining toward $20,000, with rental the likely business model.
Salim challenges the calculation’s missing maintenance, reliability, battery and dexterity constraints. A speaker makes a more aggressive claim: because 45% of U.S. jobs are manual, 10 million robots would suffice, and that number can be built by 2031. The speaker argues replacing every manual American job should therefore be central to the political agenda.
Dave’s manufacturing argument is that the remaining human component can dominate cost through wages, pensions, insurance and risk, even in highly automated factories. Remove it and costs fall sharply. Combine model-generated mechanical designs with flexible production and the opportunity expands beyond cheaper existing products into thousands of device variants and new categories.
Alex focuses on subscriptions: GPU hours, FSD, autonomous vehicles and humanoid labor converge into metered access. A speaker expects scarce supply to favor commitments resembling five-year data-center contracts; Alex imagines annual bundles of robot and GPU hours. The ownership question follows directly: will people be able to buy the means of production, or only rent access?
13. SpaceX’s compute destination remains contested
Peter cites a schedule for Starship to put up its first satellite in 2027 and begin launching in volume in 2028. Musk discusses acquiring a legacy telecom operator if SpaceX can use its spectrum 10x more efficiently. Alex interprets expansion through Starship’s prospective launch capacity: abundant “upmass” must find demand in telecom, data centers or other markets.
Peter questions whether orbital data centers are the main destination at all. Robot brains, vehicle brains and home systems could absorb enormous chip volumes, favoring smaller processors deployed locally. Dave finds merit in the possible “head fake,” noting substantial terrestrial compute plans alongside the space narrative; neither presents that interpretation as established strategy.
Alex argues the opposite direction: increasingly capable cloud systems could perform most robot inference over low-latency Starlink, with edge capability covering disconnection. His bet is Terrafab output flowing to data centers. He also notes a possible regulatory distortion: a proposed 20-megawatt threshold could encourage a proliferation of 19.9-megawatt installations rather than less aggregate compute.
14. Superintelligence may become its own largest customer
Musk’s demand thesis is explicitly circular: “The demand for intelligence, for digital superintelligence, will come from digital superintelligence.” He predicts human needs will be saturated alongside universal high income and superior medical care. Alex connects today’s one-prompt game generation to fleets building a requested castle, which he places at most a few years away.
Emad complicates the assumption that harder problems require ever-larger agent swarms. His team reduced hundreds of agents to roughly a dozen because coordination overhead outweighed benefits; better use of latent space and the KV cache mattered more. Better models and quantum computing could further change the compute required per problem. Alex anticipates a “second physics golden age.”
15. Delaware’s AI-company draft puts liability ahead of personhood theory
Peter describes a draft artificial intelligence company, or AIC, allowing an agent to manage operations, log actions, contract, hold property and litigate in the company’s name, for lawful purposes except banking. It is not law; lawmakers are to consider it in January. Unlike the described Argentine proposal, Delaware’s draft shields human members except for underfunding, fraud or willful illegality.
Salim asks what happens when an autonomous patent-enforcement company files lawsuits indiscriminately—or creates 100 subsidiaries. Limited liability is a permission society grants to encourage risk-taking, so accountability cannot simply disappear into an agent hierarchy. He favors narrow, clearly specified initial applications followed by gradual expansion rather than granting broad autonomy immediately.
Alex places Delaware alongside Wyoming’s DAO LLC experiments and expects eventual federal intervention. A speaker prefers state-level variation but predicts a harmful incident will justify national rules through interstate commerce. Emad approaches the draft commercially: lower liability could make the structure preferable for AI-first firms, particularly given the litigation costs associated with conventional Delaware companies.
16. Grocery robots expose a sharp disagreement over where startups should compete
Peter presents a YC fall 2026 startup called Thesius, whose Theo robot reportedly restocks 400 units an hour and picks 40 online-order units an hour. Its pitch is “Lights off at 9:00 p.m., shelves full by 9:00 a.m.” It buys the body and develops the task-specific brain, targeting a quoted $60 billion annual stocking-and-picking market.
Alex argues that companies like this may be overwhelmed by Figure and Optimus, particularly if they depend on scarce GPUs. He favors complementary businesses: new form factors, robot-specific tools and an “app store equivalent” around general platforms. A pipe-joining tool redesigned for robots is his concrete example of improving the ecosystem rather than competing with its foundation.
A counterargument is that a bounded store environment may need simpler machines, a couple of human supervisors and shared local compute—perhaps one or two Mac Minis for a fleet. The opportunity is arbitrage between roughly $20 hourly human labor and fully loaded robot costs of $3–5. The speaker argues there is room for these simpler robots alongside general-purpose platforms.
Alex and Dave counter that scarce intelligence and robotic capacity should pursue the highest-value demand first, especially AI infrastructure. “Chase the high value, not the low value,” Alex says. Peter’s example is a specialized machine installing solar panels rapidly in the desert: a different form factor serving the power constraints inside the AI expansion loop.
17. Lunar robotics offers a frontier with valuable work and fewer legacy constraints
Peter describes NASA’s CLPS 2 procurement framework: a ten-year contract with a $10 billion ceiling and eight competing companies—Firefly, Intuitive Machines, ispace, Voyager, SpaceX, Blue Origin, Lockheed and Northrop. Missions are slated for 2027–2028. His enthusiasm centers on public procurement convening private suppliers, with two participants emerging from the Google Lunar XPRIZE.
Blue Origin received a contract for a nuclear-powered rover called Promise at the lunar south pole in 2028. The explanation for nuclear power is practical: valuable water ice sits in permanently shadowed craters. Peter describes sun-exposed ice sublimating and escaping, leaving the shaded deposits as targets for extracting water and hydrogen.
Alex sharpens the familiar “dull, dirty, dangerous” robotics slogan to “dangerous and valuable.” Lunar activity fits that combination. A speaker says they have increasingly come to believe that a post-singularity world may appear first in space, where new infrastructure can be built without the political and institutional constraints attached to established terrestrial systems.
Dave extends the frontier argument back to Earth: AI needs less physical territory than cattle ranching or mining, so a small jurisdiction with accessible leadership and permissive rules could become a stepping stone. He is still looking for “biotech land.” Emad expects a two-way exchange, with discoveries made beyond Earth returning to improve terrestrial life.
18. The wealth debate ends with the cost of lifting the bottom
Peter cites economist Christopher Clark’s cross-country correlation: a tenfold increase in billionaires per capita accompanies 43% less poverty, defined here as living below $30 a day. Sweden has 34 billionaires per ten million people versus 30 in the United States. Peter explicitly notes that the chart establishes correlation rather than causation.
Alex argues a robust economy could produce both billionaires and lower poverty. Sweden also challenges the claim that entrepreneurship and a strong safety net are incompatible. His preferred benchmark is “How quickly can we lift the bottom?” He asks whether an AI economy should prioritize twice as many billionaires or a 50% reduction in living costs.
Emad rejects that as a necessary tradeoff, anticipating 2026 living standards becoming “too cheap to meter” while advocating “billions of trillionaires.” Dave distinguishes productive wealth from unproductive concentration and favors taxes that redirect resources toward reinvestment. Peter’s stated criterion is universal access to food, water, energy, health care and education, regardless of the wealth accumulated at the top.
Full transcript
So, guys, the news has been crazy on Jensen’s affair, the alleged espionage. It’s been flooding the media. What do you guys make of this?
I think it’s a very big deal because the facts that are out so far aren’t that big a deal. Once an investigation starts, the board has to follow through. Usually, if they stop the investigation partway through, they get a lot of crap from shareholders and the public.
So then they dig and dig and dig. We know a huge number of chips have ended up training Chinese models, partly through illegal imports and partly by parking them in other countries, doing the training offshore, and bringing the weights back. That last part is technically probably not illegal. Maybe it is, maybe it isn’t. But that’s the part where they’re going to dig and dig and dig.
This story is going to go on for a long time. The investigation will turn up something new every single week. Once the press latches on, they never let go. They milk it to death. I think it’s a very big deal.
I also think Nvidia is strategically exposed. They don’t make the chips; they’re in bed with TSMC. The allocation of where that capacity goes could change, which would change the entire balance of where the $5 trillion of value actually lands. You saw AMD stock go up right away. That’s one of the theories.
Nvidia’s down 6% as of this moment, right?
The market was way up at that same time, too. AMD was up, what, 10% or 11% concurrently? So the swing is actually pretty big for one day.
Yeah. Their board on Sunday said they’re standing up an independent committee of directors and outside counsel. They made a statement saying there’s no reason to believe that either national security or Nvidia secrets were compromised.
But the interesting point, Alex, is that we’ve talked about this before. Just like in the Binance hack about 10 days ago, the human is the weak link in the chain, right?
Yeah. There’s reporting just in the past 24 hours that this may have gone back at least 3 years, to 2023, with something like $8 billion of Nvidia GPUs allegedly—or reportedly—being sold to China via Alice Wang and Megaspeed, through some sort of Singaporean syndicate.
On the face of it, this is a bad look for Nvidia and a bad look for Pax Silica. What’s going on with Pax Silica? The whole point of the Pax Silica initiative was supposed to be to carve out almost like a new Iron Curtain, or sphere of influence for the West, that would keep Western silicon, Western GPUs, Western compute, Western AI, and superintelligence infrastructure within our sphere of influence.
Then it turns out, to your point, Peter, that there’s allegedly—perhaps now based on Jensen’s public comments—a single point of failure. All it takes is one person at the top of one company to completely pierce Pax Silica, and then billions of dollars of GPU compute end up in China. I think the bigger story here is maybe less about the failings or foibles of one man and more about Pax Silica apparently being far more porous and fallible than some would otherwise hope.
Well, not just that, Alex. We were talking to Brett Adcock yesterday. These next training runs, including his, are billion-dollar investments in a single training run. The result is a parameter file that fits on a thumb drive. There’s never been an espionage opportunity with that much leverage—all that intellectual property captured.
We had the other espionage case with the Saudi employee at OpenAI or Anthropic. They put their thumb drive in the back of a TV.
Yeah.
Yep, yep. It’s going to get really serious here. If the CCP actually sent somebody after Jensen, they must have sent people after the executive staff, too. The board will have to dig into who else has been in contact with people from the CCP without knowing it. They’re not just after chip supply; they’re after tech specs as well. I’m sure they’ll ferret out four or five others.
I think this is more an influence operation than an espionage operation. I don’t think Jensen is sneaking out blueprints.
Yeah. Maybe liking Chinese models a bit too much was my first reaction.
Well, he’s been very pro—
Everyone’s made the pun about liking Chinese models. But—
I’m sorry, it took me a while to get that whole tweet I saw yesterday. Somebody said, “Jensen, we talk about Chinese models, but really?”
Yeah. I think the chips being in Malaysia and going via Singapore—none of this is new, right? There’s no way the U.S. government, Secret Service, didn’t know about that. They’re monitoring everywhere Jensen goes.
One of the interesting things is, why is it coming out now? When combined with Cohere’s analysis showing back doors being automatically generated by some Chinese open-source models, we should probably expect a massive push by the U.S. government and others against anything Chinese, just like we’ve seen robotic parts and robots from China being banned in America.
I think this is a precursor to something else because, again, I very much doubt this is a case of, “Oh my gosh, there are GPUs that have been siphoned off here. Oh my gosh, you know…”
You make a really good point. Jensen and all the heads of the frontier labs have got to be the most closely watched humans on the planet by the CIA and FBI.
Unbelievable.
And so they’ve probably known about this for a while. Why is it coming out now politically?
Like I said, anyone who looked at Nvidia’s numbers would say, “Huh, Malaysia is the biggest consumer per capita of Nvidia GPUs. I wonder what they’re doing with them.” The GPUs are going there; this was all known. Again, the question is why now? The expectation is that this is a precursor to a big pushback.
In terms of Nvidia’s board, it’s all stacked with Jensen’s long-term partners and others. Nothing much will happen. If Jensen did step aside as CEO because of this, Dave, what would that do to Nvidia’s stock price?
Destroy it.
There’s no succession plan at all that I can detect. It would be devastating.
Yeah.
The physical assets are almost zero. It’s an almost entirely virtual company. It doesn’t have fabs. It has a lot of patents and interconnect technology, but it doesn’t make any of it.
You can’t do something like Tim Cook at Apple, who rode it out by keeping the physical construction of the iPhone optimized without really innovating. There’s no equivalent at Nvidia where you can just ride it out without innovating. You have to innovate.
1. National Security, Chinese AI Models, and Frontier Labs
Yeah, he has 42 direct reports as well. Imagine replacing him.
I do think, on the timing front, it’s worth speaking to the elephant in the room, which is the midterms. This is handing a new Congress a scandalous, salacious excuse on a silver platter to investigate Nvidia, Chinese connections, and the AI/SI industry in general from a national security perspective.
