Video · 2026-06-11 · 1h 21m · 36 moments

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

✦ AI generated

GUEST Clark Tang GUEST Brad Gerstner GUEST Andrew Fox GUEST Gavin Baker

timeline · colored by role

Claim◆

SpaceX is a must-buy, must-own, set-it-and-forget-it holding for institutional investors because compute needs and model value are being underestimated and its core space business is the best bet on the future.

Gavin Baker argues that being 'AI pilled' means believing compute needs and model value are underrated, and combined with SpaceX's core space business, it's an essential holding for institutional investors.

transcript

Gavin Baker: And I think we're all pretty AI pilled. And if you're AI pilled, that means we got to build a lot more compute than the world thinks. And that these models are going to be a lot more valuable than people think. You combine that with their core business, I don't know another entrepreneur or another business that's a better bet on the future, right, than SpaceX.

Claim◆

SpaceX is a must-buy, must-own, set-it-and-forget-it position for institutional investors who want real exposure to both the AI and space future.

Gavin Baker argues that given how 'AI-pilled' they are about future compute needs and model value, combined with SpaceX's core business, it's an essential holding for institutional investors betting on the future.

transcript

Gavin Baker: I don't know another entrepreneur or another business that's a better bet on the future, right, than SpaceX. And so I think for most institutional investors, it's a must buy, a must own, a set it and forget it, right, in order to have a real bet on both the space and the AI future.

supports · 3

Claim◆

SpaceX is a must-buy, must-own, set-it-and-forget-it position for institutional investors because it's simultaneously the best bet on the AI compute buildout and on space.

Gavin Baker frames SpaceX as a uniquely positioned 'must own' stock because it combines a massive AI compute bet with an unmatched core space business, ahead of its IPO.

transcript

extends · 1supports · 1

Claim◆

SpaceX is a must-buy, must-own, set-it-and-forget-it position for institutional investors as a bet on both the space and AI future.

Brad Gerstner argues that because compute needs and model value are both underestimated, and SpaceX's core business is unmatched, it's a must-own stock for institutions.

transcript

Brad Gerstner: I think we're all pretty AI pilled. And if you're AI pilled, that means we got to build a lot more compute than the world thinks. And that these models are going to be a lot more valuable than people think. You combine that with their core business, I don't know another entrepreneur or another business that's a better bet on the future, right, than SpaceX.

extends · 1rebuts · 1supports · 1

Claim◆

SpaceX is a must-buy, set-it-and-forget-it holding for institutional investors because it's simultaneously a bet on massively underestimated AI compute value and on the space future.

Gavin Baker opens (and later closes) the discussion by arguing SpaceX is a uniquely attractive 'set it and forget it' holding because it combines an underrated AI compute business with the best pure-play bet on space.

transcript

Gavin Baker: I think we're all pretty AI pilled. And if you're AI pilled, that means we got to build a lot more compute than the world thinks. And that these models are going to be a lot more valuable than people think. You combine that with their core business, I don't know another entrepreneur or another business that's a better bet on the future, right, than SpaceX.

supports · 2

Claim◆

The speed at which SpaceX brings terrestrial AI data centers online is one of the two most important variables in valuing the SpaceX IPO, since faster buildout directly lowers cost and increases monetization.

Gavin Baker argues that how fast SpaceX brings terrestrial AI data centers online is one of the two biggest levers for the IPO, since speed directly reduces build cost and increases monetization rates.

transcript

Gavin Baker: We do know from Jensen that uh Elon brings data centers up faster than anyone 122 days. Speed is literally cost because every day you're paying electricians and plumbers. That's cost. And they're now monetizing them at arguably the highest rate.

extends · 1supports · 1

Mechanism

The two biggest variables for SpaceX's valuation are how fast it can bring terrestrial data centers online (Elon can stand them up in 122 days, faster than anyone) and where its models sit on the intelligence-per-cost Pareto curve.

Gavin Baker identifies terrestrial data center deployment speed and the AI model Pareto curve (intelligence per dollar) as the two key levers driving SpaceX's economics.

transcript

Gavin Baker: I think the most important variable, one of the two most important, is how quickly they bring on terrestrial data centers. We do know from Jensen that Elon brings data centers up faster than anyone, 122 days. Speed is literally cost because every day you're paying electricians and plumbers. That's cost. And they're now monetizing them at arguably the highest rate.

Mechanism

The two most important variables for the SpaceX IPO are how quickly it brings terrestrial data centers online and the superior operating profit per gigawatt of its xAI compute deals with Google and Anthropic.

