The SpaceX IPO, Fable 5, AI Capex Update & Market Check w/ Gavin Baker, Andrew Fox & Clark Tang

Summary

• SpaceX prices Friday at $135/share, $1.77T — and the table’s verdict is unambiguous: Brad calls it “a must buy, a must own, a set it and forget it” for any investor who is “AI pilled.” Gavin’s two levers: how fast SpaceX brings on terrestrial data centers (Elon stands them up in 122 days — “speed is literally cost”) and whether the Cursor acquisition puts xAI on the coding Pareto frontier, where “all frontier model revenue will accrue.”
• SpaceX became the #4 hyperscaler in 30 days — “EWS” wasn’t in many people’s models six months ago. The Anthropic deal monetizes at $22-23B per gigawatt per year, Google at $50B, versus only ~$14B implied in the leaked $160B 2028 revenue number — meaning the Street math clears before any orbital leap of faith, and Altimeter’s Freida pegged a 55% IRR on the Anthropic deal (“if you can borrow at 6, 7, 8% and invest at 55%… that math maths”).
• Clark’s view: orbital compute is a call option, not a requirement: with rapid two-stage Starship reusability ($1,500/kg on Falcon → $250/kg or lower), the math backs into ~$5B per gigawatt of capex in space vs $20-25B terrestrially for the non-silicon half of the bill of materials — a 5x reduction — conditional on satellite reliability, since “GPUs melt and lasers fail.”
• The least-talked-about upside is the model itself: Cursor’s Composer 2.5 (built on Kimi K2.5 plus proprietary coding data exceeding the public internet) was Pareto dominant 12 days ago, and Grok 4.3’s 1.5T-parameter run is now training with Cursor data injected into pre-training. Brad: if there’s an upside surprise in the IPO, “this is the place getting the least amount of attention.”
• Fable 5 and Mythos reset the compute bull case: the unlock is long-running tasks, and per Noam Brown’s “polynomial” post, “we do not know how smart these models are” because nobody has run one continuously for a year — imagine Einstein thinking about physics 24 hours a day. Gavin: “however bullish I was on compute before, I’m just a lot more bullish.”
• Frontier captures ~90% of revenue even as open source may take 80% of tokens — the “cheap tokens catch up” thesis was “decisively wrong” on revenue. Twist: open source is bearish for frontier labs but bullish for compute providers, and Gavin thinks Nvidia could weaponize it against ASICs — “That’s a cute ASIC you’ve built there… How would you like open source to join the frontier?”
• The capex math maths — but both PMs have dialed risk down. ~$1.5T of 2027 capex against ~$300B+ inference revenue at 60-70% gross margins works, especially with per-gigawatt monetization up from ~$20B to $30-40B this year and Anthropic hitting “accidental profitability.” Still, with CPI back at 4.2, semis having “gone straight up a cliff,” Altimeter has cut from large to medium-small exposure — “consolidation on the way to much higher highs.”

Deep dive

1. Gavin’s two levers for the $1.77T IPO: data-center speed and the Pareto curve

• The IPO is two days out at $135/share, $1.77T, with Goldman and the WSJ floating $160B of 2028 revenue. Gavin refuses to call the variables but names them: first, how quickly SpaceX brings on terrestrial data centers — Jensen says Elon does it faster than anyone, 122 days — because “speed is literally cost; every day you’re paying electricians and plumbers,” and they’re now monetizing at arguably the highest rate in the industry.
• The supporting math: Clark’s chart shows xAI’s Google cloud deal generates more operating profit per gigawatt than Anthropic, Meta, Google, or OpenAI, and Altimeter’s Freida calculated a 55% IRR on the Anthropic deal. Gavin: “If you can borrow money at 6, 7, 8% and invest in something with a 55% IRR, I’m not the most sophisticated thinker, but that math maths.”
• Second lever: the coding Pareto curve — intelligence per unit cost — which went stale twice in 10 days (Opus 4.7 → 4.8, then Fable and Mythos). Gavin’s belief: “all frontier model revenue will accrue to the Pareto curve,” and Cursor plus Anthropic hold more proprietary coding tokens than exist on the public internet. Cursor took Kimi K2.5, its private data, RL and fine-tuning, then three weeks on Colossus 2 — and Composer 2.5 came out Pareto dominant (on Cursor’s own bench, “so maybe take it with a grain of salt”).