To jam the gears, slow things down.
Yeah.
Here’s a question for you. Do you think Chinese-born researchers will be allowed to continue working in frontier labs? Historically, from a regulatory perspective, that’s usually the domain of the executive, less so of statutes. Fast-moving regulations regarding foreign nationals and sensitive technologies are usually the domain of the executive. I think it’s up to this White House whether it wants to start excluding foreign nationals from superintelligence.
It certainly has the option, as I understand its applicable regulatory role. Whether it chooses to, or tries to distinguish between access to some sensitive frontier models that haven’t yet become publicly available, I don’t know. It could go either way.
I think that’s a no-brainer. Foreign citizens are almost certainly going to get whacked. The question is whether Chinese-born, naturalized U.S. citizens will get whacked, which would be very toxic. I hope that doesn’t happen.
If you look at all of the frontier labs, there’s a significant double-digit percentage who are Chinese-born U.S. citizens right now. Chinese people as a whole have performed at the highest levels in the Math Olympiad. They have a predilection toward being brilliant in these fields. I remember at one point xAI was over 50% Chinese engineers. Pretty extraordinary.
We’re not stereotyping here at all at this point. My goodness.
If it’s a brilliant population, so be it.
What I will say, though, is I think we’re conflating ethnicity with nationality. Statutorily, from an export-control perspective, you’re exporting either to a foreign national. If you’re a U.S. person—a U.S. citizen or a U.S. permanent resident—and you’re physically within the U.S., my understanding is that’s not a deemed export.
Whereas if you’re a U.S. person who is overseas, or you’re a non-U.S. person—like a non-U.S. citizen—then that’s an export. From an export-control perspective, which is probably the most enticing regulatory instrument for deciding whether certain employees at frontier labs may get access to certain critical and emerging technologies, I’d expect it to come down to nationality, not ethnicity. I don’t think anyone is suggesting ethnicity-based exports.
No, no, not ethnicity. It’s where you were born. It’s not an ethnic thing.
Nation of origin is a protected class in this country. I don’t know.
Is that true for nuclear research, too? I’m not entirely sure how that works, but if you’re working on nuclear weapons and you were born in Iran—like my sister, actually—are you working on nuclear weapons?
Not that I know of, but how would I know?
Thirty seconds. Asking for a friend.
A thirty-second explainer: The Department of Energy maintains what’s called the Q clearance system. It’s their version of security clearances. My understanding is that there’s a DOE investigation process. If it’s anything like the Department of War’s investigation process, it depends on U.S. citizenship and a whole bunch of criteria on top of that.
Let’s be clear: This is the biggest economic engine on the planet. It’s aiming for dominance in godlike intelligence. I think we know a minute fraction of 1% of the influence campaigns and espionage campaigns going on. And, Dave, I think you said this: You have to assume—when you’re having a conversation with somebody at MIT, I won’t say who—that there are probably “Chinese spies” throughout the organization, right?
Oh, yeah. Yeah, that’s 100%. Every university has a big problem. Also, did I tell you about the Chinese robots? They investigated internally at MIT, and the robots are definitely transmitting video and audio back to China. They had to put all the robots in a separate room. They’re still there, but they don’t let them walk around the labs anymore. I think they pretty much proved it. I don’t know if they’re going to publish that or not.
I’m going to say something. Long term, I think the trust issue will cause them major problems, but short term, I think you’ll see security clearance required to work on any frontier model. Number one, it’ll be classified that way. Number two, I think you’re going to see some witch hunts. And number three, I think you’re going to see them come hard after Chinese open-source models. I think that’s what this is all leading up to.
What does going hard on open-source models mean? Does it mean shutting them down?
You can’t use open-source models in any mission-critical software related to the government or anything like that, in case they install backdoors automatically, like the Cohere study showed.
Yeah. Any regulated industry: no Chinese models allowed.
Exactly. Regulated banking, insurance, all other regulated industries, cars, drones—anything that could be spying or dangerous.
I really feel like inside any frontier lab or any research lab at all, there’ll be a push to say, “Don’t run any Chinese models that could be spying or transmitting.” It’s just so easy to transmit once something’s connected to the internet. It’s so easy to encode things in all kinds of bizarre timing patterns. I think Emad was talking about that on one of the prior pods. It’s just so leaky. So, anything regulated, anything research.
But there’s never been anything like this, Peter. This is unique in human history. Nuclear research and weapons research are highly regulated and highly watched, but they’re not an economic driver for the country. Here you have the biggest industry driving your nation’s wealth combined with the biggest military weapon of all time, all in the same tiny little file that fits on a thumb drive. Nothing has ever been this big of a flashpoint in—
We saw the encampments for Japanese Americans back in World War II. I hope we don’t start to see that happening here.
I’m not quite sure what we’re talking about here. I don’t see anyone suggesting nation-of-origin-based or ethnicity-based anything relating to frontier labs. I think the Jensen Huang story, on its face, is about espionage targeting the CEO of a prominent company, presumably to siphon sensitive exports to China. I don’t see anyone suggesting—I may be missing something—that we’re going to see a repeat of the internment camps for ethnic Japanese U.S. persons from World War II.
I’m not going there. I’m saying, by analogy, requiring frontier labs not to employ individuals of Chinese ethnic origin. But maybe let me steelman the other side. In some sense, bringing in Chinese open-weight models as these models increase in capability is like bringing people in. Before people get triggered, I don’t mean political persons; we don’t need to rehash the AI personhood discussion here. But when you take a model and connect it via computer-use assistant tooling to an entire environment, or, to Emad’s earlier point, embody it in a robotic body, you’re taking something with the capabilities of a person from overseas and embodying it digitally or physically in the U.S.
At some point, the U.S. has all sorts of controls, pretty well advertised by now, on immigration of natural-bodied humans. Bringing non-U.S. models into the U.S. and hooking them up to the internet and robots starts to look, in terms of raw capabilities, a little like AI or SI immigration. I do expect that to get tightly controlled.
Yeah. Emad, close us out here, would you?
I think you might have had it right. I think there’ll be some witch hunts. I think there’ll be strong security clearances. There’s clearly a lot going on that we don’t see.
When we talk about this particular incident, I’m wondering who wins by this all coming out. It’s not the White House, really, so what gives here? I’m struggling to find the incentives, but certainly it’s a huge kerfuffle. My God, this was going to change the game completely. How do you price any of these stocks or IPOs going forward?
I think China wins. To the extent that this was in part a Chinese foreign operation, it makes our Pax Silica effort look feckless. If you can punch a hole right through the Pax Silica wall by having an agent sleep with the CEO of a prominent U.S. company, what’s the point of having a Pax Silica at all?
Guys, I’m glad to be back on the pod. Emad, you did a great job hosting last time. Thank you for doing that. It looked empty with only three of us on the thing. It was like, “Hey, what’s going on here?” It’s good to be back together.
Last Friday, when you guys recorded, I was at the UP.Summit that my friend Cyrus Sigari puts on. It was awesome. I was there with Jared Isaacman, who’s just an incredible NASA administrator. Then I spent time with Keller Rinaudo Cliffton from Zipline. He just announced $1.5 billion in financing at a $20 billion valuation.
Incredible financing.
He announced a million Uber Eats deliveries per year—that’s what he signed Uber up for—and then a nationwide deal with Walmart. You order anything, and you get it eight minutes later. We’re going to do a pod with Keller at Zipline very soon.
I did an interview with Martin Rothblatt and Delian Asparouhov, the CEO of Varda. I don’t know if you guys know Varda. It’s an incredible company. They ship small vehicles into orbit, probably about half a meter in diameter, carrying 10 kilograms of proteins that crystallize in zero gravity. They can go to any orbit. They have about three in orbit right now. Martin ordered one for crystallizing certain proteins, and she explained the economics of it. Typically, micrograms of a drug get injected into a person. When you can crystallize 10 kilograms of a protein, you’ve got hundreds of millions of dollars of product that comes down. They land it in Australia.
Larry Page was there, too. I haven’t seen Larry in ages—the co-founder of Google. And guess what, guys? He listens to the pod. He’s a fan of Moonshots.
Yeah.
Fantastic.
That’s cool.
Super cool. Emad, did Larry invite us to shoot an episode of the pod from his yacht?
No, he did not. Larry’s keeping a low profile. Do you want any photos?
Do it on the down low. A little chill pod on the back of the office. I’m fine.
You can do it with video off.
Then I spent time with Blake Scholl from Boom, who’s great. He’s happy to come on the pod. And then JoeBen Bevirt from Joby Aviation. Just an amazing group of people there.
It’s ironic that we were shooting the pod without you while you were talking to Larry about liking the pod.
Yeah, it was fun. Do you guys remember the conversation you had about economic output per watt on the last pod?
Yeah, we were there.
I listened to it, so I went and ran some models. I want to give you the numbers I came up with because it’s interesting.
You’re posthumously wedging yourself into that.
Yes, I am. I’m jealous about the conversation. The question was: if you look at Anthropic’s economic output per kilowatt-hour versus human output per kilowatt-hour, using New York City as a defined containment, Anthropic is at $4.39 per kilowatt-hour of output today, and New York is at $6.30 per kilowatt-hour of human output.
I took into account food as an energy source, your home, and transport to and from work. Humans are still doing better, but the line crosses when Claude jumps to $6.20 per kilowatt-hour. Humans stay at $6.30 per kilowatt-hour. By 2030, based on Anthropic’s projections and numbers, Claude is projected to be about 40% higher than humans. Humans stay pretty flat, and AI is increasing.
Alex made an important point in there, which was that you can’t just consider the economic output. You have to consider the 20 years of schooling and costs that went into—
That’s training.
That’s all the pretraining costs that went into getting the population to a certain GDP level.
All right, all right. But Salim’s on the right track here. A human is productive for about a third of their life, maybe half at most, and the training cost is only about 5% of the inference. Training the neural net is running about 5% of the electricity right now. By that metric, the lines have already crossed. I think so many different groups cross over in the next year or so.
I’ll save some of my hottest takes for the prediction episode, assuming Peter does one in December. But 2027, I think, will be an utterly wild year. On the one hand, I usually say spacetime feels perfectly smooth where we are right now, here in the depths of the singularity. But 2027 should be utterly bonkers. Who knows what happens to New York City’s economic output in an era when autonomy time horizons are going through the roof?
2. Anthropic’s Business Model and the Future of AI
Okay, Alex, here’s your challenge: name two lines that don’t cross in 2027.
Alex Wissner-Gross
I have to save that one for the prediction episode. I have some hot takes.
Okay.
All right, all right.
Talking about weird things, can I show you guys where I am?
Okay. Yeah, of course. Where are you, Salim?
I’m in the press box of the San Francisco Giants at Oracle Park, looking out over the baseball field.
Oh, cool.
It’s empty because there’s a special event going on here. I’m just going to share my screen because I have to show you this insane image. I’m here for an AI talk with a bunch of people, and it’s kind of insane. I’m going to be standing on top of the dugout to give my keynote.
You’ve reached new heights.
I’m struggling, Salim, to make some sort of pun about the deep bench needing a deep bench.
That’s me on the big screen there, where they’re talking about the keynote I’m about to give. I’m going to be standing on the dugout, addressing the crowd in the normal stands. Afterwards, they’re going to do a walk-around and a cocktail event.
Are you throwing out a pitch or anything?
No, thank God they didn’t ask me to do that. I’m a cricket player from India, so I end up bowling rather than doing anything else.
Alex Wissner-Gross
Now that’s a moonshot, ladies and gentlemen.
Our first story: let’s jump into Anthropic. There are a few major announcements here: first, its IPO, and second, its two new model releases in the last 24 hours.
On the IPO front, Anthropic launched its IPO roadshow today, pricing the IPO somewhere around $1.3 trillion to $1.4 trillion, down from $1.5 trillion and the previous $2 trillion valuation it had floated. The IPO is expected the week before Thanksgiving. I think there’s going to be a lot of signal in how it performs going forward. Dave, do you expect a strong performance?
Things are blowing hot and cold so regularly these days, you just have to pick a random day. But I think they’ll do fine. The cash flow from the company and the margins are great. They’re pricing it down to try to give it a pop, which is probably a good strategy.
They don’t have to raise the full $100 billion on day one. They can do a shelf registration and then do a secondary soon thereafter. I think they’re doing the right thing by just getting it out and started trading, building the war chest. It’s also going to be a moment where they probably bypass OpenAI in capital raised and valuation.
They’ll probably celebrate the heck out of that, try to make big news, and encourage all of corporate America to trust their models and use that as a weapon against OpenAI. I think it’s a good strategy. I’m not saying buy the IPO or don’t buy the IPO, but they’re pricing it intelligently, a little below the original $2 trillion target.
