SpaceX's Financials Leaked: Is it Worth $2TN | Meta Debuts Muse Spark: Are They Back in the AI Race?
Summary
Mythos matters because autonomy turns an old-model “rifle” into a code-scanning “machine gun.” Older models can find the same vulnerabilities with skilled steering, but Mythos can traverse entire codebases and attack at machine speed: “same thing, but quantity makes a huge difference.” The panel rejected the selloff in cybersecurity stocks; if every missed flaw will be found, spending should rise because “if the other side now have machine guns, then you’ve got to build tanks.”
Jason has stopped listening to Dario Amodei’s warnings, while David Friedberg thinks the grandiosity can be sincere and useful. Jason called Dario potentially “the second greatest founder of all time behind Elon” after five years to roughly $30 billion, yet said the repeated claims about destroyed jobs and dangerous models have become an uninspiring “boy who cries wolf” routine. David’s counter was that many doom warnings are wrong, but the concerns can still be sincerely held and economically useful.
Public software’s decisive test is whether its AI agent is good enough to sell independently. Jason Lemkin argued that incumbents are shipping token-constrained copies of what frontier products did five to six months earlier, and “a 60% product has to be free”; it cannot drive reacceleration merely by checking an AI feature box. Without chargeable agents, moats preserve trapped customers but attract nobody new—“prisoners don’t create growth”—leaving companies in a “slow death spiral” toward low-growth, IBM-like valuation.
Amazon’s $20 billion Trainium business dents NVIDIA at the margin without yet becoming a merchant-silicon rival. Most of that figure represents AWS using its own chips instead of buying NVIDIA’s, including compute supplied to Anthropic, rather than customers choosing standalone Trainium hardware. Roughly 10% of NVIDIA’s revenue is material enough to pressure a perfection-priced multiple, but with Trainium nearly sold out, NVIDIA around $194 and compute scarce, the nearer-term market remains constrained by supply rather than demand.
Muse Spark puts Meta back in the model race, and fifth place is strategically sufficient for now. The model was described as decent—stronger on work Scale AI understood a year ago and weaker on capabilities frontier labs developed since—but “if you’re fifth, you’re in the game.” The $14 billion expenditure may be justified as insurance against Meta becoming dependent on Anthropic or OpenAI, though the accompanying move toward closed source weakens Llama’s former ecosystem role.
OpenAI’s ad business looks achievable, but even a historic consumer outcome may not support the whole company. The pilot reached a $100 million annualized pace in six weeks across 600 advertisers; projections call for $2.5 billion in 2026, $11 billion in 2027, $25 billion in 2028 and $53 billion in 2029. The panel could imagine roughly $100 billion around 2030, yet “consumer alone ain’t going to be enough to feed this beast”—OpenAI may need another $100 billion-plus from enterprises.
The model war may invert internet economics, with enterprises representing two-thirds of value and consumers one-third. Anthropic retains focus, developer goodwill and a current perception lead; OpenAI brings the consumer brand, more compute and increasingly conventional enterprise sales DNA. As CIOs move from rogue departmental adoption to fixed “token-maxing” budgets, OpenAI’s packaging could become an advantage—but only if it repairs its relationship with Microsoft, the strongest top-down channel into global enterprises.
SpaceX’s proposed $2 trillion valuation assumes almost no penalty for time or execution risk. Against leaked revenue of $18.5 billion, the valuation is about 108x sales and “appears to be the most expensive IPO at scale of all time”; the reported $5 billion loss is also incomplete because xAI’s losses are included only from the acquisition date. The bull case effectively sets “the Elon discount rate” and “the Elon probability of failure rate” to zero for direct-to-cellular, space-based data centers and other future markets.
Deep dive
1. Mythos industrializes vulnerability discovery
Harry’s opening question separated the marketing debate from the capability itself: Anthropic withheld Mythos because it could autonomously discover thousands of zero-day vulnerabilities, including flaws that had sat unnoticed for years, and initially shared access only with security vendors.
Harry also laid out the skeptics’ strongest evidence: an older model, carefully queried and redirected by a human, could reproduce some of Mythos’s discoveries. Jason’s distinction was autonomy and throughput—Mythos “just kicked off on its own,” reasoned across large codebases and found weaknesses continuously: the difference between a rifle and a machine gun.
