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

Summary · 5 min read instead of 1 h 21 m

Gavin Baker and Andrew Fox work at Atreides Management. Clark Tang is a partner at Altimeter, which Gerstner leads. Baker says he is a big SpaceX shareholder, and Gerstner says Altimeter is a shareholder and expects to buy in the IPO.

Baker, Fox and Tang argue that SpaceX can grow into the revenue forecasts through rapid Starship reusability, Starlink growth and the sale of terrestrial AI compute, with orbital data centers as an extra option. Baker and Gerstner see the Cursor acquisition and xAI's model as the least-discussed source of upside. On models, the group says Fable 5 and long-running agents show that frontier models take most of the AI revenue, even if open source handles most tokens. They conclude that AI demand is outrunning compute supply, so $1.5 trillion of 2027 capex is justifiable. Gerstner says Altimeter has trimmed its position size after the run-up, citing inflation and macro risk.

Minute by minute

1. 2:22

   Two variables drive the IPO case. Baker names two key variables. One is how fast SpaceX brings up terrestrial data centers: he cites Jensen Huang's figure of 122 days and a 55% IRR that Altimeter's Freda calculated on Colossus I. The other is how Cursor data and training compute move xAI's models along the Pareto curve of intelligence per unit of cost. Baker declines to forecast either.

2. 7:03

   Rapid reusability underpins SpaceX economics. Fox calls launch the crown jewel and says rapid reusability is what makes orbital compute attractive. SpaceX aims to fly both Starship stages 30 to 50 times before a retrofit, and plans to bring back the second stage later this year and refly it next year. Fox says expectations run from about 160 to 165 launches last year to the high hundreds within several years, then thousands.

3. 10:24

   Starlink has room to grow. Fox says fewer than 1% of global households have Starlink broadband, which he thinks could reach hundreds of millions of terminals. Gerstner says bank models take connectivity revenue from about $10 billion to $50 billion by 2028. Fox calls $50 billion 0.3% of the global telecom market and says better, faster, cheaper tends to win.

4. 13:23

   SpaceX becomes an AI hyperscaler. Gerstner says the Anthropic and Google compute deals, which he calls Elon Web Services, were missing from earlier SpaceX forecasts. Tang says Musk secures power, builds sites faster than anyone and finances deals at attractive rates. Baker says SpaceX went from no AI cloud business to the number four hyperscaler in 30 days, passing Oracle. The group says Google may pay a premium partly to get first in line for space compute.

5. 19:00

   Data centers are not commodities. Baker rejects the view that data centers are commodities and says Musk designed them from first principles, as he did rockets and cars. He says this is how SpaceX built in 122 days. Tang adds that only two or three players can reliably engineer behind-the-meter data centers. Baker says speed pays everyone in the supply chain, from power to turbines.

6. 23:39

   Orbital compute is an upside option. Tang says the leaked $160 billion figure implies about $14 billion per gigawatt per year, against Anthropic at 22 to 23 and Google at 50. So he and Baker see no need for orbital compute to justify the IPO. For orbit, Tang estimates about $5 billion per gigawatt of capex, against $20 to $25 billion for terrestrial shell, power and cooling. This assumes cost per kg falls from about $1,500 to $250 or less, and that failure rates stay manageable.

7. 29:19

   Cursor may be the hidden upside. Gerstner calls the model, after the Cursor acquisition, the least-discussed source of possible upside. Baker says Composer 2.5, built on Kimi K2.5, was Pareto-dominant 12 days earlier, and that the 1.5 trillion parameter Grok 4.3 is training. He expects Cursor data to enter pre-training, not just reinforcement learning. Tang says SpaceX has secured maybe up to 20% of Vera Rubin capacity.

8. 34:54

   Can SpaceX grow 8x in four years. Gerstner describes the bear case: revenue of about $18 billion growing to the banks' $160 billion in three to four years. He finds each part of the growth plausible, and the model an upside surprise. Baker says investors should run their own numbers on each lever. Gerstner calls SpaceX a must-own for institutions, though he says he is not pumping the $1.77 trillion IPO.

9. 36:45

   Post-IPO drawdowns are uncertain. Gerstner cites a chart of post-IPO drawdowns above 50% for companies like Facebook and Shopify. Baker says no IPO this large has entered an index this fast, so selling pressure is unknown. Musk owns about 50% and is locked up for 365 days. Baker says employees and investors have had liquidity about every six months for 10 years. Gerstner says Altimeter holds a base position and adjusts its size.

10. 42:37

    Long-running models change how we measure AI. Gerstner says Fable 5 is Mythos with safeguards that fall back to Opus 4.8 when triggered. He cites Noam Brown's view that snapshot benchmarks matter less than time or compute. Baker says nobody has run Mythos continuously for a year, so no one knows how smart the models are. He compares this to an Einstein who never tires, and says he is more bullish on compute.

11. 48:29

    Fable 5 enables multi-agent work. Tang says Fable 5 handles multi-agent orchestration far better than Opus 4.8. He fed it seven company models and asked it to find contradictions in his assumptions, and it also reasoned across three years of his notes. Baker says Anthropic showed a 50 million line Ruby code base at Stripe refactored in a day. Gerstner concludes that long-running agents mean more token use.

12. 50:58

    Frontier takes value, open source takes tokens. Baker says the frontier may hold most of the economic value while open source handles most tokens. His example is Harvey, which used a router and a fine-tuned open model to beat Opus at lower cost. Gerstner says frontier labs now take about 90% of revenue, contradicting the debate two years ago. Tang argues better open source is bullish for compute providers. Baker says Nvidia could become the leading provider of open source models.

13. 58:42

    Nvidia holds share against ASICs. Tang says Nvidia kept its share over the last few years, and would have gained it if Anthropic were excluded. He says the ASIC argument has shifted to custom accelerators for specific workloads. Baker doubts that the gigawatts announced on paper will be deployed as planned. He says that when power is the limit, tokens per watt favors Nvidia, and he calls OpenAI's chip a good ASIC.

14. 1:03:22

    $1.5 trillion of capex can pay back. Gerstner cites Morgan Stanley raising its 2027 capex forecast to $1.1 trillion, and he expects closer to $1.5 trillion against about $300 billion of lab revenue. Baker thinks that revenue figure is low, and says inference revenue will be well over $200 billion this year. A guest says monetization per gigawatt rose from about $20 billion to $30-40 billion this year. Tang says rational buyers choosing to spend show real ROI.

15. 1:12:02

    Altimeter trims after a strong run. Gerstner says semis doubled and tripled this year while internet fell 16% and software 8%. He cut Altimeter's exposure from large to medium-small, citing CPI at 4.2, geopolitics and summer seasonality, though he expects higher highs. Baker likens stocks to runners that climbed a cliff and need rest. Gerstner closes by noting a forecast of another trillion in revenue from three companies in four to five years.

Published Jun 11, 2026.