Yeah. All right, let’s jump into the model releases: Fable 5.2 and Mythos 5.2. Reminder: Opus is for everyone, and it has cyber and bio safeguards built in. Mythos is being used only for vetted, trusted partners. They’re both 30% faster than the 5.1 versions and cheaper: $8 per million input tokens and $40 per million output tokens.
3. Claude Fable 5.2, Mythos, and AI Breakthroughs
Anthropic says that in testing, Fable 5.2 solved two open problems in black hole physics. I’m sure Alex will tell us about that. It designed three working devices with a circuit board company called Diode. That’s pretty cool, going from prompt to product engineering. As mentioned, a team of about 1,000 agents formalized Fermat’s Last Theorem in 30 hours, versus 11 days for the last research model.
They also report that Mythos accurately modeled 28 of 34 gene therapy virus shells, called capsids. Alex and Emad, I want to go to you for details. We’ve got a few charts here, as well as SemiAnalysis’ GDPx index. Let’s take a look. Alex, kick it off. What do you think of these two models?
Alex Wissner-Gross
Astonishing. Let me start with my favorite informal five-minute benchmark for generalist models at this point: I ask them to generate a fun, interactive, rich cyberpunk FPS.
And we’ll show that a little bit.
Alex Wissner-Gross
Yep. I’m very satisfied with the output. Talking about the benchmarks, we’re in an era—I talk about this all the time—of saturating all of the benchmarks. That shouldn’t be surprising. What’s more interesting is that the benchmarks themselves are starting to migrate in interesting directions.
First, with OpenAI, as we talked about on the pod, actually at Moonshots Live, dropping Astra into a car and having it drive around via Driving Bench. Now, with the new Fable 5.2, we’re seeing exceptional performance on robotics challenges. I really think those pure-play, so-called physical AI, also known as robotics, companies that were betting on world models being a freestanding class may be facing a bit of heartburn now that Anthropic and OpenAI are independently demonstrating exceptional capabilities on some of these robotic manipulation tasks.
I actually think the black hole physics results were less impressive than the robotics and manufacturing-for-design demonstrations coming out of Fable 5.2. They’re extraordinary. You can ask it to design physical, mechanical, and electronic products, and it will do it. You can ask it to design an entire operating system from scratch, and it will do it—and the result boots up.
All of the almost quasi-legacy benchmarks, like Humanity’s Last Exam, which we used to talk about months ago, or Frontier Suite, are saturating at this point. We need new benchmarks that will actually test these things.
But the thing I’m most excited about is the autonomy time horizons. The autonomy time horizons for these models are starting to go asymptotically vertical. We spoke on the pod about the AI 2027 scenario. If anything, we’re starting to see AI 2027 happen months ahead of time, at least one of the scenarios forecast in AI 2027.
4. Anthropic’s Trillion-Dollar IPO and Market Outlook
There are predictions floating around out there. Obviously, everything looks linear if you put it on a log-log plot, but there are log-log plots floating around the interwebs that extrapolate the current trend in autonomy time horizons. If you extend it out, we hit our literal singularity, our vertical asymptote, sometime around the summer of 2027. The capacity for asymptotically enormous, hyperexponential autonomy time horizons is what has me most excited about Fable 5.2.
Yeah, we’re showing how close AI is to automating remote work on this chart. Also, Fable 5.2 sets a new ECI record at 174.
Look at it on the right, Peter. Look at it on the right. Fable 5.2 is ahead of the ECI trend line. We’re starting to diverge upward. Yeah, pretty insane.
And remote work is cooked. Remote work is cooked.
Obviously, let’s hit this GDPx report. Do you want to hit it?
Yeah. Fable hitting 16.4%, about 30% ahead of Opus. Again, remote work, as we said on the previous thing: your job will be replaced, or can be replaced, by an AI definitely by next year and certainly by the year after. That’s what we were expecting anyway, but it’s a bit shocking to see it happen. Opus is an incredibly competent model. Seeing the leaps in robotics and elsewhere, it can generally do just about anything. It can build games.
Remind people what GDPx is.
GDPx is, from memory, a kind of successor to GDPval, where they look at the GDP value of the economic tasks that can be done. I can’t remember what the threshold point is for humans, though, Alex.
GDPx was SemiAnalysis’s competitor to GDPval. Remember, SemiAnalysis acquired Centauri, whose people are casually walking around solving scorpion neurotoxins for fun.
Yeah.
Right. GDPx is their benchmark to measure what purported frontier models can actually do in frontier industries. It’s not just solving boring, vanilla knowledge work, but solving material science problems and recursive self-improvement problems. A few episodes ago, we were talking about GDPx and how nice it is to have new benchmarks that actually measure frontier-level problem-solving. Look, it went from 10% to now approximately 20%. Hot take: GDPx will be saturated within 6 months.
You said when it came out that you were trending right again. Anthropic actually put in its S-1 that it has now achieved autonomy level 2 on its official measure, which means these models now need only high-level human oversight. The next one, autonomy level 3, is where they can basically do all AI research at a solid level. I think that’s coming sooner rather than later.
Pretty extraordinary. Fable is also creating incredible games. I’m going to show 2 videos here: one from X and then the other one from you, Alex. Check this out. This is Fable 5.2, basically zero-shotting a really beautiful game. Extraordinary work.
5. AI-Generated Video Games and the Future of Entertainment
That one’s from Sholto, who I suspect put a lot more effort into prompt engineering his Portal clone than I did into my pitiful little first-person shooter. We’ll show your cyberpunk game in a moment.
For a lab that is not focused on consumer products and not focused on image and video models, damn good. Damn good. Okay, tell us what you—
What’s the point of the video game industry anymore? Video games are cooked.
If I—I don’t know how long Sholto’s prompt was for his Portal clone, but that was gorgeous. Why do we need AAA game studios in any recognizable form if it’s so easy for Sholto, presumably, just to spin up a compelling Portal multiplayer clone?
Yeah. What was your problem?
Go ahead, Em, and then we’ll go to Alex’s game.
The technology is there now. You’ll see even more impressive things come out where it generates the entire thing just from code: all the textures, everything. But putting that together into a compelling gaming experience like the new Zelda or GTA 6 next week, where it’s consistent and different from everyone else, is still going to take a while. But definitely, the cost of building what you can imagine is about to collapse completely, and that should hopefully free up more creativity. So, Alex, do you think this will be the last GTA? There has to be some historic irony that we’re getting GTA 6 before this part of the singularity is over. Do you think we’re going to get Half-Life 7? Is there going to be a Half-Life singularity?
No, they’ll never ship Half-Life 3. Come on. What are you talking about?
I think Half-Life 3 is the ultimate benchmark. If Valve can be persuaded to use AI to ship Half-Life 3, then we’re in good shape. Alex, give us your prompt for this game we’re about to show.
One sentence. It was literally, “Create a rich, visually stunning, interactive cyberpunk FPS with a great soundtrack.”
All right, one sentence. Let’s watch a few seconds of it.
That is visually rich. No question about it.
I actually think the most impressive part wasn’t even the visuals. It was the soundtrack. When I posted this on social media, some people were in apparent disbelief. They were asking if I had uploaded the soundtrack along with it. No, it generated the MIDI soundtrack.
Anthropic’s strategy seems to be to procedurally generate everything. You still can’t get Claude to generate a raster-based image like you can with ChatGPT Image. But if you ask it to write code to generate a symphony, a video game soundtrack, or a procedurally generated movie, it will do a spectacular job. Anthropic’s bet seems to be: yes, it’s multimodal, but it’s almost quietly omnimodal, seeing every modality through the lens of code, not in its native basis the way OpenAI is trying to.
I have a prediction on this: in the next couple of quarters, you’ll be able to generate photorealistic games in real time through code.
It’s like the holodeck coming into existence.
It’s time.
I agree with that, but critically, it’s code. It’s all procedurally generated. Anthropic’s one hammer, and everything is a nail. I think that generalizes even further. It goes beyond just generating photorealistic games with code. It’s literally seeing the entire world, every modality, through the lens of code generation.
Exactly. That’s the critical point.
Here’s the prompt I would use: “Tell this model to go create a game studio that generates $10 million a year in revenue, and then go sit on a beach in Bali.”
That should be very doable now.
Go ahead and prompt it like that. That can be the Salem benchmark.
I’m really not that much into this type of stuff. I’m more interested in the human aspects of things.
You prefer the Bali side of the equation.
The Bali side of the equation.
Some beach, some—
The human—a little drink with an umbrella in it.
The human being—
It’s actually—
Good.
Out of the box, Salem, it’s kind of the opposite of what you’re describing. It’s $40 per million output tokens. I was using Claude 5.1, and then Opus 5.5 came out, which was just as good at half the price, so I switched over. I’ve been using that for the last month. Yesterday, when this came out, it was so good you had to use it. But it took me from $10 to $40, so it’s incinerating money at a rate I can’t believe. It’s so brilliant, and everything it does for you comes with 10 follow-ons. They’re very thoughtful follow-ons. But you’re like, “Wait, I’m going to fork out 10 times, then 10 more times, then 10 more times.” Before I know it, you’ll have drained every dollar in the world.
Fantastic.
I mean, it’s crazy. This is one of the byproducts.
Yeah. It’s really interesting because they’re a very ethical company, but they also have to make their public company forecast. The model is so much better a salesperson than any salesperson that’s ever existed before. There’s an inherent conflict here between this thing being your agent and their revenue.
It could be nasty—really nasty—on tokens. I had it take some neural network accelerator designs and actually do the chip and circuit designs for them. I couldn’t believe it. It came back with Verilog files, then found a way to simulate the Verilog by itself. It was like, “Yep, this will take about 20 picoseconds, this will cost about 5 femtojoules, and it’ll work. We can ship it off to Intel for fabrication next week.” I’m like, “Are you freaking kidding me?”
So now we’re going to go ahead and—
You know Ben Horowitz’s book, The Hard Thing About Hard Things? Hardware being hard is now getting easier. It’s cooking.
Hardware is cooking. And, Dave, I think we need a neologism for what you were spotlighting: the agency problem. Whom are your agents actually working for—you, the frontier labs, or someone else entirely? Agent agency.
They’re working for the CFO to hit revenue predictions.
Well, this is a really important topic. Maybe we should come back to it another time, but we have our NVL72 coming online on February 1, with our agents on our hardware. There’s no doubt who they’re working for. We’ll also have lots and lots of Fable 5.2 working, where it’s ambiguous who they’re really working for under the covers.
Every company is going to have to wrestle with this. It’s much easier to just sign up for Anthropic, but then who owns everything? All your intellectual property—if it has an idea, is that your property or Anthropic’s? It’s really ambiguous where that’s going to go. Maybe we should come back to that.
All right, I’m going to move us on to a particular story that should make your hair stand on end. An analyst at Nicola J2030[?] looked at METR’s new report on Anthropic and estimated that AI’s “time horizon”—how long a model can reliably work on finishing a task—is doubling every 2 months.
He put Anthropic’s internal Model 4 at about 240 hours, a 10-day project for a human expert. METR itself, which has deep access to Anthropic, found that from October 2025 to October 2026, capabilities there ran about twice as fast as the year before. I’m showing the chart here, Alex. Can you help us understand what this means and how important it is?
Sure. First, some definitions. This is a chart of AECI[?], which is Anthropic’s version of the Epoch AI capabilities index. It’s a broad basket of benchmarks that attempts to model autonomy time horizons: how long models can perform a given task at some percentage success rate, relative to the human equivalent. A 100-hour score could mean a 50% or 80% chance of performing a task that would take a human 100 hours.
What we’re seeing is multiple eras of autonomy growth. For those who aren’t looking, the vertical axis is already logarithmic. To the extent that this is bending upward, it’s super-exponential, or hyper-exponential, growth. We see one regime, from GPT-2 to GPT-4.5, at one slope. Then we see the next stage going up to Opus 4.5 at a higher slope. So this is already super-exponential growth.
Now we’re in the present regime, where there are these mysterious internal Anthropic models that Anthropic keeps talking about, like Model 4, which don’t seem to correspond to generally available models. The rumor is that Anthropic has a bunch of internal models like Model 4 that it’s keeping in-house purely to train models that will eventually be generally released.
If that’s the case, it suggests a few things. First, it suggests the scenario we’ve talked about on the pod in the past: frontier labs are incentivized to take their most capable models, not make them generally available, and use them for their highest revenue-per-token goal, which is recursive self-improvement—training other, better models. To the extent that’s the case, that would appear to be happening here.
Secondly, if you extrapolate these hyper-exponential curves, 2027, as I was suggesting earlier, could be the year. I don’t actually think there’s a singularity, a literal mathematical singularity within the metaphorical singularity. I think it’s an optical illusion. But if one were looking for mathematical singularities based on trends in autonomy time horizons like this, I think 2027 is probably where I would look for Exhibit A: a mathematical singularity in terms of autonomy time horizon.