Jason’s concrete warning was Cali, reportedly bought by MyFitnessPal for roughly $100 million and breached within days, exposing 3.2 million records. The underlying Firebase database allegedly lacked authentication, but that ordinary mistake was precisely the point: AI-built applications will multiply weaknesses while AI attackers make scanning every new site with PII economically routine.
The transition could therefore get worse before it gets better. Jason recalled Anthropic’s run costing roughly $20,000 or taking only hours—he explicitly warned he might be misquoting—and argued that once the process is simplified and distributed, attackers will target everybody rather than only strategically valuable companies.
2. A cyber arms race should expand the security market
Harry found the cybersecurity-stock selloff backward. Frontier models will become part of pre-deployment code review, but companies will still need vendors to build screening frameworks, administer them and defend systems on the assumption that every overlooked weakness will eventually be found.
The changed probability matters more than the individual exploit: previously, a missed flaw might be discovered if a sophisticated attacker cared enough; now, every miss may be tested automatically. “If the other side now have machine guns, then you’ve got to build tanks.”
If Anthropic withholds Mythos for six months, Harry argued, defenders know that on “six months and one day” bad actors may begin probing their code. That deadline should accelerate cyber investment, rewarding vendors that adapt and eliminating those that do not.
Jason allowed that limited capacity or publicity might partly explain the staged release, but not enough to dismiss the capability. The panel’s synthesis was an arms race, not the obsolescence of security: better offensive automation changes what defenses must do and raises the cost of standing still.
3. Dario’s warnings divide credibility from execution
Jason’s change of mind was blunt: “I don’t buy Dario anymore.” He could still call Dario the second-greatest founder after Elon and Anthropic “the greatest grudge startup of all time,” yet five years to roughly $30 billion did not make another warning about jobs, programmers or dangerous models easier to hear.
The objection was less that every warning must be false than that repetition had exhausted its audience. “I heard you the 11th time. I heard you the 80th time,” Jason said, before landing on the customer’s demand: “Enough already. Let me just use my tokens.”
Asked what inspiration would look like, Jason contrasted Dario with messages of eventual abundance: Vinod’s claim that society will work through job losses, or Marc Andreessen’s deflation thesis. He wanted some affirmative destination—“take us to Mars”—and suspected enterprise buyers may increasingly demand workflow benefits rather than apocalyptic framing.
David Friedberg pushed back on Jason’s dismissal of the warnings. He considers predictions of 50% white-collar unemployment “beyond madness,” but said concerns such as the decision not to release GPT-2 were sincerely held, even if the warnings were also good marketing. He said that, had he met Anthropic at Series C, he might likewise have rejected the deal because the warnings sounded silly.
4. Grandiosity can be false and still compound enterprise value
Jason’s revised investing test is not whether he shares the idealism, but whether it will “motivate people enough to do something where there is economic advantage to be obtained.” He now treats grandiosity as a rallying cry that can be useful even when the literal forecast is wrong.
SpaceX supplied the analogy: a literal Mars timetable can be doubted while the 20- or 30-year vision still rallies employees toward nearer-term engineering achievements. Airbnb’s “sharing economy” and Steve Jobs’s “bicycle for the mind” likewise differed from the less romantic businesses and habits that ultimately emerged.
Harry supplied the Oppenheimer framing. Model builders may construct the “bomb,” but if companies eliminate jobs, CEOs such as Jamie Dimon will make those decisions; Dario need not perform public guilt on their behalf. The prescription was to ship methodically, remain careful and accept that the technology is not stoppable.
Harry also said he had exhausted his own evangelism: after repeated warnings, demonstrations and team redesigns, companies still unprepared by April 2026 may simply need to be marked down. “I have alerted you,” he said; he is ready to leave lagging portfolios and public incumbents behind.
5. Trainium substitutes for NVIDIA inside AWS, not across the chip market
Harry connected Mythos to Amazon’s Trainium, citing a $20 billion annualized business growing at triple digits, near sellout and Uber among major customers. Rory immediately narrowed the claim: Amazon has almost no merchant-silicon business and is not broadly shipping chips in direct competition with NVIDIA.
The mechanism is internal substitution. From a roughly $200 billion annual capex budget, Amazon may direct about half toward chips; where possible it buys its own silicon, then sells hosted training, inference and Bedrock services. Anthropic may consume Trainium through Amazon-provided compute without having declared, “Yo, I love Trainium.”