Gavin Baker identifies two key levers for valuing SpaceX's AI business: speed of terrestrial data center buildout and the unusually high operating profit per gigawatt generated by its xAI cloud deals with Google and Anthropic.

transcript

Gavin Baker: Clark, who I've known for many years, made a great analysis here. And he shows that XAI's deal with Google for cloud computing generates more operating profit per gigawatt than Anthropic, than Meta, than Google, than OpenAI. Their deal actually with Anthropic also generates probably more operating profit than anyone but Anthropic.

explains mechanism · 1

Mechanism

The speed at which SpaceX brings terrestrial AI data centers online — as fast as 122 days, faster than anyone else — is one of the two most important levers for its valuation, because speed directly translates into lower cost and faster monetization.

Gavin Baker identifies data-center buildout speed as one of two critical variables for the SpaceX IPO, noting Elon Musk's 122-day construction cadence turns time into direct cost savings and monetization advantage.

transcript

explains mechanism · 1

Fact

Cursor and Anthropic possess more tokens of proprietary coding data than exist on the entire public internet, and using that data to train a model produced a genuinely strong result.

Gavin Baker explains that Cursor's private coding dataset exceeds what's publicly available online, and feeding it into training (via RL and fine-tuning) produced a model that briefly topped the coding Pareto frontier.

transcript

Gavin Baker: But my understanding is that Cursor and Anthropic have more tokens of proprietary coding data than anyone else. And they have more tokens of proprietary coding data than exist on the public internet. And so they fed Cursor fed um used ChemK 0.25, used their own private data, did some RL, some supervised fine-tuning and they got a really good model.

Mechanism

Cursor and Anthropic hold more proprietary coding tokens than exist on the public internet, and training on that data plus three weeks on the Colossus 2 cluster produced a Pareto-dominant coding model, showing xAI/SpaceX AI can be a real player in coding.

Gavin explains that Cursor's proprietary coding data, fed into training plus three weeks on the Colossus 2 cluster, produced Composer 2.5, which briefly became Pareto-dominant on coding benchmarks.

transcript

Gavin Baker: Cursor fed um used ChemK 0.25, used their own private data, did some RL, some supervised fine-tuning and they got a really good model. And then they spent 3 weeks in the Colossus 2 cluster and they got a model that 12 days ago was Pareto dominant with Composer 2.5.

supports · 1

Mechanism

SpaceX's launch business, and specifically rapid reusability of Starship, is the foundational requirement that makes orbital AI compute economically attractive.

Andrew Fox explains that SpaceX's launch business, particularly achieving rapid reusability of Starship, is the crown-jewel prerequisite for making orbital data centers economically viable.

transcript

Andrew Fox: Look, I think the thing that's foundational to everything is the launch business. This is the kind of crown jewel of SpaceX. It's something that no one else really has, notably reusability. And soon rapid reusability. This is, I think, what you need to believe in to get to the economics in AI that make orbital compute something that's very economically attractive.

Mechanism

SpaceX's reusable (and soon rapidly reusable) launch business is the foundational lever that makes every other part of the bull case, including orbital AI compute, economically attractive.

Andrew Fox argues the launch business, and specifically rapid reusability of Starship, is the crown jewel underpinning the entire SpaceX bull thesis, since driving down launch cost enables everything else.

transcript

Andrew Fox: So, look, I think the thing that's foundational to everything is the launch business. This is the kind of crown jewel of SpaceX. It's something that no one else really has, notably reusability. And soon rapid reusability. This is, I think, what you need to believe in to get to the economics in AI that make orbital compute something that's very economically attractive.

explains mechanism · 1

Mechanism

Achieving rapid, full two-stage reusability of Starship is the foundational requirement for SpaceX's launch business, and by extension for making orbital AI compute economically attractive.

Andrew Fox explains that the entire SpaceX AI/orbital-compute thesis rests on the company achieving airline-like reusability of both rocket stages, contrasting it with the old throwaway rocket model.

transcript

Andrew Fox: Um so, I think rapid reusability is the main thing that we're watching for and I think most people should watch for. Um, you know, Elon talks about it a lot, but getting these rockets to fly at a cadence that's comparable to an airline, right? And and Gavin has used this analogy before, but um, the old rocket industry was kind of like, imagine boarding a plane, flying to California, getting off the plane, the plane explodes after.

Mechanism

Rapid two-stage reusability of Starship is the key variable that determines whether orbital compute becomes economically attractive.