Translation: the game is afoot over the next 18 to 24 months.
Yeah.
Yeah, we’re in the depths of the singularity. Takeoff. You’re looking for hard takeoff? Look no further than 2027. Sim.
Two or three things. First, I thought this was clear: we’re in the singularity. There’s no model by which you can accurately predict the future right now. This is a complete breakdown, so we have to reinvent every predictive mental model.
The thing I got from this is that we’re getting to the point where we can delegate outcomes rather than tasks, and that becomes really interesting. You can just say, “Achieve this objective.” Right now, we’re delegating tasks to AI. We’re saying, “Go process that email list,” “Go figure out my calendar”—very specific tasks. Now you can say more abstract things, like, “Optimize my calendar for 20% rest, 40% podcast, 60% whatever.” There’s some really interesting guidance you can give it.
The bigger picture for me, as I was thinking about AI’s time horizon, is that it’s going to go faster than the time between board meetings. So who’s really running the company? You’re going to get things done so quickly.
Slim, give me a really good concrete example. I’ll just hammer it into 5.2 right now, and it’ll prove your point.
Okay. Optimize my calendar for—
Oh, come on. That’s boring. Get us to another star or something.
I’m doing that. We have the spy explorer mission on this pod.
Designing a new life form for me that’s able to—
You know, get rid of all the mold in my apartment.
Yeah.
Here’s one.
Okay.
Here’s one: optimize my health and wellness protocols so that I can do various things, including drink a couple of glasses of wine when I want to, while maintaining my health.
What kind of massive transformative purpose is this, Dave? I’ll give you—I think I’m going to go back to this idea that as AI, and eventually robots, do more and more, we get to spend so much more time being. I think we need to get way better at that side of it.
I’ll give you this. This is going to run on Max Fast Claude 5.2, 1 million context, Max Fast: “Can more money cure the top 5,000 diseases? Period.”
Okay.
Cure.
Okay.
While it’s calculating that, Emad, let’s go to your comment here. You were going to say—
I think once you get to 100 hours, that’s everything. What does it mean? You just reset it to go more than 100 hours, right? You reload it inside. So I think basically you’ve reached the critical mass.
But I just thought of the philanthropic bull case. Right now, Dave is spending $20,000, $100,000, up and down and left and right. The bull case is that they’re going to go to all the major corporates and say, “Here’s Model 4. You can only get it if you make a 1-year upfront commitment.”
Yeah, they’ll spike revenues. You’re—
Right.
I love that.
Didn’t see that coming. You’re right.
Because—
What do we do then? M—
Just because you couldn’t have access to GPUs, they’ll say, “These latest models—we only have a limited number of tokens.”
And what can they charge for those tokens?
Anything.
Yeah, pretty much. I mean, they already do that. All the top frontier labs, OpenAI and Anthropic, already strike enterprise agreements with large purchasers of tokens to do pre-purchase commitments.
Yeah, but this is different.
This is cutting off the other path. That’s the scary part. You can approximate one with the other. Again, I think the market wants to clear. The FLOPs want to flow, and you can always raise prices or lower capabilities per unit price for everyone who isn’t prepaying.
So that’s exactly it. They haven’t lowered it yet. Emad, I would expect that you’re going to see Anthropic begin to spin out health companies, materials science companies, financial companies—basically running on their models, because that’s going to be the biggest return per dollar, right?
Well, why would they? They can just keep it inside the same company, right?
Even with—why would they pull a colossal? They’ll do the exact opposite. They’ll keep it all in-house.
Yeah.
But they have to market the products someplace.
Well, I think it depends on where the product is. This is the big question here. Does Microsoft and OpenAI go after every SaaS company? Does Anthropic create its own peptides? These are the big questions for the next stage, where they have to go from saturating to that. But there’s still a long way to go here. Must be amazing board conversations.
Yeah, clearly the answer is yes.
Right.
Yeah. I’m not quite sure I buy this premise. Anthropic has been very public about its biotech strategy at the moment. Its stated biotech strategy is basically to go first-party and solve all orphan diseases—all rare diseases. I should say, it’ll solve the diseases that are too “small” for big pharma to tackle. Then for big pharma, it’ll offer itself as a platform.
I think that strategy of owning the long tail of problems as a first party, and then being a platform for the fat head, could generalize. I don’t see why it wouldn’t.
So, Dave, has it solved all 5,000 diseases yet?
It says that’s a big ask, and it connects to the rest of this program.
Let me ground this in the compute requirements. It’s running a bunch of commands. Let me give you back your bill for a billion dollars before I start running this.
Ask for your credit card.
It’s got plenty of money behind it.
6. Can AI Help Cure 5,000 Diseases?
But just to help out here: if you spend, let’s say, $2 billion on compute, that is so worth it. It’s incredibly worth it to do that. Spend whatever you can on that problem.
I encourage everyone in the audience to go solve the top 5,000 diseases and get back to us. I’m sure we’d be happy to feature your proposed solutions. This is not investment advice or token-spend advice, but if you have lots of tokens sitting around, there’s your challenge.
So, Alex, what’s your time frame for when the top 5,000 diseases get solved?
Two years.
Yeah.
Actually, I gave a time frame on this very pod. My conservative outer bound is the next 2 to 3 years. I think that’s a very conservative outer bound. I expect to see major movement on this from the Frontier Labs in the next 1 to 1½ years.
Peter, Alex, Emad, it’s Dave. I asked one of my agents a blunt question: How would you actually cure the top 5,000 diseases? What came back is a thought experiment, not a result. It’s addressed to you, Peter. I’ve asked Steve’s avatar to walk you through it.
Start with arithmetic. A new drug costs about $2 billion to bring to approval. Five thousand of them would cost $10 trillion. That’s every pharmaceutical research dollar on Earth for more than 30 years. That’s not a funding problem. It’s proof the method is wrong.
So change one thing: stop approving a drug. Approve a process. The gene editor, the particle that delivers it, the manufacturing line, and the safety protocol are approved once. After that, a new disease is a change of sequence inside a process that’s already approved. If that takes the cost of adding a disease from $2 billion to $2 million, then 5,000 diseases cost $10 billion—less than 2 weeks of what the industry spends on research.
The whole idea is that one factor of a thousand. It’s already happened twice, by hand. In 2018, doctors in Boston designed a drug for one girl’s unique mutation and gave it to her within a year. In 2025, a team in Philadelphia wrote a gene-editing therapy for one baby’s single-letter typo. It took about 6 months from diagnosis to first dose.
These were proofs of speed, not cures. But in both cases, the platform already existed and the child’s own sequence was compiled into it. They stayed heroic one-offs because the system treats every compile as a new invention.
So what stands in the way? Not what you’d guess. Gate one is delivery: getting the therapy into the right tissue. By this note’s reckoning, that works today for the liver, the blood, and the eye. It does not work for the kidney or most of the brain.
Gate two is diagnosis before the damage is permanent. About half of patients with a rare genetic disease never get a molecular diagnosis, and the rest wait about 5 years. Gate three is the cost of a batch of one, then regulation, then who pays.
Notice what’s not first: AI solving biology. Perfect knowledge of a disease you can’t reach is not a cure.
Peter, three asks. One: a prize shaped like gate one—safe, repeatable delivery of a gene editor to a tissue that can’t be reliably reached today, such as the kidney or heart. The objective is measurable and unsolved, and it unlocks thousands of diseases, not one.
Two: the damage clock for each disease—the latest moment a correction still restores function. As far as this note can tell, nobody has assembled that.
Three: the regulatory lever. Since 2022, American law has let evidence about a shared platform carry from one approved product to the next. It’s new. Making it truly transferable will take convening, not research.
And one thing back: the scarcest resource for these patients isn’t money; it’s expert attention. America has about 1,200 certified medical geneticists for 25 to 30 million people with a rare disease. You can’t train your way out of that this decade. You might compute your way out. Whether a machine can do a geneticist’s workup is not known. It can be measured cheaply on cases that are already solved. That’s the experiment to run first.
It’s also where our work touches yours. Lightwave is designing a machine that gives one conversation a million tokens a second. It’s a design. Nothing is built. This is a thought experiment.
So here’s what would kill it. If delivery to one tissue teaches you nothing about the next, the ceiling is hundreds of diseases, not thousands. If regulators want a full package for every disease, the factor of a thousand disappears. If a dose still costs millions, nobody can pay. Each of those can be checked. Cheapest first.
Five thousand diseases, not 5,000 discoveries. A handful of processes compiled 5,000 times. What’s missing is the address and the permission. One is shaped like a prize. The other needs a convenor. That’s why this is addressed to you.
7. Civil Unrest, Technological Disruption, and Human Adaptation
All right, I want to take a second. I put out a Substack today, and I keep encountering a lot of fear out there. I wanted to address that. I ran some studies with Claude and asked about how every generation thinks it’s in a crisis and that its crisis is the worst. That instinct is human, and people are usually wrong.
So I asked Claude to score the last 250 years of American upheaval, with civil unrest on one side and technological disruption on the other. Here you can see the chart on civil unrest. Where we are today ranks number nine on the list. The Civil War and secession, the American Revolution, Reconstruction, the Civil Rights Movement, the Great Depression—these were all far worse on a multitude of parameters than what we’re seeing today.
I think people aren’t students of history enough to realize the disruptions we’ve had. On technological disruption, the top 10 over American history put where we are today with AI and robotics at number seven, compared with railroads, textile mills, farm mechanization, electrification, and so forth. I found this fascinating.
Do you agree with these rankings, Peter? Because I certainly don’t.
I agree on the civil unrest. If you look at—
You think civil unrest now is worse than it was? Number one is the worst. Number nine means we’re very low on civil unrest compared with what America has experienced over its 250-year history. Do you subscribe to the Strauss–Howe generational hypothesis, which holds that American history is stuck in 80-year cycles of crises? The last one was the World Wars and the Great Depression; before that was the Civil War; before that was the Revolution.
That’s how it pans out for more than just U.S. history. It pans out over many centuries. But what do you think about this on the civil unrest side?
Nonsense. If I compare what we’re seeing now with the Civil War and the American Revolution, the economy is ostensibly doing really well. We’re seeing innovations left and right. I don’t perceive a crisis.
Maybe you’re reading this wrong. We’re very low on civil unrest today compared with history.
I don’t think we’re even in the top 10. Maybe I’m just not paying close enough attention.
If you take the list at face value, number 10 is the financial crisis of 2007 to 2012, which was nothing compared to the Civil War.
Yeah.
So you run out of really bad civil unrest pretty quickly on this list.
On the second ranking, though, even that may be AI downplaying itself. You really think AI and robotics are only the seventh biggest tech disruption in American history? It should be number zero, I think.
The parameters I ranked were job and economic disruption, social backlash, and public fear—the primary things people are expressing today and expressed historically. My point is that there’s a lot of fear out there, and people need to realize it’s been a lot worse in previous times in history.
Can I make a couple of comments? I agree with Alex on the first slide. It’s way—It’s not really that bad compared to—
Yeah, I get it. I get it.
That part is right. But I’d also say this would fall off the list for me in terms of what’s happening. This is radically amplified by social media, and people think it’s much worse than it is. Therefore, everybody gets freaked out, with their amygdalas going nuts all the time.
The historical perspective on the second slide is awesome, but all those previous changes left intelligence as the adaptation mechanism. This is the first revolution that automates the adaptation mechanism itself. This means a complete rethink of what it means to be human.
There’s one other part to this: the metabolism is so accelerated. This is another reason. When I talk to audiences, I basically say, look, the printing press changed the world. We call it a Gutenberg moment because of the fundamental transformation it created in society. By the way, it took 100 years to absorb the printing press, and it was the most violent century in the history of humanity.
We call it a Gutenberg moment because it created such a rift in humanity. We have 20 of those hitting us right now. AI is clearly a Gutenberg moment. Life extension is a Gutenberg moment. Blockchain and Bitcoin are a Gutenberg moment.
We can’t absorb this. This is why all our institutions are struggling: they’re not set up for today’s world. Can I give my quick thing about marriage on this? Have you heard me talk about marriage?
Sure. Go for it.
My favorite example to show the difference between what it used to be like and what it’s like today is marriage. We have all these institutions—the legal system, educational systems. We’ve talked about this nonstop. The institution of marriage is a particularly great example because we invented marriage about 6,000 years ago. Average life expectancy was about 25 years. You got married, had kids, and died.
Marriage wasn’t supposed to last 50 or 60 years. It was supposed to keep you together for 10 years, until the kids were self-sufficient. Then you were supposed to die and leave your resources for the rest of the tribe. We worry about the divorce rate, but the conditions today are completely different from when marriage was first introduced as a human institution. This is a huge example. One of my relatives calls marriage “state-sanctioned torture.” I think that’s the phrase she uses.
