Anthropic Files $965B IPO, Trump Signs AI Executive Order, and ChatGPT Crosses 1B Users | EP #262
Summary
- Anthropic’s confidential IPO filing turns frontier AI into a public-market asset class, with Polymarket assigning a 60% chance of a first-day valuation above $1.8 trillion. Dave Blundin argues investors are underestimating the resulting liquidity: a 5,000-person company could gain enough “dry powder like we’ve never seen before” to support thousands of billion-dollar acquisitions. Peter Diamandis calls the revenue growth unprecedented, while Emad Mostaque argues that useful AI products can justify extraordinary valuations.
- ChatGPT’s reported one billion monthly users make distribution—not just model quality—the central strategic moat. The show contrasts OpenAI’s 62% annual growth with Claude’s 56 million users and 640% growth, while Dave Blundin cites Sam Altman’s prediction that intelligence costs will fall 100-fold over 18 months. The next acquisition war is for the coordinating assistant beside each user: “Who is your Jarvis? That is actually the only game in town.”
- Trump’s executive order preserves the US labs’ speed while bringing national-security agencies inside a voluntary 30-day pre-release window. Alexander Wissner-Gross sees a difficult balance between preventing a 90-day handicap against China and reviewing privately developed cyber, biological, and chemical capabilities. A panelist’s geopolitical framing is “full spectrum dominance”; Blundin calls the result the right temporary choice, but “it doesn’t solve anything in the long run.”
- Biosecurity is becoming both a regulated physical chokepoint and a restricted AI product category. The panel backs mandatory screening of synthetic-DNA orders after citing a researcher’s $100,000 reconstruction of horsepox, while acknowledging that capable models can run locally. The proposed defense expands from screening orders to government-funded analysis of every sequence and environmental-DNA baselines—because pandemics move “at the speed of airplanes,” while information moves at the speed of light.
- OpenAI’s robotics hiring signals that frontier labs are closing the compute flywheel from software into construction, data centers, chips, and embodied data. Mostaque says the Sora video team moved into robotics and argues physical robots will become a larger, longer-lived market than GPUs. Blundin calls robotics a “very, very good 10-year investment theme,” especially for talent seeking the equivalent of an early OpenAI career.
- Microsoft’s seven in-house models reduce its OpenAI dependency, but the panel does not yet see a frontier-lab comeback. Its Excel model reportedly matches GPT-5.4 at one-tenth the resource cost, yet a panelist calls the broader suite “mid-tier,” while Mostaque describes it as specialized office intelligence rather than a path to AGI. The competitive lesson is that brands carry little protection: “It’s a battle of people, not companies.”
- The labor data discussed does not yet show an AI jobs collapse, though it may show a hiring freeze and widening disadvantage for new graduates. A panelist says employment across his portfolio doubled as industry experts became software builders, reversing his own year-earlier expectation of job loss. Mostaque’s hedge matters: genuinely competent AI and robots arrive “next year” in his view, so today’s resilience does not remove the need to redesign ownership and income flows.
- Longevity is crossing from speculative science into sovereign, venture, and clinical-scale capital deployment. Russia reportedly committed $26 billion, New Limit raised $435 million at a $3.1 billion valuation, and VERVE-102 cut LDL 62% for up to 18 months after one infusion in a Phase 1 trial. Wissner-Gross calls the latter “Star Trek-level medicine”; the closing discussion argues that longevity could attract huge capital “even faster than robotics.”
Deep dive
1. Washington framed AI policy as speed with a narrow national-security window
Peter Diamandis framed Trump’s executive order as a rejection of heavy regulation and permission-based development: labs are asked, voluntarily, to provide models 30 days before release, while federal agencies deploy AI-powered cyber defense. “We compete. We don’t constrain.”
Wissner-Gross called the policy downstream of the “Mythos moment”: private models are commodifying zero-day discovery that once resembled core NSA research. The same governance question will recur when privately built systems begin producing biological, chemical, and physical discoveries with national-security consequences.