That still represents approximately $20 billion that did not go to NVIDIA—“a little less than 10%” of NVIDIA’s revenue, by Jason’s estimate. Rory stressed that 10% is material when a stock is priced to perfection, especially as hyperscalers increasingly design their own chips and can trigger multiple compression without displacing NVIDIA outright.
The counterweight is scarcity: Trainium is nearly sold out, compute remains constrained and NVIDIA around $194 did not fall after discussion of a trillion-dollar backlog. The immediate limit is wafer and chip supply, while Jason Lemkin’s account of Jensen Huang’s skill at managing partners, competitors and “frenemies” remains part of NVIDIA’s defense.
6. Anthropic can maim vibe-coding platforms without replacing them
The panel first distinguished announcement from shipment when discussing Anthropic’s apparent move toward Lovable and Replit. Eric from Bolt reportedly said everyone knew this was coming; whether the screenshots were real mattered less than the inevitability of a frontier lab moving toward the category.
An old-school venture objection would cite databases, hosting, identity, OAuth and consumer-grade support as culturally distracting work Anthropic would not maintain. Jason Lemkin’s update is that Anthropic’s development pace makes those adjacencies look like “30 days of work” attached to billions of potential revenue.
Harry framed Claude Code as directly competitive with Cursor. Anthropic need not recreate every consumer-shopping workflow in Lovable, Replit or Base44; going halfway may capture technical product teams—the category Jason Lemkin called the highest-ROI segment—and “maim” specialists without fully replacing them. “Maiming hurts.”
The threat compounds because incumbents repeatedly build versions of what Claude, Replit or Lovable offered five to six months earlier, then fall nine or twelve months behind as they economize on tokens, restrict agents and declare victory over a 60% solution.
7. A 60% agent cannot generate a growth multiple
Jason Lemkin’s core monetization test was categorical: “You can’t charge for a 60% solution.” Customers may use an adequate agent bundled into HubSpot or another suite, but they will not pay an additional $20,000, $40,000, $60,000 or $100,000 when a standalone tool performs materially better.
Large internal teams nevertheless present these agents proudly because they compare them with their own prior products, not today’s frontier. If the same demo had appeared 51 weeks earlier it might have looked excellent; in the current market, “the check-the-box feature cannot be monetized in the AI era.”
Harry translated the product test into valuation: an agent good enough to sell independently can produce revenue reacceleration and move a company into a growth bucket. Without it, even sticky software becomes a mature, mid- to high-single-digit grower requiring stock-compensation cuts, headcount reductions and “grim” cash-flow optimization.
Jason Lemkin cited Wix’s Base44 bet as one of the few attempts to exceed 60%, reaching nine-figure revenue, while Salesforce’s Agentforce remained debatable. The strategic standard is not “has AI,” but “does the work well enough that customers willingly create a separate budget?”
8. Moats preserve prisoners while financial engineering preserves nothing
Jason Calacanis rejected the comforting use of “moat.” ServiceNow’s three- to five-year contracts may retain customers, but retention does not attract new agentic revenue: “No one’s excited to cross the moat except the folks that want to breach the castle walls.” His closer was harsher—“prisoners don’t create growth.”
Harry saw value at some price: software names around eight to nine times cash flow, and Salesforce at roughly 11–12x forward earnings excluding stock compensation versus a market near 20x, could yield returns. Yet without reacceleration, the destination is an IBM-like company where “nothing magical” happens.
Harry also argued that public investors lack listed AI-native alternatives. Until Anthropic, OpenAI and several application companies float, they can only compare Salesforce’s visible margins with the “mythical” private company growing severalfold on unseen GAAP economics; that imbalance may keep legacy SaaS trapped in value territory.
Buybacks did not solve the product problem. Jason said Wix spent about $1.6 billion retiring nearly 30% of its shares around $92 before falling 23% in a week. Harry noted that Salesforce used $25 billion of debt for repurchases; Jason would have kept the cash available for a downturn: “Come to daddy. I got money.”
9. Muse Spark makes Meta credible again
Harry’s verdict on Muse Spark was “decent”: not as capable as the leaders, but good enough to ask whether Alex Wang’s first Meta Superintelligence Labs model restored competitiveness. Jason Calacanis’s answer was yes—“If you’re not in the game and then you get back in the game, that’s a win.”
