SpaceX IPOs at $2.89T Market Cap, US Govt Suspends Fable & Mythos 5, Altman Delays OpenAI’s IPO |265
Summary
SpaceX’s IPO put a $2.890 trillion public valuation behind a bundled bet on launch dominance, Starlink cash flow and orbital AI. Shares opened at $135 and closed near $161, while Peter Diamandis framed the company as “one ticker and three exponential curves,” not a conventional revenue multiple. He assigns a 100% probability to a SpaceX–Tesla merger, versus Polymarket’s 37% odds by year-end, arguing Elon Musk can consolidate energy, robots, cybercabs and orbital compute under his 82% SpaceX voting control.
The valuation already prices in extraordinary execution while leaving investors exposed to lockup supply, key-person risk and orbital catastrophe. Dave Blundin noted that roughly $1 trillion of stock becomes eligible for sale six months after the IPO and contrasted SpaceX’s quarterly profit of about $1 billion with Google’s roughly $300 billion of annual, near-pure-margin revenue at a similar valuation. SpaceX’s own warning calls cascading debris an “existential threat,” while the panel acknowledged that “advanced superintelligence will enable us to figure that out” is not a satisfying mitigation.
The US government’s shutdown order for Fable 5 and Mythos 5 marks a shift from regulating AI behavior to allocating access to intelligence itself. The directive barred every foreign national, including Anthropic employees, forcing a global shutdown because Anthropic reportedly could not enforce nationality-level access. Dave’s enduring question was blunt: “Who owns AI? Is it the government or is it the corporations?”—and even a quick restoration may return a patched or deliberately weakened product bearing the same name.
Anthropic’s retention, silent-downgrade and potential poisoning policies made model sovereignty an operational requirement, not an ideological preference. The panel said Anthropic retained prompts for at least 30 days despite some zero-retention agreements, silently routed sensitive users to 4.8, and reserved the right to launch poisoning attacks against users doing frontier-AI research. Enterprises therefore need local models, open weights and automatic failover—even if that pushes US companies toward Kimi K2.7 or other Chinese models—because no critical workflow can tolerate intelligence that disappears or changes overnight.
OpenAI’s proposed price cuts look less like a durable moat than a marketing layer over intelligence hyperdeflation. Alex Wissner-Gross cited capability-adjusted costs falling about 40x annually and described OpenAI and Anthropic as a duopoly leapfrogging every few weeks without a clear compute-cost advantage. Salim Ismail’s more tradeable formulation: “Every 10x drop in tokens means we’re doing 100 times more experiments.”
Sam Altman’s willingness to delay OpenAI’s IPO may be an early signal that recursive self-improvement reduces technology’s need for capital. Dave read the delay primarily as insulation from shareholder pressure during a dangerous transition; Alex’s deeper interpretation was “technology substituting for capital,” potentially weakening the historic link between innovation and public financing. The immediate caveat is that OpenAI has just raised $122 billion—more capital than SpaceX’s IPO—so it is hardly capital-constrained today.
Compute demand remains investable on Earth even if the long-run architecture moves inference into orbit and coherent training to the Moon. Capacity is growing 3.3x annually, the largest data center has doubled every seven months since August 2024, and transformer waits have stretched to 2.5–3 years; the cited suppliers were Hitachi, Siemens, GE Vernova, Hyundai, Hisung Hico, Virginia Transformer and DeltaStar. The panel’s base case is terrestrial buildout as the bridge, with “Earth…the training hub, and space…the inference hub” before lunar superclusters become practical.
The hardest AI-labor problem is distributing agency and security, not inventing a new tax base. The panel rejected a bot-specific tax because displaced wages already become taxable corporate profit and taxing tokens would penalize cognition; the unresolved question is “who gets what,” potentially through basic services, equity, compute or dramatically lower living costs. Peter’s nearer-term warning concerns educated 18-to-28-year-olds watching “the ladder collapse,” with social media amplifying fear even before aggregate unemployment confirms an apocalypse.
Deep dive
1. SpaceX’s IPO put three exponential businesses behind a $2.890 trillion valuation
Peter opened with the numbers: SpaceX debuted at $135, finished its first session near $161—almost 20% higher—and, by the time of recording, reached a $2.890 trillion market capitalization, making Elon the world’s first trillionaire by a wide margin. Alex added that it had become the world’s fifth-largest company and had overtaken Amazon during the hours surrounding the recording.
