By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

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

Brad Gerstner sits down with Gavin Baker and Andrew Fox of Atreides Management, alongside Altimeter partner Clark Tang, to break down one of the biggest questions in tech and markets: how should investors think about the SpaceX IPO?

Brad GerstnerhostGavin BakerguestClark Tangguest

Jun 11, 20261h 20mWatch on YouTube

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

Great conversations run long. Their best ideas don't have to.

uListen is free on iPhone, iPad, Android, Chrome, and Claude.

CHAPTERS

Jump to publisher timestamps

Chapter summaries

SpaceX IPO framing: why AI compute changes the investment narrative
Brad Gerstner opens by positioning SpaceX as a combined bet on space and AI, then sets the agenda: the SpaceX IPO, recent model launches, and hardware/market takeaways. He frames the discussion around understanding the business by breaking it into clear levers rather than hype.
- •Brad says being “AI pilled” implies far more compute will be built and models will be more valuable than many expect
- •Brad introduces Gavin Baker, Andrew Fox, and Clark Tang and sets up the main questions: IPO levers, model race, and Taiwan/compute bottlenecks
- •Brad notes the IPO is imminent and asks for a simplified breakdown of what matters most over the next few years
Two key IPO levers: terrestrial data-center speed and model position on the Pareto frontier
Gavin Baker identifies two variables he thinks investors should focus on: how fast SpaceX can bring on terrestrial data centers, and where their models land on cost-vs-intelligence curves. He argues speed is central because it drives both cost and monetization.
- •Gavin highlights analysis suggesting xAI’s compute deals generate unusually high operating profit per gigawatt compared to other AI labs
- •Gavin cites a calculated 55% IRR on a data-center build and contrasts it with borrowing costs to emphasize ROI
- •Gavin stresses “speed is cost,” referencing rapid data-center deployment timelines and high monetization rates
- •Gavin explains the Pareto curve concept (intelligence per unit cost) and argues frontier revenue accrues to the curve
Cursor’s data advantage and why coding may be a strategic wedge for xAI
The conversation turns to why Cursor and proprietary coding data could materially improve xAI’s competitiveness. Gavin explains how private data, fine-tuning, and access to large training clusters can move models onto the frontier quickly.
- •Gavin says Cursor and Anthropic may have proprietary coding tokens exceeding what’s available on the public internet
- •He describes how Cursor used private data, reinforcement learning, and fine-tuning to produce strong results on its own benchmarks
- •Gavin suggests that with enough compute, rapid scaling on the Pareto curve is possible once a model is competitive
- •He ties coding to broader progress, relaying Amjad Masad’s view that coding may be a fast path to more general intelligence
SpaceX core foundation: launch economics and the challenge of rapid reusability
Brad asks to zoom back to the “core business,” and Clark argues launch is the foundation that enables everything else. The group centers on the importance—and difficulty—of making both stages of Starship reusable at high cadence.
- •Clark calls launch the crown jewel and emphasizes reusability as the differentiator
- •He explains the goal of airline-like cadence and many re-flights per vehicle to amortize cost
- •Gavin interjects that rapid reusability is a very hard problem, even if SpaceX has succeeded at other hard things
- •Clark references efforts to recover the second stage and then refly it as a stepping stone to higher cadence
Starlink growth: broadband scale, direct-to-cell, and the “better, faster, cheaper” thesis
Brad presses on whether connectivity revenues can scale dramatically, and Clark outlines why Starlink penetration is still early. Gavin adds a user-centric view: Starlink’s quality and potential cost structure could support large adoption if launch economics improve.
- •Clark says Starlink’s global household penetration is still under 1% and argues broadband could scale substantially
- •Brad discusses direct-to-cell connectivity as a major driver of expected growth
- •Gavin says Starlink has been the best connection for him in many locations and expects cost advantages with rapid reusability
- •The group frames ambitious revenue growth as small penetration of a very large global telecom market
“Elon Web Services”: surprise emergence as a major AI compute provider