His trade-off was explicit: a 90-day delay “could have meant all the difference” against Chinese models, but no pre-release visibility could also be dangerous. Whether 30 voluntary days works should become apparent “in the next few months,” because that is the technology’s operating timescale.
A panelist’s geopolitical reading was “full spectrum dominance”—intelligence joining superiority across air, land, and sea. Blundin supported the watered-down, relationship-driven arrangement as the correct temporary decision, while conceding that personal access and self-regulation are not durable ways to govern a country.
2. One billion ChatGPT users make the personal agent the prize
ChatGPT reportedly crossed one billion monthly active users roughly three years after its November 2022 launch, versus a decade for YouTube, eight years for Instagram, and five for TikTok. OpenAI’s annual growth was put at 62%: “Nothing in history has scaled this fast.”
Claude remained much smaller at 56 million monthly users but was said to be growing 640% year over year. Blundin compared Claude with Apple and OpenAI with Microsoft, then stressed the unserved market: billions of people still lack access, including users constrained in China.
Blundin cited Sam Altman’s prediction that equivalent intelligence would become 100 times cheaper over 18 months. Because the first billion arrived largely through organic adoption rather than classical paid acquisition, he expects customer acquisition after the labs’ IPOs to accelerate growth further.
Wissner-Gross recalled Altman preferring a billion users to the strongest model because distribution is the harder moat. With GPT-5.5 now, “as far as I can tell,” also the strongest model, the strategic contest moves to the assistant coordinating every other agent: “Who is your Jarvis?”
3. Security carve-outs are shrinking the “G” in AGI
OpenAI’s Rosalind BioDefense gives trusted government and public-health researchers specialized tools for outbreak detection, surveillance, and vaccine development. Peter saw such health and security programs as both useful infrastructure and political protection against broader efforts to slow the labs.
Wissner-Gross’s warning was that “the G in AGI is starting to shrink.” Biological and cyber capabilities that might otherwise live inside general models are being separated into post-trained systems available only to agencies and trusted researchers—perhaps the first of many capability carve-outs.
His technical guess combined scaffolding, unshackling, post-training, specialist databases, and tools. A trusted biodefense model could discuss smallpox or query resources unavailable to ordinary ChatGPT users, creating a tier of intelligence whose generality is preserved only for approved audiences.
4. DNA screening is necessary, but local models move risk downstream
Peter co-signed a letter urging Congress to require DNA-synthesis companies to screen customers and sequences. The load-bearing example was a Canadian researcher who reportedly spent about $100,000 on mail-order DNA in 2017 to reconstruct horsepox—a method that could theoretically extend to smallpox.
Screening already costs responsible providers such as Twist Bioscience money, creating a disadvantage when competitors skip it. That made legislation unusually straightforward to the panel: industry is asking for a common rule so safety does not remain an optional operating expense.
A panelist called synthesis controls “both inevitable and inadequate.” He mapped three possible checkpoints—model conception, physical action through DNA synthesis or 3D printing, and after-the-fact liability—and expects countries to regulate different combinations across the full thought-to-action pipeline.
Another panelist confirmed that adapted open models and specialist biological models can operate at the edge, making prompt-level enforcement difficult. The public-sector comparison was stark: government represents about 20% of world GDP but perhaps 0.01% of tokens, so governments should fund sequence screening and environmental-DNA baselines before deviations become outbreaks.
5. Robotics closes the frontier labs’ self-improving infrastructure loop
OpenAI is hiring an in-house robotics team to support skilled workers, construct infrastructure, and eventually provide personal robots. Wissner-Gross called this the “innermost loop”: robots build fabs and data centers, which build chips and host models, which then improve the robots.
With Stargate reportedly moving from owning facilities toward leasing them, Wissner-Gross wondered whether OpenAI robots will build and maintain third-party data centers. Either way, he sees compute construction as the primary humanoid use case and domestic robots as a later beneficiary.