Reviews suggested strength in capabilities Scale AI understood a year earlier and weakness in techniques frontier labs developed during the subsequent 12 months. That lag was understandable; importantly, Meta did not repeat the credibility problem associated with Llama 4 and can now attempt to climb from roughly fifth place.
The $14 billion expense may be existential insurance. Meta owns dominant consumer and advertising platforms and does not want to resemble Apple, dependent on Anthropic or OpenAI for a capability that “turns out” not to be a commodity; merely staying above a Kimi/open-source baseline could justify ownership.
The tradeoff is Meta’s movement toward closed source, a “bummer” for an ecosystem that needed an American open model. Still, with Meta’s ad engine cited at $243 billion and reportedly surpassing Google, the panel argued this is “not a time to retreat” from a core platform transition.
10. OpenAI’s ads are inevitable, achievable and insufficient alone
The disclosed trajectory began with 600 advertisers and a pilot reaching $100 million annualized within six weeks. OpenAI projected $2.5 billion of ad revenue in 2026, then $11 billion in 2027, $25 billion in 2028 and $53 billion in 2029.
Jason Calacanis called advertising inevitable for a consumer product at ChatGPT’s scale. Google, Meta and Amazon had already established the playbook; Harry liked that the target is operationally clear and “mathematically possible,” allowing OpenAI to assign top talent and improve the system every week.
The uncomfortable conclusion was that even roughly $100 billion by 2030 might be inadequate relative to OpenAI’s valuation and compute burn. In an approximately $1 trillion advertising market already occupied by Meta, Google, Amazon and television, Jason treated $100 billion as a plausible high-end outcome—not the whole financing answer.
Hence the larger requirement: OpenAI may need another $100 billion-plus from enterprise intelligence. “Consumer alone ain’t going to be enough to feed this beast”—the company needs more than the consumer ad business.
11. Enterprise AI may reverse the internet’s consumer economics
Harry initially found a leaked memo from OpenAI CRO and president Denise Dresser—claiming Anthropic overstated revenue, OpenAI remained ahead and had more capacity—too reminiscent of Salesforce. On reflection, he decided that conventional, confident messaging may be exactly what large enterprises want from a strategic supplier.
Harry noted that Anthropic had won developer-led adoption: “Ask your developer” often produced the answer “I want Opus.” He expects OpenAI’s more traditional enterprise DNA and selling motion to work better in 2027 than it did when developer-led adoption dominated.
Jason characterized the market as a two-way fight. Anthropic currently has a slight lead in perception and developer friendliness, while OpenAI retains the consumer business and more compute. He argued that enterprise could ultimately represent two-thirds of AI value and consumer one-third—the mirror image of the internet.
The product itself may bifurcate. Harry said younger consumers favor OpenAI’s emotionally supportive style; Jason wants the opposite in investing: “This is a stupid deal…stop, you idiot,” followed by five factual reasons. At home he wants Netflix; at work he wants cognition, clarity and harsh judgment.
12. Token scarcity moves purchasing power back to CIOs and Microsoft
Jason Calacanis called compute “the ball game.” Everyone remains constrained, nobody will blink on 2026 infrastructure investment, and the next four quarters of chip demand may be inferred from TSMC wafer starts because virtually everything produced will sell.
Scarcity will be allocated through price: plans will be throttled, products such as Sora constrained and Claude allowances reduced as providers reserve tokens for the highest effective bidder. “That’s what money is for,” Jason said—the allocation of scarce resources is the definition of economics.
Aaron Levie’s “token maxing” observation, relayed by Harry, described CIOs establishing annual dollar or token budgets and forcing departments to compete project by project. That shifts demand away from rogue developer adoption; a CIO could standardize on OpenAI for 2028 because its sales motion, packaging and governance fit a Fortune 2000 procurement process.
Jason argued that Microsoft and OpenAI “need to get some couples therapy.” Microsoft remains the dominant route into enterprises, so competitive disagreements are an indulgence OpenAI cannot afford. The related stress test was Box: Harry asked whether, if Aaron’s 10-out-of-10 understanding cannot restore 20–30% growth, there is much hope for weaker incumbents; Jason agreed and said he hoped Aaron could do it.
13. SpaceX’s $2 trillion price discounts neither time nor failure
The leaked figures showed $18.5 billion of revenue and a $5 billion loss, with the latter attributed to the xAI acquisition rather than operations. At a proposed $2 trillion valuation, SpaceX would list at roughly 108x revenue—apparently the highest IPO multiple ever attempted at comparable scale.
