Public transcript

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

By: Bg2 Pod
Length: 1 h 21 min
Language: English
Text from: the video's captions
Transcribed: 28 September 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

Brad Gerstner hosts Gavin Baker, Andrew Fox, and Clark Tang to break down the SpaceX IPO two days before it prices at $135 a share for a $1.77 trillion valuation. They walk through the key levers: rapid Starship reusability, Starlink direct-to-cell growth, the new AI compute business (now the number four hyperscaler after the Anthropic and Google deals), the Cursor acquisition's potential for frontier models, and the call option on orbital data centers. They then cover the Fable 5/Mythos launch and long-running agents, the open-source versus frontier token debate, Nvidia versus ASICs, and whether 1.5 trillion dollars of 2027 CapEx can be justified by inference revenue. The group closes with a market check, having both dialed back exposure to medium-small after the semis rally.

Key points

The SpaceX IPO prices at $135 a share for a $1.77 trillion valuation, with banks forecasting $160 billion of revenue in 2028.
Gavin Baker identifies two key levers: how quickly SpaceX brings on terrestrial data centers, and the performance of their models on the Pareto curve.
Clark Tang's analysis shows xAI's Google cloud deal generates more operating profit per gigawatt than Anthropic, Meta, Google, or OpenAI.
Orbital compute could cost roughly $5 billion per gigawatt in launch CapEx versus $20-25 billion per gigawatt terrestrially, but requires rapid two-stage Starship reusability.
Anthropic's Fable 5 and Mythos excel at long-running tasks, prompting Noam Brown's argument that snapshot benchmarks should be replaced by time, tokens, or compute on the x-axis.
Frontier models have captured roughly 90% of AI revenue despite open-source models consuming the majority of tokens.
Morgan Stanley raised its 2027 AI CapEx forecast from $950 billion to $1.1 trillion, and the speakers estimate the true number is closer to $1.5 trillion against about $300 billion of inference revenue.
Both Altimeter and Gavin Baker have dialed back from large to medium-small exposure, expecting a possible consolidation before higher highs.

Questions it answers

00:00SpaceX IPO framing

How should investors think about the SpaceX IPO?

They frame it as a must-own bet on the space and AI future, with the IPO at $135 a share and a $1.77 trillion valuation, and banks forecasting $160 billion of revenue in 2028.

IPO is 2 days away at $1.77 trillion valuation
Banks forecast $160 billion revenue in 2028
Seen as a set-it-and-forget-it bet on space and AI

02:42Key levers and data center economics

What are the big variables in the SpaceX story?

Gavin Baker points to two levers: how quickly SpaceX brings on terrestrial data centers, and where their models sit on the Pareto curve of intelligence per cost. Clark's chart shows xAI's Google deal generates more operating profit per gigawatt than Anthropic, Meta, Google, or OpenAI.

Terrestrial data center ramp is a massive variable
xAI monetizes compute at best-in-class rates per gigawatt
Cursor's proprietary coding data may enable a frontier coding model

07:10Launch and Starlink

What has to go right in launch and Starlink?

Rapid two-stage reusability of Starship is foundational; launch cadence is expected to go from ~160 last year to high hundreds and then thousands. Starlink broadband has under 1% household penetration and connectivity revenue is modeled to grow from ~$10 billion to $50 billion by 2028.

Second stage return planned later this year, re-flight next year
Thousands of launches means 2-3 per day
50 billion connectivity revenue is 0.3% of global telecom

13:05SpaceX as AI compute provider

How did SpaceX become a major AI compute player?

In 30 days SpaceX went from not being an AI hyperscaler to number four, passing Oracle, by signing the Anthropic and Google deals. Gavin argues data centers are not commodities and Elon's first-principles design enabled the 122-day build.

Number four hyperscaler in 30 days
Speed is money for all suppliers
Only 2-3 players can reliably engineer behind-the-meter data centers

22:00Orbital compute economics

Is orbital compute essential and what does it cost?

Clark's math shows ~$5 billion CapEx per gigawatt to launch compute into space versus $20-25 billion per gigawatt terrestrially, but it depends on rapid two-stage reusability and satellite reliability. Orbital compute is not necessary to justify the IPO valuation.