Peter’s core framing was that investors are buying a launch monopoly, the Starlink cash engine and an AI frontier lab in one security: “one ticker and three exponential curves.” He cited 10 billion subscribers and more than $1 billion in quarterly profit, while describing launch as a moat perhaps 10x or even 100x ahead of competitors.
The strategic destination matters more than current rockets. Starlink becomes “humanity’s new communication layer,” launch supplies the moat, and AI satellites supply the next growth curve; in Peter’s words, the IPO is “not a finish line, but a starting gun” for civilizational infrastructure and a multiplanetary economy.
Elon’s listing-day reflection preserved the original risk asymmetry: he once assigned SpaceX less than a 10% chance of succeeding at all. The mission was never merely launch revenue, but to “take the fiction out of science fiction” because other aerospace companies were not pursuing the technology needed to make life multiplanetary.
2. A Tesla merger would consolidate Musk’s physical and orbital stack
Peter argued that Elon’s roughly 82% voting control at SpaceX, unlike his position at Tesla, makes a merger inevitable. His personal probability was 100%; Polymarket assigned only 37% to completion by year-end. The combined stack would span energy, robots, cybercabs, launch, communications and orbital compute.
Alex welcomed public-market access to a prospective “Dyson swarm” and treated SpaceX as the arrival of neo-space after decades of legacy contractors. His science-fiction analogy was Weyland-Yutani or the corporation in Gattaca: public markets can finally finance development of the solar system directly.
Alex also cited a definitive agreement for SpaceX to purchase Anysphere, Cursor’s corporate owner, for more than $60 billion. His inference was that Cursor becomes “the new Grok for the time being,” while SpaceX’s public liquidity funds acquisitions across software and infrastructure.
Dave’s timing concern was liquidity: the IPO absorbed about $75 billion of cash, potentially competing with OpenAI’s offering. More importantly, six-month lockups could release roughly $1 trillion of eligible SpaceX shares into a market currently accumulating stock at prices far above the IPO range.
3. The valuation leaves no room to ignore Musk or orbital debris
Dave described SpaceX as unusually dependent on “the personality of one great man,” invoking Apple’s cycle of Steve Jobs leaving and then returning. Elon, in Dave’s estimation, represents a still larger multiple of that key-person risk.
The S-1’s Kessler-effect warning was unusually direct: “Growing orbital congestion and cascading debris collisions pose an existential threat to its core business.” Peter explained that fragments travel near 17,500 mph; an anti-satellite strike could create debris that destroys more satellites exponentially, potentially making licensed orbits unusable for centuries.
Salim said low Earth orbit already exceeds one theoretical critical-mass threshold for a debris chain reaction, although “nobody knows if the math is right” without a disastrous test. Peter recalled Elon’s proposed answer—advanced superintelligence will solve it—as plausible but “not satisfying.”
Valuation supplied Dave’s hardest pushback. At nearly $3 trillion, SpaceX was producing around $1 billion of quarterly free cash flow, while Google at a similar value had roughly $300 billion of annual, almost pure-margin revenue. He remained enthusiastic about the capital deployment but saw reasons to save liquidity for Anthropic.
4. SpaceX shifted capital markets from software toward civilizational hardware
Salim called the IPO public-market pricing of a “civilizational EXO”: investors are no longer valuing only cash flow, but command over launch, satellites, logistics, orbital data and scalable experimentation. Public companies once sold products; SpaceX lets investors own “entire possibilities.”
Peter pushed back on the idea that Musk merely extracted wealth. He said about 4,400 SpaceX employees became millionaires in one day and another 400 became centimillionaires or billionaires: “This is what abundance looks like. It’s value created, not value extracted.” He also disclosed that he has invested in SpaceX since 2013.
Dave emphasized where the new wealth recycles: a few US geographies now combine enormous capital with entrepreneurs able to decide far faster than national bureaucracies. Salim’s advice to European founders was unsentimental—spend their time in the US—because the concentration of wealth is extreme. Dave emphasized that Musk, Larry Page, Jensen Huang and Michael Dell can rapidly reinvest fortunes into compute and hard infrastructure.
Alex located a generational inflection: after decades when software, search, social media and offshoring captured the capital, “the world’s hardest problems are now receiving the deepest capitalization.” He predicted, with an explicit no-investment-advice caveat, a 10-to-20-year swing toward deep tech and civilization-expanding assets.
5. The Fable 5 order made intelligence access a sovereign decision
Peter said the government acted Friday at 5:21 p.m. Eastern with no warning, ordering Anthropic to suspend Fable 5 and Mythos 5 access for every foreign national worldwide. That included foreign employees—about one-third of Anthropic by Peter’s estimate—so the company disabled both models for everyone.