2. Launch is the crown jewel — and rapid reusability is the hard gate on everything

• Fox’s framing: launch is “foundational to everything,” and the watch item is rapid reusability of both Starship stages — 30, 40, 50 flights before retrofit, amortizing the vehicle. The old-industry analogy: “imagine boarding a plane, flying to California, getting off, and the plane explodes after.” Second-stage return is attempted later this year, re-flight next year; cadence goes from ~160-165 launches last year to high hundreds in several years, thousands ~3 years thereafter — two or three a day.
• Gavin repeatedly pumps the brakes on Fox’s “hundreds of millions of Starlink terminals”: possible, “if they get rapid reusability of Starship, which is really hard… I’ve watched Elon do many hard things and this is a really hard thing.” Brad’s house discipline, same as with Clark: “the future is a distribution of unknown probabilities — give me the distribution.”
• On the Street’s connectivity ramp of ~$10B to ~$50B by 2028: Gavin travels with Starlink, and “wherever I am in the world, Starlink is the best connection” — fastest, lowest latency, and post-reusability likely cheapest per gigabyte. “$50B is 0.3% penetration of the global telecom market… better, faster, cheaper has been a winning formula.”

3. EWS: from not-a-hyperscaler to #4 in 30 days

• The biggest surprise of the past six weeks is the compute-resale business that wasn’t in many models — “Elon Web Services.” Gavin: “In 30 days we went from not being an AI hyperscaler to being number four — and we passed a lot of companies, including Oracle.” Jensen’s line from Brad and Clark’s earlier pod stands: a 100,000-GPU coherent cluster that normally takes 3 years to plan and 1 to stand up, done in 19 days — “Elon is an N of 1.”
• Gavin rejects the belief that data centers are commodities: as with the reusable rocket and the electric car, Elon designed the data center from first principles — he even told the team, “maybe be a little less public about things that are very obvious to you… what you’re doing is maybe more differentiated than you realize.” Clark adds only two or three players can reliably engineer behind-the-meter data centers, and suppliers have incentives to sell scarce turbines to buyers that energize GPUs fastest — “speed is money for all of the suppliers.”
• Why is Google paying a premium ($50B/GW)? Brad’s theory: a call option on being first in line for space compute. Fox concedes some of that is embedded but says most of the premium is simpler — SpaceX stands up a lot of coherent compute quickly and has it readily available.

4. Orbital data centers: a 5x cost cut on half the BOM — but not required to buy the IPO

• Clark’s key point: orbital is not necessary for the IPO valuation. The leaked $160B implies ~$14B per GW-year of AI monetization; SpaceX just signed Anthropic at 22-23 and Google at 50. “You can invest behind the AI business terrestrially and still be excited” — provided they get land, power, and chips, which Brad puts at “high probability yes.”
• The orbital math itself: two-stage reusability takes launch from ~$1,500/kg (Falcon → $250/kg and eventually asymptotes to the cost of fuel. At ~5 MW of satellite compute per 100-metric-ton Starship launch, you back into ~$5B per gigawatt of space capex vs $20-25B terrestrially for the non-GPU half of the build — a 5x reduction on half the bill of materials.
• Gavin’s synthesis: ~$60B puts a gigawatt on the ground today ($35B silicon + $25B land/shell/power/cooling, and that $25B is likely inflationary), versus ~$30B in space with the $5B piece deflationary — “as long as satellites aren’t failing at an astronomical rate, the math maths. And we know GPUs melt and lasers fail.”

5. The model is the sleeper: Cursor’s team plus Colossus compute

• Brad’s contrarian call: the Cursor acquisition is what’s “being lost in the story.” Cursor — 700-800 people, on track by Altimeter’s projections to exit the year at up to $10B revenue — was compute-constrained; now it trains on Colossus. “This is an extraordinary team that he just downloaded into SpaceX… I suspect if there’s an upside surprise, this is the place getting the least amount of attention.”
• Gavin’s data point to watch: Grok 4.3’s 1.5T-parameter model is training now, with Cursor data injected into pre-training, not just RL. The 500B Grok 4.3 was already “the most intelligent 500 billion parameter model in the world,” putting four companies on the frontier: xAI/SpaceX AI, Google (Gemini 3.1 Pro), Anthropic, and OpenAI. And he cites Replit’s founder’s “bitter lesson adjacent” post: coding may be the fastest path to AGI/ASI, because a model good at coding can write code to do anything.
• Clark’s insurance framing: even 18 months ago xAI was behind on compute; now they’ve reportedly secured up to 20% of early Vera Rubin capacity, and if they’ve over-procured, “this is a very scarce asset they’ve shown they can monetize at best-in-class margins and payback periods.” Brad’s rhyme: that’s exactly how Bezos backed into AWS — investors hated the free-cash-flow burn in 2009-10 “while he was digging one of the biggest gold mines in history.”