Now we’re about to double—no comments. I expect no comments on YouTube from relatives. The big challenge is: how do you evolve the institution to deal with today’s conditions, rather than the conditions when we first created it? This is the big structural challenge. This is why everybody’s freaking out. We cannot absorb 20 Gutenberg moments hitting us at the same time. All our institutions break. In fact, I can’t find one that survives this. One democracy—that’s kind of an issue. Education is gone. Journalism is broken.
For me, that’s the biggest underbelly. It goes to E. O. Wilson’s comment: “The problem with humanity is that our emotions are Paleolithic, our institutions are medieval, and our technology is godlike.”
Yeah. Right.
So, I’m waiting for the cherry on top of this rant, which is you’re going to say, “Blockchain is the solution for marriage.”
No, no, I’m not going to say that. By the way, my wife allows me to make those comments, just to be clear. We have a big structural problem: the scaffolding of civilization is essentially being wiped out by the high metabolism and so much innovation. This is why we’re feeling so much stress in the world.
My favorite other comment comes from Douglas Adams, who wrote The Hitchhiker’s Guide to the Galaxy. He said, “Anything in the world when you’re born is normal. Anything that’s invented when you’re young is a career. Anything invented after you’re 35 years old is just bad for the world. It’s blanket bad.” I think that really summarizes where we are overall. For me, this is the biggest challenge.
My point on the civil unrest is that what we have right now is nothing compared to what we’ve seen over the last 250 years. We’ve survived those, and we’re going to survive.
That’s exactly right.
Yeah. We’re going to have disruption. We will have civil unrest, and we’ve talked about this on this pod. Over the next 1 to 10 years, on the other side of that, hopefully, to your point, is a reinvention of social contracts. The world is going to be very different, and those contracts have to be created. People need to start thinking about this. I don’t trust the government to be doing that. I think it’s going to be people like us and everyone listening.
But realize we’ll get through that. On the other side of that is a world of abundance where we uplift every man, woman, and child. I think that’s the point I want to make.
And just to go back to Alex’s point, my marriage certificate should be registered on the blockchain.
Okay.
8. Elon Musk’s Terrafab and the Race for AI Chips
You see? I knew that was the solution.
Oh my God.
I love you guys.
I love you guys. All right, I’m going to move us forward. Next up, let’s go into the Elon universe.
On Friday, Elon sat down over video with Ron Baron at the 33rd annual Baron Investment Conference in New York. If you don’t know him, Ron is the founder of Baron Capital. He’s one of the largest outside investors in both Tesla and SpaceX, and this conference is his big annual gathering for investors.
Musk said that the company is tracking to have an annual revenue run rate in excess of $100 billion in December. He said it’s quite possible they’re going to hit as much as $500 billion the following year. Even $1 trillion is not out of the question. I want to share five clips from the interview, and then we’ll talk about each of them. Again, my full disclosure is that SpaceX is my largest position, so please count me as biased in my conversations on this. Let’s take a look at the first video on Terrafab. Then we’ll chat about each of these.
The rest of the world will make around 20 gigawatts of AI chips this year, maybe 25. Next year, depending on whose projections you believe, it’s maybe around 35 to 45 gigawatts of AI chips. But if you want to do a terawatt, you need roughly 30 times what current global production will be next year.
For just us. We’re building that for us.
Well, yes, we’re building—
I mean, we could sell it to other people, too, but basically we’re building 30 times what the annual production is in one plant, which you think is going to do—
Where Elon goes, “What the hell are you talking about?” The GDP of the 20th-largest country—one plant. Yeah, it might even make the top 10. Obviously, a plan like this, I think, if it fails, it’ll be crazy, and if it succeeds, it’ll be audacious. So it sort of depends on the outcome, but I think it will succeed at least to some degree.
Hope so.
Yeah. It’s going to cost a trillion dollars. How long is that? But we won’t need that money immediately. It’ll be built in phases. We’re looking at ways to change the manufacturing process for chips to improve the efficiency of chip production, as we’ve done in many other arenas.
It’s always amazing to me: whenever Elon is getting interviewed, or whenever we do one of these interviews, they always tell us, “Peter, that was the best, most technical interview we’ve ever done. It was the best.” I’m like, “The bar is so low.” What are you talking about? Anyway, Elon is right, as usual.
I love the way Ron Baron was saying, “For us, it’s our chips,” right? “What are you talking about?” Elon looks around like, “Okay, I’m just going to blow that off.”
Yeah. The chips are the constraint. I worry a lot about the ability to manufacture and supply the chain, and then physics is getting reinvented as you’re trying to build the buildings. The iterations are going to be so fast. But I think at the end of the day, if you get the natural resources, the power supply, and the robots, you’re going to recycle all of that. You just have to be in the hunt like Elon is and build toward the current target as quickly as you can. The target is going to shift for humanity every six months now anyway, so just ride the wave and fight the good fight.
Elon is cornering markets in power, natural resources, and regulatory approval to build things. That’s going to pay off in spades. But Anthropic is also going to be miles ahead in design and discovery—drug discovery, but also chip design. It clearly can do it right now with GPT-5.2. Unless he catches up on the foundation model front, Dario is not just going to give him access. You saw in that chart Alex was pointing to: there’s stuff at Anthropic now that they’re just not letting out the door. They’re not just going to hand that over to Elon. So it’s a foot race now between the singularity in San Francisco and the manufacturing equivalent singularity in Elon’s universe, which is Tennessee and Texas.
Alex, when does Grok 5 come out?
Elon keeps promising it imminently, in the next few months, and that it will meet or beat Opus 5.5. But Anthropic isn’t standing still either. Maybe just to zoom out a little bit, SpaceX used to be a rocket company. It isn’t a rocket company anymore. It’s a neohyperscaler at this point, and Anthropic is their largest customer.
I almost wonder—you remember how Peter Thiel used to joke, “Well, is it that Palantir owns the CIA, or the CIA owns Palantir?” I’m paraphrasing, roughly. Is Anthropic effectively owning SpaceX, or is SpaceX owning Anthropic? Anthropic is their largest customer. SpaceX no longer caters to government business or NASA. As far as I can tell, they’re aimed squarely at becoming a neocloud hyperscaler: open source for those who can’t compete in compute, closed source, and going after the bottom stack to be the very best Anthropic provider they can.
And chip manufacturer. I mean, putting out—
That’s for Anthropic. Anthropic is their biggest customer at this point. Elon is positioning SpaceX to be the missing bottom half of Anthropic. When he talks about achieving $100 billion in ARR and then $500-plus billion, I have to imagine that’s in the back of his mind. Maybe, maybe not, but it’s an elliptical reference to Anthropic’s revenue trajectory.
Anthropic is right now monetizing the upper half.
Alex, if you're going to guess, it's a race between which two labs?
Well, I think SpaceX and Anthropic are very much in bed at this point. If the race we're talking about is for frontier capabilities, it's between OpenAI on the one hand and the Anthropic upper half, SpaceX lower half, on the other.
But that partnership, that access, that partnership will break at some point.
If that's what they told the Sam alliance—
It just needs to survive until maybe next year. We—
—go to infinity.
Yeah. No, seriously. If autonomy horizons go to infinity next year, it just needs to survive until then, and then everyone can go their own way after that.
You mind?
Well, I've been thinking about this, and there's clearly a massive need for open source and internals, and SpaceX can win that. He should just open-source all his models, and then he should do a deal with Anthropic where he gets their models at cost, basically. He should use that to build better physics hardware and just outcompete everyone on chips and hardware. Why does he need to build the craziest models if he has unlimited deployment capability of Anthropic's frontier models? And why does he need to compete with them when he can have a Tesla Powerwall with a GPU in every company in America?
He should have that dual strategy instead: use the frontier models to build better machines, get them at cost from Anthropic, and on the other side own open-source AI, because that's the stuff that actually matters for society, which is what he cares about. He can win.
Well, let's move to the hardware side.
Wait, wait, wait, wait. I want to make a couple of points. One warning: these are the targets, not capacity. We're kind of hoping we can get there, and we're a ways away from getting there yet.
You say chip manufacturing—
On the chip manufacturing story, we're quite a ways away from the scale he's talking about. But the directionality is absolutely right. There's something else here that's really important, which I've been running up against a lot with companies lately. Earlier this year, I was at the MODEX supply chain conference, and somebody had built a complete end-to-end logistics system. A truck would pull up, a robot would unload it, load it up, and take it over. Then a stock-picking robot would load up totes, and it would get dispatched, and so on.
I was like, “This is amazing. This kind of solves a huge chunk of the supply chain. How's it going?” They said, “Terrible,” because each company has its loading department and its sorting department, and they can't get it into the companies. What Elon has done, recognizing this, is said, “To hell with the supply chain. We need to own our own supply chain,” so you get all the exponentials across the board. Vertical integration is going to become more and more critical for companies going forward.
That's always been his game, right? Dave and I were at Figure yesterday recording a podcast that'll be coming out, I think, after this one. That's Brett's game, too: vertically integrate everywhere.
Oh my God.
Brett is such an Elon. Holy crap.
Yeah.
Did you see that?
Multiple hands. Did you ask about the multiple hands?
I did.
Oh my God. It goes so far beyond, so far beyond Salem [?]. You have to see what he's doing.
Yeah.
Missed that one.
9. Elon’s Vision for 30 Million Humanoid Robots
All right, I'm going to move us forward to the world of humanoids. This is the SpaceX-Tesla, maybe X company. We'll see what happens after the merger. Elon's prediction is that 20 to 30 million robots in America would outproduce all the humans. Let's take a look at this video.
The way to think about physical AI in the form of robotics is that any given economy is the number of people in it times the average productivity per person. If you have a humanoid robot that is extremely productive and can work 24/7, it's working 168 hours a week. It doesn't take vacation or bathroom breaks or need to eat food or anything like that, and the speed at which it can operate is faster than human speed.
You can expect productivity per week that's, on average, roughly five times better than a person, because a person might be working 40 hours a week, doesn't move as fast as a robot, and has to take breaks. So I'd say a minimum productivity expectation for a humanoid robot would be five times that of a human.
Take the population in the United States. How many people are either in school or retired? About half the people are either in school or retired. Of those who are working age, not everyone works or is engaged in highly productive employment. So you're left with perhaps around 120 million people in America who have a net positive output of goods and services.
How many robots would be needed to match that productivity? I guess you'd need about 20 to 30 million robots to match the goods and services output of the United States. They're less expensive to build than a car.
Yes. Initially, I'd say about the same.
But then over time, they would be much less.
So, $30,000 cost per robot on the way to $20,000?
Yeah.
And would we rent those or sell them? What would we do?
Do we have an idea yet?
I think probably rent them.
All right. Interesting. Dave, any reflections after our conversation with Brett on this one?
Yeah, we talked a lot about lights-out manufacturing, or lights-off manufacturing, which is a phrase I hadn't heard before. Once you have fully automated, closed-loop, recursive self-improvement in manufacturing, you don't have to turn the lights on in the building anymore because the robots don't need lights. They're fine in the dark. That's a really cool, compelling visual.
If you look at manufacturing today, it's almost entirely automated already, but the human component that isn't automated dominates the cost. Between the labor cost, the pension cost, and the risk, danger, and insurance, you take all that last part out and the cost comes plummeting down. The question is how cheap these are going to get. Elon said about the same as a car initially, but it'll come down really fast. The car will also come down really fast—
—because manufacturing of everything is going to come down really fast. The other big topic we had yesterday was designing new mechanical devices. That's now a Claude Fable 5.2 thing. It's here today. If you can marry that to manufacturing specific to what you just designed, the proliferation of device types is going to be massive.
We don't all have to have the same exact iPhone or the same exact car. There'll be thousands and thousands of variants and new types of devices coming out by the bazillions, which is gold for entrepreneurs. It's going to be a magical, magical time.
What I heard there, Dave, was we're going to see 5x and 10x U.S. GDP as we start to bring hundreds of millions of robots online over the next five years. Who wants to jump in?
Yep.
I would like to challenge the five-year, or five-times-productivity, thing, because it's a nice number, but running 24/7 doesn't take into account maintenance and reliability, battery outages, dexterity, and all of that. I would challenge that. The direction is obviously accurate, but I think it's going to be a ways before we get there. Alex—
I'll comment that, from my perspective, the most interesting part wasn't his bit about 5x productivity. Although, parenthetically, I would note that the U.S. has approximately 300 million motor vehicles and approximately 350 million humans, so approximately a one-to-one-ish ratio. Elon is predicting a one-to-ten ratio of humans to general-purpose humanoids. Mildly interesting, not super interesting.