Andrew Fox explains that achieving airline-like reusability of both Starship stages — not just the booster — is the foundational lever that unlocks the economics behind orbital AI compute.

transcript

Andrew Fox: This is, I think, what you need to believe in to get to the economics in AI that make orbital compute something that's very economically attractive. Outside of the idea that we are in shortage for power, shortage for chips. Right? Um so, I think rapid reusability is the main thing that we're watching for and I think most people should watch for.

Mechanism

SpaceX's plan to fly both Starship stages 30-50 times before retrofit will amortize vehicle costs and is the foundational lever that drives down launch costs.

Andrew Fox explains that SpaceX's goal of reusing both Starship stages dozens of times before retrofit will amortize the vehicle's cost across many flights, which is foundational to the entire bull case for SpaceX's business.

transcript

Andrew Fox: What SpaceX are ultimately trying to achieve is have a Starship fly both stages, not just the booster, 30, 40, 50 times before you have to retrofit that ship. And when you do that, you're amortizing the cost of the vehicle over many flights, right? And that's what brings the cost down significantly.

Data

Launching AI compute into orbit costs roughly $5 billion per gigawatt of CapEx, versus roughly $20-25 billion per gigawatt for terrestrial data center infrastructure — a 5x reduction on half the bill of materials.

Andrew Fox lays out the back-of-envelope math showing that once Starship achieves reusability, launching AI satellites into space could cost about $5B/gigawatt versus $20-25B/gigawatt for terrestrial infrastructure like power, cooling, and shell.

transcript

Andrew Fox: But the math you get to is it's about $5 billion per gigawatt of CapEx to put these in space. For comparison, terrestrially, talk about the switch gears, the generators, the transformers, the shell, getting the power, that today is about 25 20 to 25 billion per gigawatt. So we're talking about a 5x reduction in cost on half of your bill of materials for the data center.

explains mechanism · 2supports · 1

Data

Once SpaceX achieves rapid Starship reusability, launching AI compute into orbit will cost roughly $5 billion per gigawatt of CapEx, versus about $20-25 billion per gigawatt to build terrestrial data centers today.

Andrew Fox lays out the math showing that reusable Starship launches could put AI compute into orbit for about $5B per gigawatt of CapEx, a fraction of the ~$25B per gigawatt cost of terrestrial data centers.

transcript

Andrew Fox: Launch this compute into space. And the math that you get to before you account for things like bad GPUs, bad satellites, right, these will all be things that happen. But the math you get to is it's about $5 billion per gigawatt of CapEx to put these in space.

Data

Once Starship achieves rapid reusability, launching AI compute into orbit will cost roughly $5 billion per gigawatt of CapEx, versus $20-25 billion per gigawatt for terrestrial data center build-out today.

Using specs from Elon Musk's satellite reveal, the speaker calculates that launching AI compute into space could cost about $5B/gigawatt versus $20-25B/gigawatt terrestrially, a roughly 5x reduction on half the data center bill of materials.

transcript

Clark Tang: The math that you get to before you account for things like bad GPUs, bad satellites, right, these will all be things that happen. But the math you get to is it's about $5 billion per gigawatt of CapEx to put these in space. For comparison, terrestrially... that today is about 20 to 25 billion per gigawatt.

explains mechanism · 1

Data◆

Once two-stage rapid reusability is achieved, launching AI compute satellites into orbit will cost roughly $5 billion per gigawatt of capex, versus roughly $20-25 billion per gigawatt for the equivalent terrestrial data center infrastructure.

Andrew Fox walks through the math showing that with reusable Starship launches, putting compute satellites into orbit could cost about $5B per gigawatt versus $20-25B per gigawatt for terrestrial switchgear, transformers, shell, and power infrastructure.

transcript

Andrew Fox: The math that you get to before you account for things like bad GPUs, bad satellites, right, these will all be things that happen. But the math you get to is it's about $5 billion per gigawatt of CapEx to put these in space. For comparison, terrestrially... that today is about 25, 20 to 25 billion per gigawatt.

explains mechanism · 1

Claim

The x.ai/SpaceX AI model business, supercharged by the Cursor acquisition, is the most underappreciated source of upside in the SpaceX story, more so than the well-understood cloud compute deals.