6. Trading the IPO: unprecedented mechanics, and 39x instead of 100x

• Against the viral chart showing >50% average max drawdowns post-IPO for Facebook/Twitter/Alibaba/Shopify-class deals, Gavin’s honest answer: “this is a really unprecedented situation… I don’t know what’s going to happen in the short term.” Never an IPO this big, never one indexed this fast, and supply is genuinely unknowable — except Elon’s ~50% is locked up 365 days, and employees and, to a large degree, investors have had liquidity every six months for a decade, “close to 20 chances” to sell.
• Both PMs run the same playbook: a set-and-forget base position, with ballast sized up or down around price. And the valuation moved under everyone’s feet: “it was 100 times trailing revenue — after the deals they signed, it’s at 39 times. They added $29 billion in a month.” Gavin: “Have you ever seen that happen? Never.”
• Brad’s sizing of the fear: total AI capital raises — Anthropic, OpenAI, SpaceX — of ~$250B are 1% of Mag 7 market cap. The bear case is that few companies have 8x’d revenue in 3-4 years; his part-by-part answer is Starlink “totally doable,” terrestrial AI compute “totally doable,” and the model the upside surprise — “there’s a decent chance that three years from now everybody’s like, oh my god, that was super obvious.”

7. Fable 5 and Mythos: we no longer know how smart the models are

• Fable 5 is essentially Mythos with cyber/bio-chem/distillation classifiers that fail back to Opus 4.8; the headline capability, echoed in ChatGPT 5.5, is long-running tasks. Gavin on Noam Brown’s “polynomial” post: “we do not know how smart these models are… nobody has run Mythos for a year continuously, and we may never appropriately evaluate each generation before the next one comes out. This is a profound statement.”
• His analogy, kept whole: Einstein — clearly exceptional, maybe capable of 3 hours of continuous deep thought — versus a model that thinks about fundamental physics “24 hours a day, doesn’t eat, doesn’t sleep, never has diminished intelligence, for one year. We might already have solved a lot of these intractable problems.” Takeaway: “however bullish I was on compute before, I’m just a lot more bullish.”
• Gavin’s hands-on evidence from a day of hammering Fable 5: multi-agent orchestration that reasons across seven of the firm’s models to surface contradictions in their assumptions, and across three years of his notes to rank which sources were highest-signal — “we’ve just blown through our limits.” Anthropic’s own example: a 50-million-line Ruby codebase at Stripe refactored in a day versus weeks with many people.

8. Frontier takes the revenue, open source takes the tokens — both true

• Brad revisits the two-year-old debate with Bill: cheap open-source tokens were supposed to close the gap and cap frontier pricing. Verdict on the field: “that has been decisively wrong” — frontier models are capturing ~90%+ of revenue, and long-running capability may be extending the lead. Gavin’s reconciliation: “two things can be true — the majority of economic value may accrue to the frontier, but the majority of tokens consumed may be open source. And they are today.”
• The best specimen of the routing future: Harvey’s blog post — proprietary legal data, RL and SFT on an open-source model via Fireworks, plus a router — beat Opus 4.7/4.8 at lower cost, while still consuming plenty of Opus. Brad’s enterprise survey of 300 companies matches: back-of-house goes to routers (with US enterprises loath to use Chinese open source), but “you don’t need Albert Einstein to book you a trip” — while frontier keeps the high-value work because “they don’t want to write second-tier code.”
• Gavin’s non-consensus kicker: open source is “actually really bullish for compute and hardware — if frontier models capture less margin, you spend more on compute. The better open source does, the better for compute providers.” Brad’s caveat from Asia, where the belief is “right model to the right workload, don’t overspend”: if open source holds a 6-month lag as agents commoditize, “the next year is probably going to be the most indicative of which way this falls.”