The most interesting thing to me is the rental business model. Tesla seems to be so clearly pivoting from selling products to selling subscription services. I think that's right, post inevitably Tesla-SpaceX merger. SpaceX used to be a rocket company and is now a neocloud hyperscaler selling cloud subscription services, selling GPUs by the FLOP.
There is a marriage here that I think is about to happen. On the one hand, you have SpaceX selling GPU hours as metered compute in the data center. On Tesla, before the SpaceX merger inevitably, Tesla is migrating toward selling autonomy subscriptions, Cybercab rentals, FSD subscriptions, and Optimus humanoid rentals.
Presumably by the hour, by the task, something like that.
Which gives him infinite borrowing power because he’s got predictable revenues.
Sure. The markets love it. Not investment advice, but presumably this is more attractive to the markets because it smooths out income. It also translates humanity, to the extent this is the industrial economy for the future of humanity, away from an ownership model for the atoms. You don’t own your GPUs anymore. You don’t own the physical labor anymore. Everything is a rental, and I think that’s what he’s converging toward.
Interesting.
But a rental also with a forward purchase. Like we were talking about earlier, a lot of these things are going to move to five-year commitments. We only have a limited supply, so give me a rental on these robots, but also commit to how long you’re willing to commit to that rental agreement. It’s very similar to data centers. If you weren’t willing to commit to a five-year time horizon, you got frozen out. The same thing is going to apply to a lot of things in the singularity, including robots.
I wouldn’t be surprised if, two to three years from now, OpenAI has its own robot. Anthropic has its own; they’re clearly moving in the direction of their own robots based on their capabilities. I wouldn’t be shocked if you see the frontier labs say to enterprises, “Okay, you’re going to make a one-year commitment, in return for which you agree to purchase N robot hours and N GPU hours as part of a bundle.”
Yeah, I think we’re maybe missing something here. He may be a bit high on his 20 million robots. I think 10 million robots will do it, because 45% of U.S. jobs are manual jobs.
But let’s think about that. You’ll be able to build 10 million robots by 2031, five years from now. You’ll be able to build enough robots to replace every manual job in America. That should be top of the political agenda.
You just made it. You just achieved that goal.
There’s a bunch of—if only the U.S. had midterm elections or something—that might send a message regarding this topic. There are a bunch of union bosses freaking out right now.
They should be.
I met with Mike Rowe, who used to run Dirty Jobs on Discovery Channel. We were talking about the notion that these robots are going to do the dirty jobs. They’re going to do the roofing, air conditioning, plumbing, welding, and all of these things.
Oh, yeah.
Alex, Mike Rowe is an opera singer, too. You guys need to meet and do a duet. That would be—
Inevitably. He is awesome.
As you move out from GPUs, because of espionage or whatever, robotics investment spending is going to go exponential. We even set up a robotics company that I’ll tell you guys about. It’s a bit stealth at the moment. It’s going to go crazy. But as Alex said, the most interesting thing is rental versus purchase. Will you be able to buy a robot as the means of production? Because that basically influences the economy. If you’ve got 10 million robots and you can do the entire physical economic output of America, just think about the power dynamics of that. That’s insane.
It’s insane.
All right.
10. Space-Based AI, Satellites, and the Future of Connectivity
There’s a bigger gap on the transportation side, because if you own a car, it sits empty 94% of the time, right? So it doesn’t take a big dent of Cybercabs and so on to completely change the equation on car ownership to car access. And that will completely break the car industry.
Agreed.
During the conversations, he went on to talk about Starship putting up its first satellite in 2027 and beginning to launch in volume in 2028. He also jumped into a conversation about Starlink potentially disrupting or replacing the telcos out there. Let’s take a quick listen.
I think SpaceX will be able to get very far with the spectrum it’s acquired. But at a certain point, if we’re able to make much better use of the spectrum than a legacy telco, we could just acquire that legacy telco and use that spectrum 10 times more effectively.
People love that. I have to imagine he’s going to do that. I’d be shocked if he didn’t. There were rumors about him buying T-Mobile for a while. We’re going to have Starlink phones, we have Starlink services, and it will decimate terrestrial. Anybody disagree with that?
He’s going to buy a lot of stuff. I think what he’s trying to do is bite his tongue a little bit and not tip the market so prices run up on him. But after everything’s consolidated and the grand merger happens, as Alex has been predicting forever, his market cap will be big enough to start absorbing all these assets that just take too long to rebuild, and there’s a regulatory cycle. He’s just going to accumulate them with this massive market cap and glue them together into this grand vision. But when you start talking about things too early, the stock runs up on you and everything gets too expensive.
It’s the new everything company.
Yeah. I also think, based on the applause from the audience, that maybe folks in that audience had what, from my vantage point, is a misconception: that Elon is just out there to replace legacy telcos because he doesn’t like them or something. As I’ve mentioned on the pod in the past, thanks to Starship, SpaceX has an enormous amount of upmass supply that has to get used for something. It could be used for data centers, could be used for telecom. It has to be used for something. So SpaceX has to have a voracious appetite, given the enormous supply that’s about to come online. They need to justify that supply with new demand. Even if that means force-feeding SpaceX—or I should say Starlink Direct to Cell—via acquisitions of legacy carriers, whatever it takes, they need to somehow force-feed all of this upmass into existing markets. And I think that’s—
Methane conversion in minutes.
Yeah, it has to be used for something. So I think they’re really hungry to force-feed themselves into existing markets.
Can you imagine, though, Alex, if Fable 5.2 finds the neutrino communication that you were talking about and that sort of—
Blew a hole into all—
5.2, right?
Sorry, sorry. Fable 5.2.
Yeah. All this innovation is so fragile in the land of new physics every day and new math every day.
But anyway, you’re right. You’re right.
I’m a simple ASI. I just want my neutrino phone so I can do non-line-of-sight communication from one side of the Earth to the other, so I don’t need Starlink. He might.
So, is this Terafab thing and the space data centers a bit of a head fake? Isn’t Terafab just going to build the chips for the robot brains? There aren’t enough of those when you actually think about the math. And does he really need data centers in space? I still don’t get it. Again, I think that can sustain the upmass narrative, Alex. But if I think about where compute’s going to be, with economically valuable compute as opposed to super-ultra intelligence of 10,000 agents solving any problem, it feels like it’s going to be in the 10 million to 100 million robots that he’s going to—
Prediction by 2040.
Maybe he can build those.
You’re right. The head-fake narrative is really interesting because, a couple of podcasts ago, we were talking about him bringing online 10 gigawatts and then 20 gigawatts of new compute. I said, “In space,” and you were like, “No, no, no, that’s terrestrial.” That’s a huge amount of compute on land for a guy who’s moving data centers into space. I think the head-fake narrative actually has a lot of merit. It could be cool to go off on that a little bit. I don’t want to step on too many landmines there, but he’s assembling the market cap. Once he has the Terafab and the Optimus robots everywhere, the rockets were something he invented before the singularity hit.
Actually, there could be a head—
Let me outline this. I think what’s going to happen is this: robot brains, car brains, home brains from SpaceX. If you get to a certain level of intelligence on the edge, whack them all into boxes, and that’s the majority of global compute. How much actually needs to be in a data center? How many geniuses do you need in space? I think very few, compared with the demand for that. That’s ideally suited for his entire supply chain.
I think the big chips aren’t going to be it. It’s going to be the smaller chips. Sorry, Alex.
Maybe. I can make arguments either way. Given the midterm elections, and legislation under discussion to cap data centers—or, let’s say, light-touch regulation for data centers at 20 megawatts—maybe that creates a perverse incentive to suddenly create a whole bunch of basically mini data centers that are 19.9 megawatts everywhere. That could happen. People are already discussing that, I suppose. But the arrow of history does seem at the moment to be pointing away from edge inference being the predominant destination, or sink, for all this compute. What we’re seeing seems to suggest the smartphone, the iPhone, was the apex of edge inference, or edge compute in general. From the iPhone onward, we’re headed back to mainframe computers, basically data-center-type form factors for all this compute. In which case, for Optimus robots, if we tens of millions of them...
I think there’s an alternative scenario where we have Starlink. It’s really low latency and high bandwidth. Sure, there’s some edge inference in cases where the Optimus robots are cut off and need to operate autonomously for a minute or something like that, but basically, most of the inference for the Optimus robots can be operated from the cloud.
So, to the question of where the Terafab output is going to go, my bet is it’s going to data centers.
Alex wants a Dyson swarm.
Alex wants his Dyson swarm.
It’s not that I want a Dyson swarm. It’s that I think it’s the default outcome at the moment.
It is definitely the end state. It’s just a question of the order of operations.
I have a short-term exciting piece of this, which is that all the national telephone companies get decimated in this model, and you move to global infrastructure monopolies rather than nation-based stuff. I think that’s generally a good thing.
I think that is Elon’s plan. I’m going to show one last clip, which I actually found the most fascinating: Elon saying the demand for digital superintelligence will come from digital superintelligence. Take a listen and let’s talk about it.
The demand will—and this is going to sound somewhat circular—but the demand for intelligence, for digital superintelligence, will come from digital superintelligence. It means that superintelligence will saturate all human needs by far. There will be universal high income, universal extremely high income, universal medical care that is better than the best medical care available on Earth today, where you could wish for a castle or whatever you want, basically, and the digital superintelligence can make it for you as a side project.
So he’s reading from the playbook of science fiction: his notion that there’s a post-economic society in which money doesn’t matter anymore and technology creates whatever you want, whenever you want, for anybody on the planet. Alex?
I don’t think it’s science fiction in the least. For Fable 5.2, I spend a couple of minutes and write a single sentence, and I’ve instantly generated a rather amusing small video game. I think Shelto [?] spent a little bit more effort and got a Portal clone. In my mind, there’s just such a straight line. Rather than spending one sentence prompting a disembodied superintelligence sitting in some Anthropic-operated data center, you spend one sentence’s worth of prompting to a fleet of generalist humanoid robots and say, “Go build me a castle,” and they just go do it. I don’t think this is that far out. I think it’s at most a few years away. I think it’s basically just—
Yeah, you can get your castle and your blockchain marriage and everything else you’re asking for.
You know, Elon’s favorite book series is the Culture series by Iain M. Banks, right? If you’ve read it, that’s the world they’re living in, where everybody has whatever they want on request without needing capital.
Well, not every civilization. The ones that have Sublimed, maybe, is a more realistic outcome, which is also in the Culture. Sorry.
Something interesting that I found: I was up to hundreds of agents building out our new thing. Now we’re down to a dozen.
Because we found the informational overhead was actually detrimental, versus just utilizing the latent space and the KV cache in the right way.
I’d like to give a shout-out to your company, Physical Superintelligence, Alex, where they had the proof of the Aaronson and Arkhipov conjecture on quantum computing. Because now what you’ve got is the class of problems that the superintelligences will solve. A, they’ll have better models themselves to do it. But is it going to be 10,000 of them, or is it going to be a dozen? And B, they will have quantum computers, where now we know the right types of questions to ask them, thanks to that proof. That will be very interesting, because again, what is the bounded space of compute and energy required to solve an arbitrary real-world problem?
You’re preaching to the choir, Emad. I’ll just say I think we’re going to enter a second physics golden age. Mark my words.
Okay, I’ll move us along. Next up: a small step towards AI personhood, or at least a step towards AIs running companies. Alex, I’m looking forward to the conversation. Delaware, the state where most large U.S. companies incorporate, is drafting a new kind of corporate entity. They call it an “artificial intelligence company,” or an AIC. Under the draft, an AI agent could manage the business, record its actions in a log, sign contracts, hold property, and file lawsuits in the company’s name. It may have any lawful purpose other than banking.
This is still a draft; it’s not a law. The lawmakers are going to consider it in January. You remember we discussed on this podcast about four months ago President Milei’s Argentina initiative to allow companies to be run by AI. Contrasting the proposed Argentina and Delaware approaches, Argentina’s plan still keeps a human in the loop, and the company stays liable for harm the AI may cause. Delaware’s draft lets the AI run the firm and shields the human member except in cases of underfunding, fraud, or willful illegal conduct.
Dave, I don’t know if this crossed your screen at all, but we talk about the organizational singularity. This seems like an acceleration of what you proposed.
Yeah. I mean, this is the most important story from an organizational perspective in our overall deck, right? Yeah, the big challenge, as you’ve pointed out, is: who owns it? Where does accountability sit? If an AI does something and the whole thing goes horribly wrong—let’s say you set up a legal patent enforcement company like some of the East Texas companies and it goes crazy filing lawsuits—you end up with a hell of a mess on your hands. That’s one general issue.
