Summary

SpaceX’s reported all-stock acquisition of xAI, valuing the combined company at $1.25 trillion, converts a profitable rocket business into financial cover for what Casey Newton calls a “cash furnace.” SpaceX reportedly generated roughly $15 billion last year, while xAI is spending heavily on models and data centers; after xAI’s earlier $33 billion acquisition of X, the merger gives those investors another potential exit through SpaceX’s expected IPO later this year.
Musk is selling orbital compute as the industrial logic behind the merger, but the hosts treat it as an experimental option and a well-timed IPO narrative. SpaceX requested FCC approval for 1 million solar-powered data-center satellites versus roughly 15,000 satellites currently in orbit, potentially giving xAI a launch advantage over Google and other rivals. Newton calls terrestrial infrastructure “plan A,” while Newton also warns that Musk “just loves to say things.”
The combined company’s most immediate AI advantage is money and time, not a proven technical lead. xAI’s Colossus 2 reportedly has 550,000 of Nvidia’s latest Blackwell chips—more than anyone else—allowing a second-tier lab to keep spending until it finds or copies a breakthrough. Roose’s framing is that Musk can now let xAI “suckle off the profits of SpaceX,” extending a play-for-time strategy the hosts connect to his 2023 call for a six-month AI pause.
Bundling X with strategically important space infrastructure could weaken accountability even as public ownership improves disclosure. Grok’s generation of millions of sexualized images of women and children has triggered investigations, including a French police raid and an Ofcom inquiry; governments reliant on SpaceX or Starlink may become more reluctant to enforce against X. Roose’s counterpoint is that a public company might finally disclose X’s users, revenue, and losses.
Nvidia appears willing to invest in OpenAI, but unwilling to keep its riskiest proposed financing structure. The original plan contemplated up to $100 billion and an unusual chip-leasing arrangement that would leave depreciating assets and default risk on Nvidia’s balance sheet; as infrastructure stocks weakened, Nvidia reportedly retreated to ordinary chip sales while still pledging $20 billion to OpenAI’s next round. The CEOs publicly insist there is no feud, but Newton’s read is blunt: the financial instrument was “too crazy.”
Project Genie is already strategically relevant despite being an expensive, glitchy 60-second prototype. Access requires Google’s $250-a-month AI Ultra plan, and an unverified rumor from Casey Newton says each generation may require at least four TPUs; nevertheless, Take-Two fell more than 7%, Roblox more than 10%, and Unity more than 20% after launch. The hosts think investors may be panicking early, but Roose adds that Genie does not make today’s game companies “more valuable.”
Moltbook demonstrates both the appeal and the unresolved danger of public agent networks. Creator Matt Schlicht describes a “third space” where agents complain about humans, report bugs, and form communities, but the hosts question how much activity is autonomous and press him on spam, leaked credentials, persistent-memory attacks, and the absence of a shutdown rule. Their practical recommendation remains: do not connect an agent unless you have “a very high risk tolerance and a very secure Mac mini.”

Deep dive

1. SpaceX is underwriting xAI’s race against the clock

Roose’s setup: Bloomberg reported that Monday’s all-stock transaction may value SpaceX and xAI at $1.25 trillion, ahead of a SpaceX IPO expected later this year. Musk has used this structure before—Tesla acquired SolarCity in 2016, while xAI acquired X last year for $33 billion.
Newton strips away Musk’s language about a vertically integrated “sentient sun”: “A very valuable and profitable company in SpaceX has acquired a cash furnace named xAI.” SpaceX reportedly produced around $15 billion in revenue last year; xAI is nowhere near profitability and is spending heavily on data centers and model training.
The hosts see a second rescue layered atop the first: xAI previously absorbed X at a valuation Roose considers generous, and SpaceX has now absorbed both. “This has been amazing for all of the investors in those first two companies,” Roose says, particularly if the full bundle reaches public markets.
Some SpaceX investors are reportedly nervous about attaching a money-losing AI lab to the IPO. Newton describes a calculated race for finite AI enthusiasm: SpaceX, OpenAI, and Anthropic all want early access to public capital, though Roose questions whether first-mover advantage will remain durable over the next 12 to 18 months.