5 MW per Starship launch, ~$5 billion per gigawatt in space
Terrestrial is $20-25 billion per gigawatt
Reliability and maintenance are the key risks

29:30x.ai and Cursor

What does Cursor change for xAI?

The Cursor acquisition gives xAI a team with more proprietary coding tokens than exist on the public internet, plus massive in-house compute. Grok 4.3 is training and the speakers see the model as the biggest potential upside surprise.

Cursor could exit the year at up to $10 billion revenue
Grok 4.3 is a 1.5 trillion parameter model in training
Model capability is the least-discussed upside

43:30Fable 5 and long-running agents

What did Fable 5 and Mythos show about the frontier race?

Fable 5 and Mythos excel at long-running tasks, and Noam Brown's point is that we don't know how smart these models are because nobody has run them continuously for a year. Frontier models have captured roughly 90% of revenue while open source may hold the majority of tokens.

Benchmarks should measure time, tokens, or compute, not snapshots
Frontier captures ~90% of economic value
Open source is bearish for frontier labs but bullish for compute

59:00Nvidia and ASICs

What did Computex reveal about Nvidia vs ASICs?

Nvidia has maintained share handsomely, and Clark notes a more nuanced ASIC landscape with MediaTek's V8T versus Broadcom's V8I. Gavin suggests Nvidia could release a frontier open-source model and become a dominant cloud provider faster than people think.

Nvidia likely holds ~10 of OpenAI's 27 gigawatts on paper
Tokens per watt favors Nvidia in a watt-constrained world
OpenAI's Jalapeno chip is good but needs more cooling

Notes

SpaceX IPO

IPO priced at $135/share, a $1.77 trillion valuation; banks (Goldman Sachs, per WSJ) model $160 billion revenue in 2028 versus $18 billion last year.
Elon owns ~50% and is locked up for 365 days; employees and investors have had liquidity roughly every 6 months for 10 years, which is unprecedented.
After the recent deals, the multiple fell from ~100x trailing revenue to ~39x; SpaceX added $29 billion in a month.

Key Levers

Gavin Baker highlights two variables: speed of terrestrial data center buildout and position on the Pareto curve (intelligence per unit of cost).
Clark's chart shows xAI's Google cloud deal generates more operating profit per gigawatt than Anthropic, Meta, Google, or OpenAI; Frieda at Altimeter calculated a 55% IRR on Claude's one.
Elon stood up 100,000 GPUs in 19 days (per Jensen, "Elon is an N of 1"); speed is cost.

Core Business

Rapid two-stage reusability of Starship is foundational; second stage return is planned later this year, re-flight next year.
Launch cadence expected to go from ~160-165 last year to high hundreds, then thousands of launches (2-3 per day).
Starlink broadband has penetrated less than 1% of global households; connectivity revenue modeled to grow from ~$10 billion to $50 billion by 2028, only 0.3% of the global telecom market.

Orbital Compute

Clark's math: ~5 MW of capacity per Starship launch, ~$5 billion CapEx per gigawatt in space versus $20-25 billion per gigawatt terrestrially (a 5x reduction on half the bill of materials).
Reliability and maintenance of satellites remain the key risk; GPUs melt and lasers fail.

Models

Fable 5 (Mythos with safety classifiers, falling back to Opus 4.8) excels at long-running tasks; Noam Brown argues we don't know how smart these models are because nobody has run them continuously for a year.
Cursor brings proprietary coding data (more tokens than exist on the public internet); Composer 2.5 was Pareto dominant 12 days ago; Grok 4.3 (1.5 trillion parameters) is training.
Frontier models capture ~90% of revenue while open source may be 80% of tokens; open source is bearish for frontier labs but bullish for compute providers.

CapEx and Market

2027 CapEx forecast raised to $1.1 trillion by Morgan Stanley; speakers estimate closer to $1.5 trillion against ~$300 billion inference revenue with 60-70% gross margins.
Both firms have reduced exposure from large to medium-small, citing inflation (CPI 4.2), seasonal weakness, and the need for consolidation.

Text from the video's own captions. Summary and notes written by AI from the transcript, so check anything important against the recording.