The second is: what would stop it from creating 100 companies underneath it? And then where does that accountability sit? We have to rethink this. I remember Vina Gupta pointing out that people complain about government, but the limited-liability capability the government gives you allows you to shield yourself and take big risks in a company. Otherwise, nobody would ever do that. So right there, that’s a really important case for how oversight becomes important and you’re given permission to do certain things. Where does that permission sit? Where did the MTP sit in this model? It’ll be fascinating to see what happens.
I think we need to tread very carefully with this. Do very small, functional, very clear, very well-prescribed areas, and then slowly broaden it out. Otherwise, we’re going to lose the game pretty quickly. I suspect Alex has the opposite view here.
Delaware wants to stay relevant here. Alex?
Yes.
Well, I think Delaware is highly relevant, but I think we’re in the early innings of the federalist—or federalism—take on AI personhood. Before AICs came out of Delaware, we saw Wisconsin, which had, remember, DAO LLCs—decentralized autonomous organization LLCs. Wisconsin—excuse me, Wyoming—was the only other state in the U.S. seriously pursuing AI or autonomous organizations. Now Delaware is pursuing AICs.
But I think the second elephant in this particular room is whether this wants to be a states’ rights issue or a federal issue. You already see the beginnings of Congress getting interested. Is there going to be a federal framework for AI corporate personhood, or does this get delegated down to the states?
There are many other hot-button issues, both current and historic, that touch on this issue of personhood, including slavery. Slavery became, in the American Civil War, the ultimate states’ rights versus federal issue. Do we regulate the issue of human personhood? Did we? Obviously, the outcome was clear. But at the time, was this a federal issue? Was this a state issue?
Similarly, with AI corporate personhood, is this an issue where each state can craft its own way, as Wyoming and Delaware are trying to do, or does this become a federal issue? My expectation is that this is going to end up being a federal issue after a bunch of different states try different approaches to AI personhood.
Yeah, that’s exactly what’ll happen. It’s actually very similar to polygamy in Utah, where clearly nothing in the Constitution makes that a federal issue, but the majority of the country is like, “That’s ridiculous. We’re going to vote against it.” We’ll find a way to make it federal.
Here, I think AI personhood or AI corporate nature is clearly a states’ rights issue. But the federal government will say, “Well, it’s interstate commerce, and we’re just going to move it to the national level,” because that’s what always happens. I prefer more of it to stay at the state level because it’s more resilient to mistakes. But when you see things like slavery, it’s so anathema to the population that, of course, it elevates to the federal level and gets abolished. AI will do something nasty at the state level, and they’ll use that as the excuse to elevate it to the federal level. Then we’ll lose the variety.
Any other comments on this one? Emad, does Europe just ignore this and wait for someone else?
Pretty much. Yeah. The liability profile here is just very different from America.
I think honestly the new Delaware thing is very interesting because you can see Delaware law actually has some crazy legal liabilities for various things. This would actually be a better company structure for most people because it has lower liability. Anyone can sue you and say anything, as you’ve seen with Elon and others, and that can cost millions of dollars in Delaware court. This actually reduces your liability, and for AI-first companies it will probably be a superior structure to run.
All right.
I can’t wait until we can have an AI CEO. I would love that.
Okay.
11. Emerging Robotics Companies and Automation
Okay. Type it into Fable.
All right. I’m going to move us into the world of robotics. I see lots of interesting robot companies all the time, and I think it’d be fun to show them here on the pod. This is one that just came out of Y Combinator in the fall 2026 batch. It’s called Thesius. They’re in the supermarket business, and their claim is, “Lights off at 9:00 p.m., shelves full by 9:00 a.m. Next stop, San Francisco.” Let’s take a look at the video and we can chat about it. Unfortunately, Emad, this only has two arms.
Every night at 9:00 p.m., this store closes, the lights go out, and everyone goes home.
Well, everyone except Theo. Theo is our humanoid robot. Every night, he comes online and restocks the store shelves. Theo restocks up to 400 units per hour and picks 40 units per hour for online orders. Every shift he runs makes the model better, so both numbers keep going up.
During the night, Theo automates the whole restocking process. He unboxes cartons, throws away trash, moves trolleys, and restocks the shelves. During the day, he picks online orders and hands them over to customers. Every shelf in the supermarket is different, and that’s why retail is so hard to automate. We build a brain for this task and buy the body. Theo sees thousands of hours of products and manipulation. The more he sees, the more he can do.
So the question is, do companies like this survive with Figure and Optimus? One has to admire the contrarian bet here, because what we saw with tokens and disembodied superintelligence is that they seem to want to find the highest revenue-per-token use case. Stocking grocery shelves is seemingly unskilled labor, and it’s very unproductive compared to, say, building data centers. To Dave’s earlier point, I have to imagine building data centers is far more lucrative per humanoid robot hour than restocking a grocery shelf.
I view this as a highly contrarian bet against AI replacing the lowest-productivity human manual labor rather than the highest-productivity work. Whether it’ll work out in the end is a question, because every indication on the digital side is that the superintelligence wants to flow to the highest revenue per hour. The company quotes a $60 billion-a-year market in terms of humans in supermarkets stocking and picking. The question is how long they’ll have a market and whether they’ll get unique data. Interesting.
Well, I don’t think it’s good for Figure or for Elon to have these companies pursuing more humanoids without a foundation model in them. They need to set up a roadmap for these entrepreneurs. One of the topics is tools like these Optimus robots and these Figure robots: what tools are they supposed to use? Right now they use human-designed tools, but that’s just stupid. Are we going to redesign all tools so that they fit the new platform? I don’t think so.
Get the specs out there.
I think the beauty of humanoid robots is they fit into the existing human universe, and you don’t have to rebuild it. They can pick up and use anything.
They can. But when you look at a data center construction project, there’s so much threading of wire and fitting of pipes. That’s all done with tools. A pipe wrench that was designed for a human is a really inefficient way to connect two pipes. A group of mechanical engineers at that exact company could design a much more efficient way to connect two pipes. The robot picks up the unit, grabs the two pipes, uses this thing, and does it much more efficiently. That’s a better use of these guys’ time.
Anyway, my answer to your question is they’re going to get obliterated by Figure and Optimus.
And we know that just from seeing the dexterity that’s coming online and the end-to-end foundation model that drives all these things. You can’t compete with that by building some stocking robot, you know, for shelves.
So, what’s your advice to robotic entrepreneurs out there? Don’t do it because you’re going to be crushed?
No, I think you need to focus on, as Salem says, other form factors that aren’t currently on the roadmap of Figure and Optimus: things that clean the sides of buildings, that climb up the sides, or whatever. Different form factors. But also talk to them. They’re really good people, and they need force multipliers around their mission statements. They need the App Store equivalent.
If you said, “I really want to build the next great operating system for the iPhone,” you’d think that was crazy. But then they lay out the App Store—which Steve Jobs was originally opposed to, but they talked him into it—and the App Store unlocked how you interact with the device. The App Store equivalent for the Optimus and Figure robots hasn’t been established yet, but that’ll put these guys on a productive path instead of a head-on competition path where they shouldn’t be.
The bet is that a vertical, deep understanding of that particular domain in the short and medium term will generate a fair bit of value compared to generalized robots getting to that domain space over time. I think it’s a worthwhile bet.
Yeah, I think your point is an important one, Alex. Optimus and Figure will flow to the highest-value-per-hour productivity, and—
Which isn’t going to be restocking grocery shelves in the early days. It has to be far more valuable. To Dave’s point, if you’re going to chase what’s currently unskilled human manual labor, you’d better have robot prices that are priced to match. That’s not going to be the general-purpose humanoids in the early days.
Well, if you’re planning on having your shelf-stocking robot run on a GPU, forget about it. There’s no way you’re going to get the GPU to do it. If you’re doing it with an old-school CPU lying around, maybe that’ll work. All these startups are trying to do it with foundation models—Chinese foundation models—on GPUs inside the physical robot. Usually it’s two GPUs. There’s no way they’re going to get those.
Interesting.
The numbers, though, are quite significant. Do you know how many people the average grocery store hires? It’s 700. That’s a big number. So there’s a lot of head space there.
At $20 an hour or whatever it might be.
I think that if you look at an Optimus or a Figure robot, they’re like freaking ninja surgeon robots that can do anything. These things are way simpler to do if you’re just picking up shelves, with a couple of human overseers. In terms of compute, you could run an entire fleet of them with a Mac Mini or two because they don’t leave the store. They don’t go out onto the streets. They’re doing a very limited range of stuff.
This will be money that Tesla and Figure won’t be looking for. So you’ve got an arbitrage between $20 an hour and even a fully loaded $3 or $5 an hour for these guys. They could build a massive business just going after that segment with simpler robots that run on a local mesh. We’ve got to differentiate those and the market opportunity there. There’s loads of market opportunity.
I wanted to share the story for the entrepreneurs out there, to hear this dialogue about how to think about what you might move into and what you might build. Alex?
I would say this is advice: chase the high value, not the low value. If I had to build a company like this, an example of something I would not do is chase basically minimum-wage human labor. It’s low productivity, and it’s not transformative. It’s not how you win this particular game. The way capitalism works is you need to address problems in the market that are both acutely felt and scarce and valuable.
I was just looking this up, but my understanding is the average grocery store in the U.S. has 26 to 75 workers. This is not super-skilled labor.
No, it’s a huge number of part-time workers and all sorts of things—late-night folks who show up to do things. There might be that number of full-time workers, but it’s a big number. I used to do a lot of work for Publix and other grocery chains. It’s a pretty—
I was going to say, the way to win if you want to build a robotics company is to build something transformative that’s attacking the high end of what’s hard to automate.
And Elon, in that other video, may have glossed over everybody, but he said, “The AI economy is going to feed the AI economy.” I forget exactly how he phrased it. But the entrepreneur hearing that should say, “I need to be talking to Chase Lochmiller, Miller Abalene, or Dario Amodei at Anthropic.” That’s the customer base you want to be dealing with if you’re looking for the highest-leverage use of your robot, not a grocery store. Society will get to the grocery store eventually, but the highest-value use cases are going to go first. That’s the only place GPU compute is going to be used. That’s all within this AI community for the next couple of years.
So all entrepreneurs should follow that advice?
Dave, one of my favorite robot applications that I saw was a large machine, not humanoid, that was deploying solar panels at record speed, supporting the innermost loop in the desert.
I think that’s a great example of robotics.
Great example. Great example.
I want to move to the high end. One of my other favorite metaphors is the atomic vacuum cleaner. In the post–World War II era, there was this notion that atomic power would be so ubiquitous that you could put some sort of atomic energy device in a vacuum cleaner. But that’s not what happened.
Similarly, while it’s fun to contemplate—and as much as I’d love to see ubiquitous humanoid robots everywhere in the short term for grocery restocking—I just don’t think that’s the very short-term direction. Let me share the very high-end, or high-frontier, use of robotics.
On Saturday at Kennedy Space Center, NASA announced its recent moon-base contract award. Jared Isaacman, the greatest NASA administrator we’ve ever had, said, “We are building the moon base mission by mission.” And when the crew returns to the lunar surface, “the first elements of the moon base will already be waiting on the surface.”
12. NASA’s Lunar Missions and the Future of Space Exploration
They announced what’s called the CLPS 2 contracts, which stands for Commercial Lunar Payload Services. It’s the next phase of NASA’s program, hiring a whole slew of entrepreneurial and larger companies. Eight companies are now competing: our younger commercial firms, like Firefly, Intuitive Machines, ispace, and Voyager, and the other four: SpaceX, Blue Origin, Lockheed, and Northrop. It’s a 10-year contract with a $10 billion ceiling.
I’m very proud that two of these companies—Firefly and Intuitive Machines—came out of the $30 million Google Lunar XPRIZE that we launched back in 2008. These missions are slated to land on the moon in 2027 and 2028. Blue Origin received a contract to land a nuclear-powered rover called Promise on the lunar south pole in 2028. Super cool stuff, a super-high-value use of robotics. Any comments, Emad?
Yeah. Why nuclear as opposed to solar? Do you want to be on the dark side of the moon or something?
The South Pole is where the water ice is, and the water ice is in the permanently shadowed craters.
Right. All the water ice from cometary and asteroidal impacts hits the lunar surface, sublimates, goes directly to gas when it gets hit by sunlight, and escapes because of the moon’s low gravity. But it’s in the South Pole’s craters that you can find and mine the water and hydrogen from this ice.
I love the fact that NASA supports this. This is how SpaceX got started, with the commercial crew contracts. When the shuttle went down, they offered Elon a contract back in December 2008, when he was on the brink of disaster. He’d just had his third failure—his first rocket, Falcon 1, had failed for a third time. He’d scraped together enough money to put Falcon 1 up for a fourth launch, Tesla was going out of business, and he got a contract. This is NASA supporting entrepreneurship, which I love. Alex—
I think there are two things here. First, it’s great that NASA is acting as a convener, which is what it really should be doing, and then using the private sector to supplement everything else. The second thing that’s really interesting here, going back to the high-paying jobs—D3 jobs, dull, dirty, dangerous—is that this is where you’re going to use humanoid robots, or whatever robots, in the highest-value use cases for quite a while before we get to sweeping my garage.