Brad Gerstner argues that the market is underrating how much the Cursor acquisition has advanced SpaceX's frontier AI model capability, calling it the biggest overlooked source of potential upside.

transcript

Brad Gerstner: I think the thing that's getting lost is I think they've dramatically advanced their capability when it comes to building a frontier model. People outside Silicon Valley may not know, you know, Michael and the team at Cursor as well. This is an extraordinary team that he just downloaded, right, into SpaceX.

extends · 1provides context · 1

Claim◆

We genuinely don't know how intelligent the newest AI models are because no one has run them continuously long enough to properly evaluate their capabilities before the next model ships.

Gavin Baker highlights Noam Brown's point that intelligence benchmarks are becoming obsolete because no one lets a frontier model run long enough to discover its true capability ceiling before it's superseded.

transcript

Gavin Baker: Because nobody has run Mythos for a year continuously. And we may never know how smart each generation of models actually is or was, but because we don't have time to appropriately evaluate their intelligence before the next model comes out. I mean, this is a profound statement.

explains mechanism · 1supports · 1

Claim◆

We may never truly know how intelligent each generation of AI models is or was, because no one has run a model continuously for long enough (e.g., a year) to properly evaluate it before the next model replaces it.

Referencing a Noam Brown post, Gavin Baker argues that because frontier models are replaced so quickly, nobody has time to actually evaluate their full capability, meaning true intelligence levels remain unknown — an insight he calls profound.

transcript

Claim◆

Because no frontier model has ever been run continuously for a full year, we genuinely don't know how intelligent each generation of these models actually is before the next one replaces it.

Reacting to Noam Brown's post, Gavin Baker argues that because models are evaluated only briefly before being superseded, nobody actually knows the true ceiling of their intelligence — a point he calls profound.

transcript

Claim◆

No one has run a frontier model continuously for a long stretch (e.g. a full year) because each generation is superseded so quickly, so we may never actually know how intelligent these models truly are or could become.

Citing a Noam Brown post, Gavin argues that intelligence should be measured on a time/compute axis rather than snapshot benchmarks, since models are replaced so fast that none has been allowed to run long enough to reveal its true ceiling.

transcript

Gavin Baker: Nobody has run Mythos for a year continuously. And we may never know how smart each generation of models actually is or was, but because we don't have time to appropriately evaluate their intelligence before the next model comes out.

Claim◆

We may never actually know how intelligent frontier AI models are, because no one runs a model continuously long enough to properly evaluate it before the next generation replaces it.

Gavin Baker relays Noam Brown's point that intelligence evaluation can't keep pace with model releases, since nobody runs a model like Mythos continuously for a year, meaning true capability levels remain unknown.

transcript

supports · 1

Claim◆

Because no frontier model has been run continuously for a meaningful length of time, we genuinely do not know how intelligent these models actually are, and letting a model like Einstein-caliber intelligence run tirelessly for a year could already have solved intractable problems.

Gavin Baker argues that since no model has been run continuously for a long stretch, its true intelligence ceiling is unknown, illustrating with a thought experiment of an ever-focused, tireless Einstein-level intelligence running for a year.

transcript

Gavin Baker: imagine Albert Einstein had just thought about fundamental physics 24 hours a day. He doesn't have to eat, he doesn't have to sleep, he doesn't have to relax, he doesn't drink, never gets old, never has diminished intelligence, and he thought for 1 year. I mean, we might already, you know, have solved a lot of these intractable problems.

extends · 1rebuts · 1

Claim

Frontier AI models are capturing the vast majority of economic value even though open-source models may account for the majority of raw tokens consumed.

Gavin Baker argues two things can be simultaneously true: frontier models keep capturing most of AI's economic value while open-source models handle most of the actual token volume, countering the thesis that cheap open-source tokens would close the revenue gap.

transcript

Gavin Baker: Two things can be true. The majority of economic value may continue to accrue to the frontier and man has it ever accrued to the frontier thus far and for sure the first 6 months of this year, but the majority of tokens consumed in the world may be open source.

rebuts · 1

Claim

Frontier AI models capture the vast majority of economic value even though open-source models may account for the majority of tokens consumed worldwide.

Gavin Baker argues two things can be simultaneously true: frontier models keep capturing the lion's share of AI economic value even as open-source models handle most of the world's raw token volume.

transcript

Gavin Baker: The majority of economic value may continue to accrue to the frontier and man has it ever accrued to the frontier thus far and for sure the first 6 months of this year, but the majority of tokens consumed in the world may be open source.

gives example · 1rebuts · 1supports · 1

Data

Frontier AI models capture the vast majority of economic value even though open-source models may account for the majority of tokens consumed globally.