Mostaque said Aditya Ramesh and researchers from the Sora video effort moved into embodied robotics. His thesis is that robots will exceed the GPU market, depreciate more slowly, and remain productive for years once hardware reaches the competence demonstrated by systems such as the Unitree G1.
6. Anthropic’s IPO would create unprecedented acquisition currency
Anthropic reportedly filed confidential IPO paperwork and could become the first major frontier lab to list. Polymarket assigned a 60% probability that its first-day market capitalization exceeds $1.8 trillion, roughly the valuation the hosts associated with a prospective SpaceX offering.
Blundin’s response was not to become numb to trillions: “This flow of money is the biggest in the history of the world by an order of magnitude or more.” With only about 5,000 employees, Anthropic could finance an enormous supplier ecosystem and “thousands of billion-dollar acquisitions.”
Peter floated Anthropic revenue at roughly $5 billion to $6 billion and described the valuation as about 20 times revenue; Mostaque said such multiples had previously been closer to 50 times and that $100 billion to $150 billion in revenue next year would not surprise him. Both Anthropic and SpaceX, he argued, would be oversubscribed.
Wissner-Gross called public listing a public good because retail investors have largely missed the private appreciation of frontier companies. Peter added that disclosure requirements may be especially valuable for a lab presenting itself as safety-conscious.
7. AI economics are breaking the link between headcount and enterprise scale
Peter compared the time to a trillion-dollar valuation: 42 years for Apple, 21 for Google, 24 for SpaceX, about 10 for OpenAI, and roughly five for Anthropic. Anthropic’s reported $9.4 million of revenue per employee was nearly four times Apple’s $2.5 million.
Wissner-Gross predicted a one-person centacorn or teracorn within 10 years, but questioned whether revenue per agent will remain meaningful. End-to-end teams may blur into one collective system, erasing the boundary between individual agents and a team-level agent.
Wissner-Gross described an AI economy capable of financing itself: after an IPO, a lab could spend $100 billion directly on disease-solving agents without passing through banks, Main Street, or legacy businesses. “It could entirely build an economy of its own.”
8. Microsoft rebuilt the stack but not yet the frontier
At Build 2026, Microsoft introduced seven internally trained models spanning reasoning, code, images, video, and transcription. Its Excel model reportedly matches GPT-5.4 while running 10 times more efficiently, and its Mayo Clinic collaboration targets a frontier healthcare model.
A panelist’s answer to whether Microsoft is back was categorical: “No, they’re not in the game.” He viewed the releases as mid-tier systems comparable with models OpenAI and Anthropic had shipped months earlier, constrained by insufficient frontier compute and talent.
That panelist’s historical analogy put OpenAI in Microsoft’s old role and Microsoft in IBM’s: the platform owner let its partner run away with the next wave. Mostaque agreed, calling Microsoft’s destination “humanist office-based intelligence” optimized for hundreds of millions of Teams users, not general superintelligence.
Another panelist reframed the contest as people rather than corporations. If Mark Zuckerberg could offer OpenAI researcher Mark Chen a reported $1 billion package, Microsoft’s missing ingredient was the will to recruit “the five or 10 cannot-miss great AI researchers” needed for a genuine frontier team.
9. Institutional backlash mixes valid error concerns with self-preservation
The hosts rejected a New York Times analysis finding Elon Musk completed only 19% of 602 public goals over 15 years. Their preferred datum was 75% completion for goals set in 2015; their defense was that audacious portfolios work like venture capital, where roughly 10% of bets produce 90% of returns.
Wissner-Gross read the 130-signatory Leiden Declaration as a rear-guard response to AI progress in mathematics: “AI is going to cook math. AI is cooking math. AI has cooked math.” Mostaque sympathized with displaced mathematicians but argued overspecialization leaves humans unable to combine fields.