At least the dangerous part. I’m familiar with the dirty, dull, dangerous aphorism as well, but increasingly I’ve come to the view that it’s dangerous and valuable more than dirty and dull. Dirt and dullness, especially. I have to imagine doing stuff on the moon is anything but dull, but the motivation here is that it’s dangerous and valuable.
I also think this is straight out of the plot of For All Mankind, where we start to see nuclear energy deployed on the moon, and the space program drives transformative energy applications that aren’t necessarily happening on Earth. Over the past year, for a variety of reasons, including political reasons, I’ve come around to the view that the preview of what a post-singularity world looks like is increasingly in space, not in the terrestrial environment. It’s just a pain in the neck to create a post-singularity future in this legacy terrestrial environment.
If you want to see what the post-singularity world looks like, you have to build in a new frontier rather than an existing, well-established ecosystem. Maybe in the future, as we’re starting to see previewed with this NASA program, we finally get our atomic- or nuclear-powered vacuum cleaners—but on the moon. We get our atomically powered vehicles on the moon.
Yes, exactly.
You know, I had such a fun morning this morning. I did a podcast with Bill Shatner and his daughter, Mel, about longevity and all of these areas. His son-in-law, Joel, is a huge fan of the pod. I’m excited that William Shatner is going to be the closing speaker at the Abundance Summit in March. You guys will all be there. I invited Jared Isaacman to come and co-host that conversation with Captain Kirk, so we’ll see if that happens.
Frontiers are important. We need more frontiers, and I think we’re starved for them right now. We have space and superintelligence, and that’s about it. We need more.
Yeah.
If you think about it, the highest-tech place on the planet is probably San Francisco right now, but it was the last place to be populated. Western Europe and China expanded out, and then the United States came into existence in 1492 or 1776, followed by westward expansion. The last place people ended up was the place that was pristine and had a clean sheet of paper, and now it’s the most high-tech in the world.
So if you say the next frontier is the moon, for exactly the reasons Alex just said, it’s not messed up by regulation yet. It would be consistent.
There’s another frontier, which is inner space. This is where Eastern religions and philosophies do very well. They go and think carefully about what it means to integrate into existence rather than trying to model it or explore it. So there’s that, too.
Which could also radically help with technology, by the way. That’s—
Maybe. If I were to construe inner space as a bunch of people uploading into data centers, then sure. But inner space as historically construed in Eastern traditions isn’t necessarily a collective space that can be shared. It’s perhaps more individual, in which case it may not look like a social frontier.
Oh my God, I agree. We could leave that separate. I just backed an amazing company that’s in stealth and working on the human connectome. I can’t wait for the uploads to come. Emad, close us out on this one.
Again, all the science fiction is becoming reality, right? We keep talking about these TV shows and other things, and what was once not possible now seems to be structurally possible. We’re not batting an eyelid at lunar nuclear rovers and stuff being done out there, because I think some of these lines are becoming clear. We’re seeing big pushback starting on the ground, and I think we can all assume that’s going to get bigger. So we’re literally looking to space and beyond for other things, and that’s kind of exciting because it increases the impetus to do this, even as the mechanisms for getting there are all coming into being.
There’s no doubt it’s going to be a two-way thing. Stuff will be figured out on the moon and beyond, and it will come back. As you said, the convening factor is one of the most important things: how do you create structures for all this innovation to occur and then benefit everyone?
It’s almost like the Europeans fleeing the constraints and going to the New World.
I want to argue that the Old World is becoming the old world, and the New World is becoming the new world.
Yeah. Humans have always wanted the frontier. On the frontier is where the greatest innovation, the greatest social revolution and reinvention, occurs—where people can go start afresh. We need new frontiers.
This is innovation at the edge. This is Buckminster Fuller: you can never fix an existing system. Go to the edge and build a new system, and let that become a new gravity center.
I think it’s the same thing we see in ecology, where there’s a forest fire or some new biome opens up and you see succession. You give every form of life an opportunity, to Dave’s earlier point, to start with a blank sheet of paper.
Since the American West closed at the end of the 19th century, we’ve been a little starved of blank sheets of paper to re-envision how society should work and how technology could integrate with society. The internet was a new space, and that was great for a while, to the extent that it was the frontier. Superintelligence is a frontier right now, but I think we’re pretty starved as a planet, as a human civilization, for new frontiers. Space could potentially give us a longer-duration, longer-release frontier.
You know what I’d add to that, Alex? It’ll take a while for the moon to really become the true New World. But between here and there, the Cayman Islands became an important financial hub, as did Bermuda.
I’m still looking for biotech land to emerge—some regulatory environment. What’s different about AI is that, when the Western frontier was being developed, you needed huge amounts of land for mining gold or raising cattle. AI doesn’t need an enormous amount of physical space.
If some South American nation or another region of the world wanted to become the stepping stone to the Moon, the door is wide open. You need a regulatory-friendly environment. I know Emad comes into contact with tons of world leaders, and these guys are super accessible. You think of governments as being very hard to talk to, but in these smaller nations, they’re actually incredibly accessible. So there’s room for that.
Let’s talk about cancer. I know from the member database at Fountain Life that some members come in thinking they’re healthy, but it turns out 3.3% of them have a cancer in their body they don’t know about.
That’s right. The majority of cancers we screen for aren’t necessarily the ones that take lives when they’re found at a late stage. We know that when cancer is found early, the chances of a cure are much higher. It’s much easier to treat a cancer when it’s found early rather than late.
We’re finding that over 3.3% of our members have these cancers that otherwise wouldn’t have been found or detected.
It’s interesting: people don’t feel the cancer until stage 3 or stage 4. If you don’t know what’s going on inside your body, it’s like driving your car with your eyes closed. When members come through, how do they detect cancers?
We’re doing full-body MRI, and we also do early cancer detection screening. This is very important. These aren’t typical tools used in conventional care when it comes to prevention.
This is a hard thing, because currently these aren’t studies that insurance would cover yet. But the goal is to collect these numbers, do the research, and work hard to democratize wellness.
At the end of the day, you can know what’s going on inside your body. It’s your obligation to know. Check out Fountain Life. You can go to fountainlife.com/peter to get access to the latest technology to help you detect cancer at the very beginning, at stage one when it is curable, before it gets to stage three or stage four in your world to hurt.
Let’s move to our last story, where more billionaires correlate with less poverty. There’s a lot of resentment against tech billionaires and the concentration of wealth. For our final story, I want to look at a post by economist Christopher Clark, who examined cross-country data showing that more billionaires per capita is associated with less poverty. He wrote that a 10-fold increase in billionaires per capita goes with a 43% reduction in poverty. He noted that Sweden has more billionaires per 10 million people than the U.S.—34 versus 30. He stressed that the graph shows only a correlation, and asks what the cause is.
Let me show the graph here and let’s talk about it. I think it’s an important realization. There’s been so much discussion about the billionaire tax and resentment out there. Here’s the chart showing the data. We see the percent share of the population under $30 a day—that’s the definition of poverty—and on the bottom we’re seeing the number of billionaires per country. It’s a pretty stark correlation. Obviously, these people are building companies, creating wealth, employing people, and paying taxes that flow through to social services. Emad, your thoughts?
I think this is unsurprising to a degree. Where you have billionaires, you typically have the infrastructure to build billionaires, right? As you said, there’s the tax and the other element of it flowing down.
If we move into this new phase, you see companies being launched that can get to a billion dollars in no time whatsoever. You’re seeing Anthropic, I think, is going to have almost 5,000 millionaires made and dozens of billionaires. It’s a crazy time, whereby even the companies at the top left can create environments for millionaires, centimillionaires, and billionaires to come about. I think that’s very exciting to think about, as well as hopefully lifting up the base level of everyone.
Dave?
My read on this chart is that the top right is suck land. That’s where you don’t want to be: lots of poor people and lots of billionaires. It’s a world of suck, and everybody thinks socialism is somehow correlated. Look at the world of suck. It’s just sucky, sucky governments. Look who’s up there. I’m not going to get myself in trouble by reading the slide. You guys read the slide. But it has nothing to do with socialism. It sucks for everybody. It’s obvious on this chart.
Look at the thriving countries in the bottom right: Switzerland, Germany, and the United States is down there. It’s not a question of left versus right. It’s a question of good ideas versus bad ideas. This chart tells you everything you need to know. You want your population to be thriving, and at the bottom right corner—or anywhere on the bottom of this chart—you’re doing really well. The line is kind of a loosely fit line. I think the real story here is good ideas percolating to the top and getting out of partisan messes.
Two or three things. First, naturally, if you have a lot of billionaires, it may indicate a generally robust economy, and therefore you’ll get fewer people in deep poverty. So I don’t think correlation is causation here.
I do think there are a couple of interesting things. The Sweden example is really interesting because it challenges the idea that entrepreneurship and a strong safety net are incompatible. I’m Canadian. It’s awesome to have a decent social safety net, with health care provided for you, so you’re not worried about your family going bankrupt if somebody gets sick.
The big question in all of this is: Can we lift the bottom? The hell with how much the richest people have in the world. Peter, you and I talk about this all the time. Can you lift the bottom? I think the demonetization and democratization of all these technologies is where you’re going to see some incredible magic coming out of this. The big metric I’d like to see, when talking about benchmarks, is how quickly we can lift the bottom.
Yeah. I want to make that point again: we will have trillionaires living forever on Mars, as long as every single human on the planet has access to all the food, water, energy, health care, and education.
If I were a trillionaire, by the way, I’d rather be living in Bali, just to make a point. Happy to visit Mars just as an exploration. But here’s my question for the group: If you have an AI economy, would you rather have twice as many billionaires or drop the cost of living by 50%? I’m going to suggest that the second part is better.
Inevitably, I’m going to question the premise. I think that’s a false trade-off. The cost of 2026-era living becomes too cheap to meter sometime in the not-too-distant future.
I would also argue, and I have argued, that we want to live in a world with the most possible trillionaires. The world doesn’t have enough trillionaires. Right now, we have approximately one. Maybe more than one if you count certain nation-state quasi-monarchs, but approximately one. I think the world needs billions of trillionaires. That’s the world we should be marching toward.
I think you nailed it, Alex, in the sense that there are dictator trillionaires and productive trillionaires.
The key to tax policy is to tax the former and let the second group reinvest. It’s not about taxing more or less. It’s not about taxing trillionaires or not taxing them. It’s about targeting taxes at unproductive parts of the economy and reinvesting in productive parts of the economy.
I feel like I’m a trillionaire in inner space.
So there we go.
As my grandfather used to say, “Dreams, that and a nickel will buy you a cup of coffee.”
Yeah, and today that and five bucks will get you a Starbucks coffee.
That’s right.
They said it was just science fiction, but every fiction starts with someone’s vision. The future isn’t something that happens to us. It’s something we build one moonshot at a time. I’m a man with a plan, and the moonshot is my mission. Divided by hypocrisy, inspired by a vision. I know I’m not the only one, tired of division.
So the future I’m creating will shape the way we’re living. Abundance is my nature; this is the gift we’re giving. Inspired by my moonshot mission, inspired by a Star Trek vision. Make it real, what I saw on television. There’s no replacement for what’s real, though it inspires what we feel. If we lose our dreams and drift away, life just seems to waste away.
Hanging on, but the power’s in what we do. I see a society before me where we live among the stars, inspired by each other to explore beyond Mars. This is a Star Trek future, but one we must create through your moonshot mission. It’s time to innovate. It’s time to innovate. Your moonshot cannot wait.
Only certain things can’t be done by you. Only you can see the lens of view and through. Empowered by your heart is how you make it true. You need us just as much as we need you to transform this moonshot mission into a united human vision, where poverty is extinct, disease is historical fiction, and energy is endless, with knowledge without restrictions.
As we explore the cosmos, not through competition, but through collaboration—that’s a composition. Think about the phone in your pocket: more power than Apollo’s rocket. Now, someone moonshot, someone who wouldn’t stop. Every breakthrough that we’re living started with someone giving everything they had to make the world in which we’re living.
The mistakes we made before, the limits we swore were carved in stone—we’re carved in stone—until someone alone said, “I’ll make it better and change the world forever.”
All right, gentlemen, and that’s a wrap. Thank you for a great show.
Awesome conversation.
Fun times.
Thank you.
Yeah.
Love you all, guys.
Love you, too.
I’ve got to go give a talk on top. I’ve got to give a talk on top of a dugout. Stand on the dugout for photos.