Brad and the group describe how SpaceX became a meaningful AI compute seller through large deals, changing forecasts quickly. They emphasize execution speed, supply relationships, and the ability to monetize compute at premium rates.
- •Brad says the market wasn’t modeling SpaceX as an AI compute player until recent deals changed expectations
- •A speaker describes visiting a data-center site and emphasizes the engineering depth behind rapid deployment
- •The discussion highlights supplier relationships, financing advantages, and credibility that compounds over time
- •Gavin argues the premium is mainly about speed, coherence, and availability, with an additional “call option” value tied to future space compute
Data centers aren’t commodities: first-principles design and behind-the-meter execution
Gavin pushes back on the idea that data centers are commoditized, arguing SpaceX has differentiated design and execution. The group stresses that engineering competence and speed influence supplier behavior and economics across the ecosystem.
- •Gavin says he does not believe data centers are commodities and compares SpaceX’s approach to first-principles rethinking in rockets and electric cars
- •He suggests SpaceX’s data-center design choices are more differentiated than outsiders assume
- •Brad and Gavin discuss behind-the-meter complexity and the limited number of players who can execute reliably
- •Gavin argues suppliers prefer the buyer who can energize GPUs fastest because “speed is money” for everyone involved
Terrestrial vs. orbital compute: why orbital isn’t required for the IPO but could be transformative
Brad asks whether orbital data centers are essential to the IPO case. Gavin and Clark argue the terrestrial AI business can justify excitement on its own, while orbital compute is an important upside option tied to two-stage rapid reusability.
- •Gavin says orbital compute isn’t necessary for the IPO valuation if terrestrial compute, Starlink, and coding progress hit expectations
- •Clark estimates implied monetization rates per gigawatt and contrasts them with reported deal terms to support the terrestrial case
- •Brad reframes orbital compute as a potential additional lever that could further lower costs and expand capacity
- •The discussion repeatedly returns to the gating factor: achieving reliable two-stage and rapid reusability
Orbital compute math: cost per kilogram, megawatts per launch, and $/gigawatt comparisons
Clark lays out a first-principles economic argument for orbital compute once launch costs fall enough. Gavin adds a comparison between today’s terrestrial build costs and a future orbital alternative, while noting reliability and maintenance risks.
- •Clark explains two-stage reusability drives down cost per kilogram and moves toward fuel-dominated cost over time
- •He references a design approach for AI satellites and estimates capacity per Starship launch in megawatts
- •Clark estimates about $5B per gigawatt of CapEx to place compute in space versus roughly $20–$25B per gigawatt for non-silicon terrestrial infrastructure
- •Gavin adds a broader bill-of-materials framing and cautions reliability/maintenance could change the economics
xAI model upside: Cursor acquisition, compute access, and potential frontier acceleration
Brad argues the model story may be underappreciated relative to Starlink and compute resale. Gavin describes why model base quality plus proprietary data injection and abundant compute could produce rapid gains, and the group notes xAI was already on the frontier in recent snapshots.
- •Brad says Cursor brought a strong team and revenue momentum, and that compute constraints could be relieved by integration with SpaceX/xAI compute
- •Gavin discusses expectations around a larger parameter Grok model as a stronger base model and the impact of injecting Cursor data earlier than RL-only approaches
- •Brad argues compute monetization creates a “call option” to redirect capacity in-house for training and inference if model opportunity rises
- •Gavin notes xAI was on the Pareto frontier recently and that the number of frontier players is small
Bull vs. bear on 8X revenue and how to think about post-IPO drawdowns and lockups
Brad outlines the skeptical view: projecting massive revenue expansion in a short window is historically rare. Gavin and Brad discuss portfolio sizing, the likelihood of volatility, and why this IPO’s liquidity dynamics may differ from typical cases.
- •Brad frames the bear case around the difficulty of scaling revenue multiple times in a few years and investor discomfort with valuation
- •Brad explains his firm’s approach: take a base “set it and forget it” position and adjust around market moves
- •Gavin calls the IPO situation unprecedented in size and index implications, and says he doesn’t know the short-term trading outcome