Mostaque preserved the declaration’s strongest point: frontier mathematical outputs can become highly convincing while remaining subtly wrong. His analogy was not psychosis but “a grad student that’s really super talented and convincing,” requiring verification even when the reasoning sounds polished.
The American Federation of Teachers’ 10-point plan proposed no screens through grade two, K–12 safeguards, limits preserving teacher responsibility, and a big-tech tax. The panel favored mandatory AI literacy instead; Wissner-Gross pointed to Math Academy and reports of eight-year-olds completing high-school mathematics through adaptive learning.
10. A 50% AI equity tax would entrench power as much as share it
Bernie Sanders proposed a one-time 50% stock tax on major AI companies to seed an American sovereign wealth fund. His premise was that AI depends on a public resource more valuable than oil, so citizens deserve ownership and a voice in how its trillions are used.
Blundin’s objection was liquidity: equity cannot fund public benefits until government sells it, and dumping trillions of dollars of stock could crush the market. He expects the first administration facing a budget or election problem to liquidate the position within “exactly one election cycle.”
Wissner-Gross separated the mechanism from the objective. He supports exploring a sovereign wealth fund or universal basic equity—potentially holding a broad index containing OpenAI, Anthropic, and SpaceX—but rejected forced divestment of half the frontier labs’ shares.
A panelist estimated that a $1 trillion pool—framed as half philanthropic and half from OpenAI—would amount to about $2,000 per American, while making the labs too big to fail and their fund controller immensely powerful. The alternative proposed was placing AI shares into every child’s Invest America account; the discussion also anticipated governments licensing tokens or taxing their GDP value.
11. AI is creating builders before it eliminates workers
The Washington Post’s policy menu ranged from taxing robots and expanding unemployment insurance to retraining, public dividends, or waiting. Against that anxiety, Apollo economist Torsten Sløk was cited calling AI a net job creator, while Cognizant planned to hire 20,000 graduates.
A panelist said employment across his companies doubled because domain experts can now write code and build products without engineers. “A year ago, I would have said it’s going to go down”; instead, converting non-builders into builders has outweighed automation losses so far.
Mostaque’s caution was temporal: society has only just reached competent intelligence, so current data show neither mass job loss nor robust hiring. He expects much more capable AI and robots next year, making early planning for ownership, income, and a life beyond compulsory work essential.
12. Nvidia’s laptop chips are a bid to own Jarvis at the edge
Nvidia’s N1 and N1X move beyond discrete GPUs into complete PC processors. The N1X was described with 20 CPU cores and 6,144 CUDA cores, delivering roughly RTX 5070-class graphics in a laptop package.
A panelist was surprised Nvidia waited so long given its value relative to Intel and its unsuccessful attempt to acquire Arm. One explanation was that its growing share of leading-edge TSMC capacity may let laptop vendors access advanced nodes they could not secure independently.
Peter proposed a consumer beachhead: laptops become the test bed for AI-native operating systems, giving Nvidia direct user data and profiles before assistants disrupt Apple’s position. He also recalled legacy bad blood after Apple stopped using Nvidia GPUs in Macs.
Another panelist called the device a defense against AMD’s Strix Halo integrated chips and a route to “Jarvis throughout the home.” RTX 5070-level capacity could run a three-billion-active-parameter model—or Liquid AI’s LFM-1B model—while Nvidia supplies the open intelligent substrate.
13. Closed-loop cooling weakens the data-center water attack
Satya Nadella said Microsoft’s newer cooling loop is filled once and then operates with effectively zero ongoing water consumption; its average daily use over a year is roughly that of one restaurant. Peter praised him for meeting the criticism directly.
Peter contrasted figures presented for California almond farming—1.3 trillion gallons annually—with 150 billion gallons for all US data centers. The panel’s broader message was that protest narratives overlook design changes and vastly larger agricultural uses.