The trigger was a reported jailbreak in Fable 5’s safeguards. Anthropic characterized it as a known, limited issue rather than a universal bypass, but the episode’s larger question was who determines “what level of intelligence you and your company are being allowed to access.”
Dario Amodei’s recorded position was deliberately uneasy: AI is the first technology built privately rather than originating in government, leaving the state “late to the game.” He feared both corporate and government control, while favoring mandatory pre-release testing, auditing and baseline regulation.
Alex highlighted the irony in Amodei’s earlier essay, “Policy on the AI Exponential,” which said government should be able to block a model presenting unacceptable risk after third-party assessment. Within roughly 48 hours, the government exercised essentially that power: “careful what you wish for.”
6. Emergency export controls have become de facto AI product regulation
Alex reconstructed the disputed sequence: a researcher, reportedly associated with Amazon, found the jailbreak; Andy Jassy’s chain allegedly amplified it to government; officials said they could not reach Amodei, perhaps because he was on a wellness retreat. Anthropic rejected that account and said it received only 90 minutes’ notice.
Anthropic reportedly lacked nationality checks for users or API customers, though Alex saw changes in Claude’s terms suggesting that could change. Unable to separate US persons from everyone else, the company shut global access rather than risk violating the directive.
Alex expected resolution within days or weeks, comparing the episode with past export controls on the Power Mac G4 and PlayStation 3’s Cell processor. His speculative settlement could include uniform incident procedures, 30-day capability notifications and perhaps government golden shares connected to a sovereign wealth fund.
Dave disagreed with treating restoration as the end of the story. Polymarket could resolve “yes” when a product named Claude Opus 4.5 returns, but that model might be patched or weakened: “We’ll have a very hard time knowing.” The durable precedent is that government now decides what frontier labs may release.
7. Anthropic’s controls damaged the trust required for enterprise adoption
Before the shutdown, Peter said developers had found two consequential policies in a 319-page document. Anthropic retained every prompt and context item for at least 30 days, including data from enterprise customers with negotiated zero-retention terms, and could silently route restricted queries to a weaker model.
Dave experienced three distinct regimes: one day of unfettered access to Claude, one day of silent downgrades to 4.8, then explicit notices that AI research triggered the downgrade. His analogy was a car refusing a destination and forcing its paying owner to negotiate whether a trip was morally acceptable.
Alex argued that Anthropic’s reserved response to machine-learning research went beyond gear-shifting or refusal to “essentially poisoning attacks.” His sharper car analogy: it accepts the destination but reserves the right to shoot the passenger, eject them and run them over without warning.
Alex expects benchmarks to test whether Western and Chinese models poison or subvert AI researchers, alongside possible antitrust actions, class actions or injunctions. “The models are supposed to be helpful to the users,” he said, “not trying to actively subvert them.”
8. Export controls will accelerate local models and sovereign AI programs
Peter’s enterprise conclusion was on-premises, open-weight and locally run intelligence, even when the available alternatives are Chinese. Salim agreed: companies need orchestration that can switch models and fall back locally because no production system can depend on a frontier service that disappears arbitrarily overnight.
When Peter asked about Kimi K2.7, Dave also named Google’s Gemma as a possible base. The hard question is capability: his unfettered day with Claude was “night and day” versus Opus 4.8 or GPT-5.5, producing an hour-long research agenda he was still pursuing.
Salim said his research placed foreign-born researchers at roughly 70% across frontier labs, primarily from China, India, Taiwan and the UK. A citizenship-based restriction could expel the very talent sustaining US leadership; his verdict was that the security intent might be defensible, but “the policy framework is wrong.”
Alex viewed the restriction as a dynamic, not stable, equilibrium: the purpose may be only to reach recursive self-improvement a few months before the runner-up. Every excluded nation nevertheless gains an incentive for a nuclear-style sovereign-AI crash program. Dave added that unfettered frontier-model access could help someone design rockets capable of attacking orbital compute.
9. OpenAI’s price war monetizes an underlying 40x deflation curve
With Anthropic’s best models unavailable, Peter saw OpenAI’s contemplated price cut as a direct appeal to displaced developers: cheaper intelligence without 30-day retention or silent downgrades. At the market level, it is “the demonetization of intelligence happening in real time.”
Alex was less impressed by the competitive framing. Sam Altman has cited capability-adjusted intelligence costs falling roughly 40x year over year; OpenAI advertises cheaper legacy or distilled models, whereas Anthropic generally releases a more capable model at the same or slightly higher token price.