- •Gavin argues employee/investor liquidity has existed for years via recurring windows, which may reduce forced selling pressure
Fable 5, Mythos, and why snapshot benchmarks are breaking as models run longer
Brad shifts to recent model releases and their implications for the superintelligence race. Gavin argues long-running task performance changes how we measure intelligence and increases the importance of compute, because models can think continuously without human limitations.
- •Brad describes Fable 5 as Mythos with safeguards and fallback behavior and highlights commentary emphasizing long-running tasks
- •Gavin says it’s hard to argue Anthropic isn’t “up,” citing releases and traction
- •Gavin relays the idea that we may not know how smart models are because there isn’t enough time to evaluate before the next release
- •He compares human attention limits to models that can run continuously, arguing this makes him more bullish on compute demand
Frontier vs. open source: value accrues at the frontier while tokens may route to cheaper models
The group revisits a prior debate about open source closing the gap. They argue revenue has concentrated in frontier models so far, but model routing and fine-tuned open source can still dominate token volume—while being bullish for compute providers overall.
- •A speaker argues revenue has empirically accrued to the frontier because each frontier step unlocks new use cases
- •The same speaker describes multi-agent orchestration and large-context reasoning as newly practical capabilities
- •Gavin says two things can be true: frontier captures most economic value while open source captures most tokens via routing
- •Gavin cites an example where proprietary data plus fine-tuning and routing improved outcomes at lower cost, while still consuming some frontier tokens
Taiwan and the ASIC landscape: nuance beyond “Nvidia vs. ASICs,” plus execution as the differentiator
The discussion turns to hardware dynamics and takeaways from time in Taiwan. The group argues the narrative is shifting from a binary battle to workload-specific accelerators, while Nvidia’s execution and efficiency per watt remain central in a power-constrained world.
- •A speaker says discussion in Taiwan included excitement around new ASIC efforts and more workload-custom designs
- •They argue expectations of Nvidia share loss have not played out as strongly as predicted
- •Gavin notes that what’s “on paper” in deployment plans may differ from what is actually installed, and is skeptical of low Nvidia share outcomes
- •Gavin emphasizes tokens-per-watt as a revenue driver and says watt constraints favor the most efficient solutions
AI CapEx math: $1.5T spending versus AI revenues, margins, and demand-driven pricing strength
Brad asks whether massive CapEx can be justified by inference revenue levels. Gavin and others argue margins and accelerating demand make the economics more plausible than skeptics claim, with monetization per gigawatt rising due to supply constraints and new agentic workloads.
- •Brad contrasts estimated CapEx levels with projected inference revenues and asks what would make the group more nervous
- •Gavin asks about gross margins, suggests they may be higher than assumed, and argues revenue projections may be too low
- •A speaker says monetization per watt appears to be increasing because demand is outstripping supply, reversing deflation assumptions
- •The group discusses how higher willingness-to-pay and fixed-cost leverage could accelerate profitability even if labs want to reinvest heavily
Market check and closing: positioning, seasonality, and why they remain optimistic despite risks
Brad and Gavin close with a candid portfolio/market discussion: semis have surged, dispersion is high, and macro/geopolitical risks remain. Even with a possible consolidation and seasonal slowdowns, they argue model capability gains make it hard to become too bearish long term.
- •Brad says his firm reduced exposure from “large” to more moderate due to valuation, geopolitics, inflation concerns, and the need for consolidation
- •Gavin uses a runner/cliff analogy to describe how sharply some stocks have moved and why a rest period may be needed
- •They discuss summer seasonality in token consumption and the possibility of shifting share to cheaper open-source tokens
- •Brad ends by describing rapid revenue scaling among major tech firms and a belief that AI will transform a meaningful share of global GDP, followed by final thanks and wrap-up remarks

Publisher timestamps

Get more out of YouTube videos.

Read along on your phone, search it beside the video, or ask your favorite AI app what was said.