This conversation features Gavin Baker from Atreides Management and Andrew Fox, joined by host Brad Gerstner and partner Clark Tang from Altimeter Capital, discussing the SpaceX IPO and AI market dynamics. Baker is a major SpaceX shareholder while Altimeter also holds positions and expects to participate in the IPO.

The discussion centers on SpaceX's $1.77 trillion IPO valuation and Goldman Sachs' projection of $160 billion revenue by 2028, breaking down three major business lines: Starlink connectivity, terrestrial AI compute infrastructure, and the xAI model business enhanced by the Cursor acquisition.

The conversation explores the frontier AI model race following Anthropic's Fable-5 launch, long-running agent capabilities, and the broader AI infrastructure buildout requiring over $1 trillion in annual CapEx by 2027. The team also examines NVIDIA's continued dominance despite ASIC competition and the revenue dynamics between frontier and open source models.

SpaceX IPO: Three Business Lines Drive $160B Revenue Target

SpaceX IPO priced at $135 per share for $1.77 trillion valuation, with Goldman Sachs projecting $160 billion revenue by 2028 across three major business lines

Starlink connectivity business could 5x from $10 billion to $50 billion by 2028, representing just 0.3% penetration of global telecom market with "better, faster, cheaper" positioning

Launch business targeting thousands of launches within 3 years, requiring rapid Starship reusability where both stages fly 30-50 times before retrofit to achieve airline-like economics

Direct-to-cell connectivity enabled by Starlink V3 satellites could reach hundreds of millions of terminals, though dependent on achieving rapid two-stage reusability

Elon's 122-Day Data Center Advantage Creates AI Compute Empire

Elon builds data centers in 122 days versus industry standard 3+ years for planning and deployment - "Speed is literally cost" because daily electrician and plumber costs add up - Gavin

Google deal monetizes at $50 billion per gigawatt while Anthropic deal at $22-23 billion per gigawatt, both generating higher operating profit than Meta, Google, or OpenAI's own infrastructure

SpaceX became #4 hyperscaler in 30 days, passing Oracle and multiple NeoCloud competitors after signing major AI compute deals with Google and Anthropic

Data centers are not commodities - Elon "re-engineered from first principles" like he did with rockets and electric cars, creating fundamental design advantages others cannot easily replicate

Cursor Acquisition Transforms xAI Into Frontier Coding Leader

Cursor's 700-800 person team brought proprietary coding data exceeding what exists on public internet, plus revenue projections up to $10 billion for 2024 exit

Composer 2.5 was Pareto dominant for coding 12 days ago using Cursor's private data with reinforcement learning, trained on Colossus 2 cluster in just 3 weeks

Grok 4.3 (1.5 trillion parameters) now training with Cursor data injected into pre-training process, not just reinforcement learning, potentially creating superior base model

"Once you are at multiple places on that Pareto curve, if you have compute, you can scale really rapidly" - revenue could surprise significantly upward - Gavin

Orbital Compute: $30B vs $60B Per Gigawatt Cost Advantage

Orbital data centers cost $30 billion per gigawatt versus $60 billion terrestrially - $35 billion for GPUs/silicon, $25 billion for land/power/cooling that becomes nearly free in space

Two-stage Starship reusability reduces launch costs from $1,500 per kg (Falcon) to $250 per kg, with 5 megawatts capacity per 100-metric-ton Starship launch

Space provides free power (solar), free cooling (radiation), unlimited land (space), and lower operating costs, requiring only satellite reliability to match terrestrial GPU failure rates

Orbital compute not necessary for IPO valuation but provides massive call option - terrestrial monetization at current rates already supports leaked revenue projections

Long-Running AI Agents Redefine Intelligence Measurement

Fable-5 and ChatGPT o1 demonstrate test-time compute capabilities where models can think continuously for hours or days, solving previously impossible long-running tasks

"We do not know how smart these models are" because "nobody has run Mythos for a year continuously" - Noam Brown's insight that snapshot benchmarks are becoming irrelevant - Gavin

Imagine "Albert Einstein thinking about fundamental physics 24 hours a day" without eating, sleeping, or aging - continuous deep thinking for a year could solve intractable problems

Multi-agent orchestration now enables models to reason across years of notes, identify contradictions, and refactor 50 million line codebases in days versus weeks with human teams

Frontier Models Capture 90% of AI Revenue Despite Open Source Token Volume

Frontier models capture 90%+ of AI revenue while open source handles majority of token volume - "All revenue will accrue to the Pareto curve" for coding and high-value tasks - Gavin

Enterprise survey shows companies optimizing with model routing but still expecting to consume significantly more frontier tokens for critical workloads like coding

Harvey achieved better outcomes than Opus using open source models with proprietary legal data and routing, but still consumed substantial Opus tokens for premium tasks

"You don't need Albert Einstein to book you a trip" but enterprises won't accept second-tier code - high-value work continues flowing to frontier models despite optimization efforts

NVIDIA Maintains Dominance as ASIC Competition Fragments

NVIDIA maintained market share despite predictions of dramatic ASIC gains, with new accelerators like MediaTek V8T and Broadcom V8I becoming more workload-specific rather than general replacements

OpenAI's Jalapeño chip performs well but requires more cooling and power, raising questions about focus versus vertical integration for frontier AI companies

"If all of his customers are going to compete with him, then why not compete with his customers?" - Jensen could release frontier open source models and become major cloud provider - Gavin

In watt-constrained environment, NVIDIA's superior tokens-per-watt performance drives revenue advantages even when competitors offer lower upfront costs

$1.5 Trillion CapEx Justified by $300+ Billion AI Revenue Growth

Morgan Stanley raised 2027 CapEx forecast from $950 billion to $1.1 trillion, likely $1.5 trillion including SpaceX and NeoCloud investments not captured in traditional forecasts

AI lab revenue projected at $300 billion by 2027 with 50-70% gross margins, while less than 0.2% of Earth's population currently uses AI agentically

Token pricing increasing rather than deflating as expected - monetization rates rising from $20 billion to $30-40 billion per gigawatt as demand outstrips supply

Dario Amodei predicts "trillions of dollars in revenue before 2030" with country-level genius AI systems by 2028, supporting massive infrastructure investment thesis

Resources Mentioned

Atomic Habits An Easy & Proven Way to Build Good Habits & Break Bad Ones

James Clear's habit formation framework was referenced when discussing how small daily improvements compound over time, relating to SpaceX's rapid execution capabilities and systematic approach to building infrastructure.