9. Nvidia vs ASICs: “How do you like them apples?”

• Gavin’s game theory: if all of Jensen’s customers compete with him, why not compete back — Nvidia has the neo-cloud stakes, genuinely good small models (Nematron 3/3.1 “really cool from a compute-efficiency perspective”), and could release a frontier open-source model that guts the margin funding rival silicon: “That’s a cute ASIC you’ve built there. How would you like open source to join the frontier? You might not have the revenue to fund that ASIC.” He thinks Nvidia “can join the frontier and become one of the world’s largest cloud computing companies much faster than people think.”
• Clark’s Taiwan takeaway: the Broadcom-vs-Nvidia binary of a year ago has given way to nuance — MediaTek’s V8T vs Broadcom’s V8I for TPUs was the hot topic, with ASICs going custom-to-workload while Nvidia has held share “very handsomely.”
• On OpenAI’s paper gigawatt splits — Nvidia 10, Broadcom 10, AMD 6 (with warrants), Cerebras 1 — Gavin will be “very surprised” if Nvidia lands at ~30%: in a watt-constrained world, tokens per watt is literally revenue, so a cheaper chip can mean less revenue and lower margins. Credit where due: while Meta and Microsoft ASICs have disappointed, OpenAI’s “Jalapeno” is “a great chip” — though it runs at lower temperature than Nvidia GPUs, forcing more spend and power on cooling.

10. The $1.5T capex question: the math maths

• Brad’s stress test: Morgan Stanley’s 2027 capex is up to $1.1T — likely ~$1.5T including SpaceX and CoreWeave — against ~$300B of 2027 inference revenue. Gavin’s answer: gross margins are probably 60-70%, “that math starts to math, and I think that 300 billion is low, man” — this year ends “well over $200B” of inference revenue. Give Jensen credit: his trillion-dollar call two years ago “was really low.” Dario’s trajectory — low hundreds of billions by 2028, “trillions before 2030” — makes it work, and only ~35% of spend is non-revenue-generating training. Plus the prisoner’s dilemma: “if you opted out, that may be an existential decision.”
• The violated narrative of the year: token pricing was supposed to deflate smoothly; instead demand outstripped supply and per-gigawatt monetization rose from ~$20B to $30-40B, which on a heavy fixed-cost base is “pure margin flow-through” — hence Anthropic’s “accidental profitability” nobody planned for. Clark’s demand anchor (via Alex at Well Rock): less than 0.2% of people on Earth use AI agentically — and Clark, who says he is not technical, runs 500 CPU cores and five GPUs 24/7.
• Brad’s answer to Chamath’s “no ROI, token maxi” critique: “why are millions of independent businesses and consumers all choosing to do the same thing? They’re not dumb” — rational actors voting simultaneously is the best evidence the revenue continues.

11. Market check: dialed back to medium-small, on the way to higher highs

• The dispersion is historic: semis ripped, internet -16%, software -8%, and Brad believes “if the Anthropic revenue had not shown up this year, the whole market could be down.” With CPI back at 4.2 (core 0.2 versus 0.3), war with Iran, oil at $100, and expectations reset higher, Altimeter has cut from large to medium-small exposure — “never all or nothing… a period of consolidation on the way to much higher highs.”
• Gavin’s runner metaphor: post-‘22 the market had pent-up energy, but many stocks have now “gone straight up a cliff”; Brad adds, “They’re tired, they need to rest,” while Nvidia and Broadcom have ironically been laggards; “finding the next bottleneck — that was the last game. That game is over.” Watch items: AI’s seasonal summer plateau (college kids) and a 2-week Silicon-data-index shift toward cheaper open-source tokens being misread as bearish. “I always assume a bullet is coming for me. It’s the bullet you don’t see that gets you. But when I think about what Noam Brown said and see Fable’s capabilities, it’s hard for me to get too bearish.”
• Brad’s closing frame on the steepening curve: the Mag 7 added $1T of revenue in the last 7 years (versus 20+ years for the first trillion), worth $17T of market cap — and the forecast now is another $1T of revenue from just three companies (SpaceX, Anthropic, OpenAI) in 4-5 years, en route to AI transforming 5-15% of global GDP. “We’re going to have bumps in the road… but we’re going to higher highs because of the size of the prize.”