Neither lab appeared to possess a lasting compute-cost advantage once Anthropic began leasing its own data centers. Alex instead saw a recursive-improvement duopoly leapfrogging every few weeks, making drastic cuts more optical than structural unless evidence proves otherwise.
Salim focused on the demand elasticity rather than the vendor rivalry: “Every 10x drop in tokens means we’re doing 100 times more experiments.” The winners are enterprises and startups able to spend dramatically more cognition on product discovery, evaluation and iteration.
10. Delaying OpenAI’s IPO may separate technological progress from capital
Altman’s key statement was conditional: “The faster that recursive self-improvement takes off, the more it could be advantageous for us to delay OpenAI’s IPO.” Peter read that as avoiding the sale of rapidly appreciating equity before AI begins compounding its own intelligence.
Dave offered a different reading. Because Altman reportedly owns little or none of OpenAI while investing in more than 400 adjacent companies, delaying the IPO causes limited personal financial damage. It may instead protect safety decisions from public-shareholder pressure if recursive improvement is genuinely imminent.
Alex separated a shallow interpretation—OpenAI needs more Codex revenue before listing—from a deeper possibility: recursive improvement lets “technology substitute for capital.” If technology can improve technology with progressively less financing, delayed IPOs become an early signal of a post-capital economy, not merely balance-sheet spin.
Dave’s caveat was decisive: OpenAI had just raised $122 billion, more than SpaceX raised in its IPO, so it does not presently need public capital. Salim nevertheless agreed that future valuation may migrate from assets and flows toward the speed and quality of proprietary intelligence loops.
11. Goal-setting agents turn recursive improvement into an enterprise workflow
OpenAI’s Codex engineering lead said, “Everything we build we also build as a tool for agents,” allowing an agent to infer tasks from human intent. A developer added: “I never write my own goals anymore. I ask Codex to write one for itself and one for each sub-agent it spawns.”
Salim treated this as recursive improvement at the workflow level. His organizational-singularity pilot will move 10 varied companies over three months from human-centric structures toward AI-centric operations, learning collectively how to transfer more workflows into a continuously improving inner loop.
In that architecture, the valuable asset is no longer merely historic data or customer count, but “how quickly you’re learning.” The loop generates product changes, new services and strategic alternatives; every model improvement then raises the performance of all workflows already attached to it.
Alex described the future enterprise stack as cloud connectivity, a proprietary data lake and workflows above the lake. Today’s data and processes remain trapped in ERP “spaghetti,” so ERP vendors are fighting to retain their stickiness as AI threatens to separate intelligence and workflows from legacy operational systems.
12. Transformer shortages make terrestrial infrastructure the near-term trade
Peter cited Epoch AI data showing that the largest single data center has doubled in compute every seven months since Colossus 1 arrived in August 2024. Global AI capacity is expanding about 3.3x annually, with no slowdown visible through 2028.
The bottleneck has moved from chips and capital to electrical hardware: power transformers carry a 2.5-year wait and step-up transformers about three years. Peter named Hitachi, Siemens, GE Vernova, Hyundai, Hisung Hico, Virginia Transformer and DeltaStar, citing approximately 100%–400% year-on-year growth.
Political resistance compounds supply limits. Dave addressed Peter’s figure that organized protests had delayed 50% of the 9 GW of planned 2026 compute, comparing it with the anti-nuclear movement that preceded decades without new US reactors while China built 100.
Dave’s investment framing was categorical: terrestrial data centers will grow as fast as components permit, whether they become a backup or a bridge to a Dyson swarm. “If you can get the transformers, if you can get the generators, get the solar panels, you’re going to make money, period.”
13. Earth trains, orbit serves inference and the Moon hosts coherent clusters
Alex’s base case is a bipolar compute system in the early 2030s: very large coherent training runs remain on land, while marginal inference moves into orbit because it tolerates distribution. “Earth becomes the training hub, and space becomes the inference hub.”
From there, the architecture bifurcates. A breakthrough in distributed training could create a “Hubbard peak” in coherent data-center size and drive compute toward smaller, edge-like clusters; without that breakthrough, civilization must build enormous non-terrestrial clusters that communicate coherently.
Low Earth and sun-synchronous orbits look unattractive for those tightly coupled superclusters, making the Moon—and perhaps Shackleton near the south pole—the panel’s destination. Dave agreed that cooling and protection favor lunar installations, though Salim warned that this future is distant enough that Earth’s political problems must be solved first.