Mostaque alleged that a widely repeated estimate began with a 1,000-fold mathematical error. Wissner-Gross advised hyperscalers to co-locate data centers with water-production or desalination facilities if they remain on Earth, while retaining his preferred outcome: orbital compute where water objections disappear.
14. Media’s revenue collapse is becoming a trust collapse
Trust in news was placed at 19%, down from roughly 80% in the mid-1970s. Peter projected the trend toward zero around 2030, arguing that audiences increasingly choose individual people whose incentives and worldview they understand.
A panelist’s diagnosis was economic rather than conspiratorial: internet competition stripped newsroom revenue, managers substituted low-cost controversy for expensive reporting, good writers left, and the resulting quality decline accelerated the loss of trust. “It starts this spiral.”
Peter proposed an AI-native newsroom using open models, visible reasoning traces, public contribution, and trusted reporters; another panelist called it a potentially massive opportunity. Peter’s criterion was simple: “Trust is built in one way. It’s by helping people.”
15. Longevity now has sovereign-scale budgets and venture-scale milestones
Russia reportedly committed $26 billion to longevity, targeting 3D-printed tissues, transplantable organs, and epigenetic reprogramming by 2030, with a goal of saving 175,000 lives. Peter linked economics to healthspan: lifespan near 79 but healthspan near 63 leaves 16 costly, painful years.
Wissner-Gross imagined Russia pursuing international prominence through longevity rather than war. Another panelist said wealthy audiences would spend nearly everything for 30 additional healthy years, implying “hundreds of trillions of dollars” in long-run demand.
New Limit, co-founded by Coinbase CEO Brian Armstrong and Blake Byers, reportedly raised $435 million at a $3.1 billion valuation. It is targeting human studies next year for epigenetic-reprogramming therapies, beginning with alcohol-related liver disease because the FDA does not recognize aging itself as an indication.
Peter’s Healthspan XPRIZE offers $101 million within an effort he said raised $157 million; 830 teams are pursuing a sub-one-year therapy that restores cognition, immune function, and muscle toward levels 20 years younger, with a winner expected by 2030.
16. VERVE-102 makes one-shot cardiovascular editing tangible
VERVE-102 uses a single infusion to switch off PCSK9 in the liver, preserving LDL receptors that clear cholesterol. In the cited Phase 1 trial, the highest dose reduced LDL 62% and PCSK9 protein 88%, with the effect sustained for up to 18 months so far.
Wissner-Gross compared it with Leonard McCoy’s miraculous medicine in Star Trek IV: CRISPR base editing plus mRNA delivery producing, “if you squint,” something like a one-time shot against a major cause of heart disease. “This is Star Trek-level medicine.”
The wider search space comes from naturally protective human variants: if some people rarely develop high LDL, Alzheimer’s, or cancer, comparable edits might be deliverable through one-shot mRNA lipid nanoparticles. Another panelist described this as “editing human software”—correcting biological prompts that have “gone a bit wonky.”
17. Compute, distribution, and orchestration remain the compounding moats
Even a perfect AI algorithm would not end compute demand, Wissner-Gross argued. It might produce several orders of magnitude of efficiency and a “DeepSeek demand crash on steroids” for about a year, but Jevons-paradox-style demand would recover and hardware pressure would intensify once algorithmic improvement saturated.
Blundin is testing that future with 170 agents, including parallel particle-simulator builds and thousands of neural-network research ideas. Most experiments fail; the point is learning how to synthesize competitive parallel work before users command “billions” of agents.
A panelist believes an unknown lab can still catch OpenAI or Anthropic technically, but calls it “very difficult” because data and distribution compound. Frontier incumbents can now spend hundreds of billions locking those advantages down, turning the contest into go-to-market execution more than algorithms.
Blundin said incubators produced under 10% of unicorns 15–20 years ago but about 70% now because founders cannot waste time assembling infrastructure. Peter’s routes into the upside remained simple: own public AI assets, use AI to multiply earning power, or build as startup costs collapse.