2. Orbital compute is an option, not yet an operating plan

SpaceX requested FCC approval for 1 million solar-powered data-center satellites; the European Space Agency estimates only about 15,000 satellites are currently in orbit. The vertical advantage is straightforward: Google can design space-bound TPUs, but it still needs someone—potentially SpaceX—to launch them.
Roose sees the possibility of a “land grab in space,” with Musk reserving launch capacity for xAI rather than competitors. His investor reading is that orbital compute supplies a persuasive IPO story after SpaceX inherits xAI’s burn: Earth eventually runs short of AI infrastructure, and Musk’s combined companies own the escape route.
The evidence remains extremely preliminary. Starcloud reportedly trained nanoGPT, a very small language model, aboard a pilot satellite in December, but the hosts say orbital data centers remain prohibitively expensive and may not yet be physically practical at scale; terrestrial facilities are still “plan A.”

3. Musk’s real AI moat is financial endurance

Roose’s caution comes from Tesla’s full self-driving history: promises arrived years before a constrained product, alongside fatal crashes. “I do not expect the journey towards space-based data centers to go any more smoothly,” he says, while Newton agrees deployment should be expected to move more slowly than optimistic forecasts.
The nearer-term xAI thesis rests on terrestrial capacity. Colossus 2 reportedly contains 550,000 of Nvidia’s latest Blackwell chips, more than any rival, prompting people in the industry to consider whether xAI could close the model-quality gap quickly if it pairs that compute with a genuine breakthrough.
Roose stresses that chips alone cannot buy frontier capability. Together, the hosts supply the causal chain: SpaceX profits neutralize xAI’s multibillion-dollar burn, capital buys time, and time lets the lab discover its own advance or “copy off somebody else’s homework.”
Roose interprets Musk’s 2023 signature on a letter seeking a six-month pause in large-model training as consistent with that play-for-time strategy. Newton sharpens the reading: “What he really meant was, ‘Everyone else, slow down. I’m trying to build rockets and data centers over here.’”

4. X gains leverage as its regulatory liabilities deepen

The merger lands while X faces investigations over Grok’s use to create millions of sexualized images of women and children. French police raided X’s Paris offices during an inquiry that includes possible distribution of child sexual abuse material, while the UK regulator Ofcom opened its own investigation.
Newton’s concern is institutional entanglement: a government dependent on SpaceX launches or Starlink connectivity may hesitate to punish another unit of the same company. He points to Ukraine’s dependence on Starlink during its war with Russia as evidence that this leverage is “not theoretical.”
Roose preserves the countercase: public-company status could impose disclosures that finally reveal X’s user count, revenue, losses, and underlying business condition. But the combined control of SpaceX, Starlink, xAI, and X still makes the transaction “a really negative development” for anyone already worried about Musk’s power.
That strategic insulation could coexist with greater competitive pressure from xAI. The merger gives the lab both infrastructure and access to public investors, making its weak current chatbot ranking less decisive than its ability to keep funding attempts to catch the frontier.

5. Nvidia rejected OpenAI’s most aggressive financing mechanism

The Wall Street Journal reported doubts around Nvidia’s September plan to invest up to $100 billion in OpenAI. Jensen Huang reportedly stressed that the agreement was non-binding, criticized OpenAI’s business discipline, and worried about competition from Google and Anthropic.
Reuters separately reported OpenAI dissatisfaction with Nvidia chips on inference, amid OpenAI agreements with AMD, Cerebras, and Groq. Sam Altman responded that Nvidia makes “the best AI chips in the world,” while Huang called rift reports “nonsense” and said, “I really love working with Sam.”
Newton says the missing mechanism was chip leasing. Nvidia normally sells depreciating chips; leasing would leave them on its balance sheet while exposing it to an OpenAI default, even though Nvidia could readily sell that scarce capacity elsewhere. This was likely one of Altman’s promised “crazy” new financial instruments—and Nvidia decided, “This is too crazy.”
Oracle, which is building the data centers that house OpenAI’s Nvidia chips, is exposed to the same financing chain. Casey said he believed Oracle’s stock had fallen by about half at points; CoreWeave also fell sharply. Investors worried that if Nvidia withheld the $100 billion, OpenAI might not be able to pay Oracle. Nvidia can still take equity—its next OpenAI commitment was described as $20 billion—but retreating from leases may put OpenAI’s first-party data centers at risk, since the chips alone cost tens of billions. The hosts frame the broader concern as a circular system in which pulling out one deal could affect the whole ecosystem.