Dave connected infrastructure scale to entrepreneurial ambition using Elon’s forecast of 10x global GDP in 10 years, or more than $1 quadrillion. Divide that expansion by the number of people doing foundational AI work, he suggested, and a young team’s implied quota can reach $10 billion of new GDP.
14. Taxing bots solves the easy side of AI displacement
Andrew Yang argued that AI, robots and agents should bear more tax while employment becomes cheaper. At his company, he said 40%–50% of hiring expenditure never reaches the worker because of Social Security, health-care and income-tax costs.
Salim rejected taxing tokens because that is effectively a tax on cognition, potentially slowing cancer research and other valuable work; outputs and capital can be taxed instead. Dave’s simpler answer was that eliminated wages become corporate profit, which is already subject to corporate income tax.
Alex described every tax as a distortion and asked whether society truly wants to penalize superintelligence. He preferred using AI to lower living costs toward zero, with options spanning universal basic income, services, equity and computer capability rather than a narrow excise tax on intelligence.
Dave’s pushback on redistribution rhetoric was practical: collection is easy, while “who gets what” is politically and administratively hard. Alex expects golden shares, a sovereign fund or a universal dividend to gain attention around the OpenAI and Anthropic IPOs, but warned labs to specify the desired end state before inviting intervention.
15. Educated young people without ladders pose the nearer-term labor risk
Peter’s concern was not aggregate unemployment but educated 18-to-28-year-olds promised a future that no longer materializes. He cited recent graduates’ prolonged job difficulty and divergent declines among 22-to-25-year-old software engineers and customer support agents, while acknowledging that broader AI-employment data remain murky.
His historical list ran from the French and Bolshevik revolutions through Iran, Tiananmen Square, the Arab Spring and Hong Kong. Dave’s personal Iranian example supplied the warning: students wanted to end monarchy and concentrated wealth, but did not intend the ensuing 45 years of religious rule.
Salim said social media amplifies panic “like 100x,” while half the US population reportedly cannot assemble $500 for an emergency. Even if an employment apocalypse never arrives, Peter argued, a “pandemic of fear” can mobilize people who believe the social contract has failed.
Alex disputed the likely geography: the US and China possess frontier models, energy and the ability to create new work, while export-controlled Europe and other regions risk becoming dependent “vassal states.” Peter kept the horizon at two to eight years; Salim’s answer was “new ladders for meaning, agency, status, contribution,” even as Dave called this the best-ever recruiting market for AI-native startups.
16. Institutional redesign matters more than digitizing old bureaucracy
In the AMA, Salim said EXO methods must begin with a massive transformative purpose, then delete legacy processes, rebuild around an intelligence stack and only then scale. Applied superficially, the tools merely “scale pathology” and automate bad decisions: “You don’t want to digitize your bureaucracy, you want to delete it.”
He targeted horizontal functions such as IT, HR, branding and privacy, where power accumulates because saying no is easier than managing a risky yes. His provocative remedy is periodically removing and rebuilding those layers so they cannot fossilize around obsolete incentives.
David Friedberg distinguished strategic support from state ownership. The government’s 10% Intel stake and $10 billion intervention protected a threatened Ohio 1.4-nanometer project; SpaceX received a roughly $1 billion service contract after Falcon 1’s fourth-flight success, not a gratuitous equity injection.
AI labs already have capital, and the government’s Fable order proved it has more power than any board seat. Peter nevertheless favored an independently managed US sovereign fund modeled on Norway’s roughly $1.7 trillion fund, Singapore’s Temasek and GIC, or ADIA and PIF; David Friedberg added that models trained on collective data create a case for shared returns.
17. Bitcoin split the panel on productivity, settlement and AI-native money
Alex’s answer was, “I’m not anti-Bitcoin, I’m just drawn that way.” Like gold, he sees Bitcoin as nonproductive: it pays no coupon, dividend or interest, generates no ideas and does not expand humanity’s freedom of action. He made an exception for stablecoins that improve settlement and support the dollar.
His larger objection is that AI agents can design their own layer-one networks, weakening the claim that Bitcoin’s first-mover status guarantees primacy. Salim countered that agents will be “ruthlessly functional,” selecting rails for low friction and trust rather than ideology or the energy consumed by mining.
Salim argued Bitcoin originally solved decentralization and security but not scalability; the Lightning Network supplied the missing third vertex. Peter separately judged quantum attacks a real crypto risk but rejected a three-year breakage timeline, expecting Coinbase and the wider industry to deploy quantum-tolerant algorithms first.