Gavin Baker argues two things can be true simultaneously: frontier models keep capturing most of the economic value generated by AI, while open-source models may still handle the bulk of raw token volume worldwide.

transcript

Gavin Baker: Yeah, I would just say two things I two things can be true. The majority of economic value may continue to accrue to the frontier and man has it ever accrued to the frontier thus far and for sure the first 6 months of this year, but the majority of tokens consumed in the world may be open source.

supports · 1

Claim

Two things can be true at once: frontier AI models keep capturing the large majority of economic value, even though open-source models may account for the majority of total tokens consumed worldwide.

Gavin Baker reframes the open-source-vs-frontier debate: revenue overwhelmingly accrues to frontier models even as open-source likely wins on raw token share, and he expects this split to persist.

transcript

Claim

The thesis that cheap open-source models would close the gap on frontier models and capture most AI economic value has been decisively wrong; frontier models have captured roughly 90% of AI revenue even though open-source models may account for the majority of raw tokens consumed.

Gavin states that despite predictions open-source models would erode frontier revenue share, frontier models have actually captured the vast majority (~90%) of economic value this year, even though open-source models may handle most raw token volume.

transcript

Gavin Baker: That has been decisively wrong. Probably more than 90%, and it may continue to be decisively wrong. Frontier might be 90% of the economic value. Open-source might be 80% of tokens.

rebuts · 3supports · 1

Data

The industry is on pace to spend roughly $1.5 trillion in AI CapEx by 2027 against only around $300 billion in combined AI lab inference revenue, raising the question of whether that math actually works.

Brad Gerstner lays out the widening gap between projected 2027 AI CapEx (~$1.5 trillion, revised up from Morgan Stanley's $950B-$1.1T) and projected inference revenue (~$300 billion), framing it as the key risk that could crash the semiconductor complex if confidence in the ROI breaks.

transcript

Brad Gerstner: we're spending 1.5 trillion of CapEx on 300 billion of inference revenue. Does that math math for you? And what would cause you, you know, to to get more nervous again about our ability to continue to make these investments? Because the second we get nervous about it, the entire semi complex is going to come down a lot.

extends · 1provides context · 1

Data

The math on massive AI capex does work out, because inference revenue and gross margins are being significantly underestimated — likely well over $200 billion in inference revenue this year at 60-70% margins.

Responding to Brad Gerstner's worry that $1.5 trillion in projected 2027 CapEx looks unsupported by ~$300 billion of inference revenue, Gavin Baker argues the revenue and margin figures are actually understated, so the spending math holds up.

transcript

Gavin Baker: I would guess they're probably a little bit higher than that. I might say 60 or 70. But, I mean, that math starts to math, and what I would just say is I think that 300 billion is low, man. I just think it's low. I think we end this year well over 200 billion in inference revenue, well over. And so, I think the math really maths.

Prediction

With inference revenue tracking well over $200 billion this year and Jensen Huang's once-outlandish forecasts proving conservative, the massive projected AI CapEx spend is justified by the underlying revenue trajectory.

Gavin Baker argues that inference revenue running well above $200 billion this year vindicates aggressive AI CapEx, noting Jensen Huang's supposedly bullish forecasts have actually understated real growth.

transcript

Gavin Baker: I think we end this year well over 200 billion in inference revenue, well over. And so, I think the math really maths, and I do think we have to give Jensen some credit because he said some things that seemed outlandish. He said a trillion two years ago. And I mean, he was really low.

Prediction

AI inference revenue is on a trajectory toward $300-400 billion by 2028 and potentially trillions of dollars by 2030, which justifies the industry's massive CapEx spending.

Brad Gerstner points to Dario Amodei's and Jensen Huang's revenue forecasts to argue that inference revenue is scaling fast enough to justify the roughly $1.5 trillion in projected annual AI CapEx by 2027.

transcript

Brad Gerstner: Dario did the podcast with Dwarkesh when he was talking about country geniuses in the data center. He said that will be here by 2028. He said revenues will go into the low hundreds of billions by 2028. So, let's call that, you know, 3 400 billion of revenue by 2028.

supports · 1

Highlight slides

Related episodes

Video · All-In Podcast

OpenAI vs Anthropic IPOs, Anthropic $3T, Zuck's Price War, China Ends Open Source?, Trump Accounts

→

Video · Machine Learning Street Talk

Intelligence is collective, not artificial — Prof. Michael I. Jordan (UC Berkeley / Inria)

→

Video · The TWIML AI Podcast

Teaching Computers to Smell | Alex Wiltschko

→