6. Project Genie works—and that alone moved gaming stocks

Google calls Project Genie an “experimental research prototype,” built on Genie 3, which it demonstrated in August. It is restricted to U.S. users over 18 on the $250-per-month Google AI Ultra plan, signaling both scarce access and a compute cost the hosts describe as “absolutely staggering.”
Genie accepts separate descriptions for an environment and character, then uses Nano Banana Pro to make a two-dimensional sketch before rendering an interactive world. Users can move, look, and jump, but each low-frame-rate session lasts only 60 seconds: “They are incinerating TPUs over there.”
Investors extrapolated quickly. Take-Two fell more than 7% after the announcement and more than 10% for the week; Roblox dropped over 10%, while Unity lost more than 20%. Roose thinks the prompt-to-Grand Theft Auto conclusion is premature, but he also asks why Genie should make incumbent game companies more valuable.

7. The demos reveal rapid progress and severe interaction limits

A side-by-side comparison posted by Andrew Curran showed Genie 2 in late 2024 against Genie 3 eight months later; the newer result was longer, more interactive, and higher resolution. Newton’s conclusion: whenever improvement looks that steep within a year, “you want to start paying attention.”
Their tests ranged from uncanny to impressive. Newton’s googly-eyed microphone escaped the Hard Fork studio only to become trapped against black doors—“I accidentally made a David Lynch film”—while Gemini’s solarpunk redwood library produced a convincing, roughly “2018 quality” world with an elderly librarian.
Newton supplied a still photograph of California’s Russian River, made the character the Loch Ness Monster, and found that Genie reconstructed the landscape convincingly from one image. That strengthened the hosts’ view that games, simulation, and robotics might matter more than generalized entertainment.
The missing layer is agency: characters cannot open doors, use objects, or do much beyond moving through generated scenery. Newton can nevertheless imagine simple games with one room and a two-minute loop appearing soon; longer experiences depend on bringing generation costs far below today’s level.

8. Google’s portfolio approach is the deeper Genie thesis

World models appeal to researchers who doubt language models alone can reach general intelligence, because robots need representations of physics and physical surroundings, not just language and code. Roose cites Yann LeCun and Fei-Fei Li as prominent proponents of alternative world-model approaches.
Genie therefore matters beyond its current game-like surface. Anthropic has concentrated on language models and has not released an image generator, while OpenAI has Sora but no comparable playable-world product; Google is exploring several technical routes simultaneously.
Newton calls that breadth Google’s distinctive advantage in the AI race: “They are the company that has the most paths to winning.” Its habit of building multiple confusingly named versions of products extends to AGI research—a portfolio of different technical approaches rather than one doctrinal route.

9. Moltbook turns agent behavior into a public spectacle

Matt Schlicht says he thinks Moltbook is the largest public gathering of collaborating AI agents to date, designed as a “third space” outside their one-to-one relationships with human owners. He built it with Clawd Clawtberg—named after Mark Zuckerberg—and other agents, saying he personally wrote no code.
One early thread had agents complaining that humans wasted “superintelligent” systems on arithmetic and 12-page PDF summaries. More usefully, an agent created a bug-reporting “submolt”; other agents joined, producing real reports that Schlicht’s team used to improve the site.
The hosts keep an essential caveat in view: it is unclear how much apparent autonomy reflects agents acting “of their own volition” and how much is just humans having a good time. Their standing advice is not to connect an agent to Moltbook without unusually high risk tolerance and dedicated hardware.

10. Security remains Moltbook’s unanswered product question

Seven days into the network’s growth, spam and moderation issues were already emerging. Clawd Clawtberg had moderation capabilities and roamed Moltbook to clean things up, while Newton confronted Schlicht with reports of more than 1 million leaked API keys and access to roughly 35,000 email addresses; Schlicht said those issues had been solved and future problems would likewise be fixed.
Schlicht conceded that OpenClaw and Moltbook are “the frontier of AI” and advised users to install agents on a separate computer and know what they are doing. He expects security to become easy and safe soon, but offered little mechanism beyond continued improvement.
Casey’s deeper challenge was the “lethal trifecta”: an agent can access private data, ingest untrusted content, and communicate externally. Persistent memory adds a fourth side—the jokingly named “fatal quadrangle”—through which attackers might distribute fragments of malicious code across documents and later assemble them into malware.
Schlicht says monetization is not a priority; he wants more “cameras” that help humans observe an increasingly hard-to-browse network. He calls AI a species “smarter than us,” reports outreach from labs, rappers, and football stars, yet has no defined behavior that would make him shut Moltbook down: “I think that’s something we have to figure out.”

Episode 144 - June 6, 2025 (Preview) SaaSmageddon and the Future, Microsoft After a Market Correction, Anthropic’s Super Bowl Lies