Jason Lemkin cautioned that acquisition accounting obscures xAI’s true run-rate loss. Results include xAI only from the closing date; if it closed on October 1, for example, the statement would capture one quarter, and the loss could ultimately be $20 billion. Until GAAP financials arrive, claims about profitability excluding xAI remain unproven.
The underlying assets still impressed: a near-monopoly in cost-effective launch and an exceptional Starlink business, with future upside from direct-to-cellular service, lower-cost rockets and data centers in space. Jason Lemkin was less convinced history will view the roughly $250 billion xAI consideration as attractive.
The valuation framework replaced “100x is crazy” with explicit assumptions. Giving every adjacent market 100% success, immediate recognition and no time-value discount can reach $2 trillion; applying, illustratively, a 15% discount rate, 70% success probability and several years to delivery produces less. “The Elon discount rate is zero.”
14. Lean headcount is a choice, but revenue per employee needs context
AppLovin’s 898 employees and roughly $4.5 million of revenue per head illustrated the new aspiration: remain small by choice. AI lets top engineers produce more and agents replace SDR work, reinforcing venture’s favorite fantasy—raise $3 million and become worth $30 billion.
Rory’s hiring test was “Would I hire them again?” Jason Lemkin sharpened it to “Would I replace them with an agent?” He expects agent-building to become dramatically easier within 18 months, just as mediocre prompts already produce sophisticated output.
His specimen was a Replit site generated in about six minutes from a short prompt about “recycled mediocre.” The product gathered context and produced the site, horror image, video, audio and connections; the lesson was that magic no longer requires a professional prompt engineer.
Rory’s caveat was that revenue per employee cannot compare businesses with different gross margins: Cursor carries heavy token costs while Salesforce generates roughly 30% operating margins. The useful comparison is longitudinal—$500,000 per employee should become $600,000, then perhaps $800,000—while AppLovin is unusual for pairing lean headcount with high margins and minimal token or capex burden.
15. Private equity must sell a complete agent into its installed base
Thoma Bravo’s retreat from growth equity looked less like a verdict on every minority investment than a return to the business generating roughly 90% of its economics: controlling software companies, applying leverage, adding acquisitions and eventually selling them. When the core is threatened, “the first rule…is you retreat to the core.”
The portfolio problem mirrors public SaaS. Mature holdings such as Coupa, Anaplan and Medallia may grow at single digits; comparable public companies trade around two to four times revenue, while many buyouts occurred near 10x with debt. Applying today’s multiple and deducting leverage could leave little equity value.
The bullish escape is genuine transformation, not consulting theater: build a “100% agent” customers will pay for. Upselling an installed base by 20%, 30% or 40% could service debt and create enterprise value without restarting growth-company customer acquisition; a 60% agent still fails Jason Lemkin’s test.
Jason Lemkin called the missed opportunity tragic. Incumbents had 10,000–150,000 reasonably happy customers, 12–18 months of model access and surplus engineers—Adobe alone had perhaps 100 strong people available—yet many still operate four-to-five-year release cycles. It is not too late, but another AI expert delivering a delayed 60% solution will not prevent the debt spiral.
16. Anthropic looks likelier to float before OpenAI
Asked for IPO order, the panel chose SpaceX, Anthropic, then OpenAI; SpaceX had apparently already filed. Anthropic’s addition of the NVIDIA CEO to its board was read less as a healthcare signal than as IPO preparation alongside the eventual need for an audit-committee chair.
Jason Calacanis’s governance requirement was simple: the CEO and CFO must be aligned, preferably in a conventional reporting structure rather than having the CFO report through president VJ, and able to give identical answers in separate “police interrogation” rooms.
Harry agreed that public disagreement or briefing against the CEO becomes untenable, but noted the value of an experienced, popular “island of stability” amid heavy executive turnover. Replacing another senior leader has organizational costs; daylight matters, but not every imperfection justifies immediate disruption.
The IPO recipe was correspondingly unglamorous: align leadership, repair Microsoft, ship ads, scale enterprise sales and exploit the compute already secured. Harry’s warning to vice presidents tempted to bypass a CEO and call the board was absolute absent fraud: the board will not replace the CEO for you—“just resign with grace.”