The SpaceX IPO, Fable 5, AI Capex Update & Market Check w/ Gavin Baker, Andrew Fox & Clark Tang | BG2

SpaceX is reportedly pricing its IPO at $135 per share, implying a $1.77 trillion valuation. Goldman Sachs and the Wall Street Journal both project $160 billion in revenue by 2028, up from $18 billion last year, roughly an 8x increase in three to four years. Gavin Baker frames it as a must-own set-it-and-forget-it position for most institutional investors, representing a simultaneous bet on the space and AI future, though he notes average post-IPO drawdowns across roughly 20 comparable companies including Facebook, Alibaba, and Shopify have exceeded 50 percent. Elon Musk owns approximately 50 percent of SpaceX and is locked up for 365 days post-IPO, and employees and large investors have had liquidity roughly every six months for about ten years, meaning most who wanted to sell have already had close to 20 chances to do so.

Andrew Fox argues that rapid full reusability of both Starship stages is the single foundational capability underpinning the economics of every other SpaceX business line, including Starlink V3, Direct to Cell, and Orbital Compute. SpaceX aims to fly both stages 30 to 50 times before retrofitting and plans second-stage recovery later this year with a reflight targeted for next year. SpaceX conducted approximately 160 to 165 launches last year, with cadence expected to reach the high hundreds within several years and potentially thousands of launches in the three years after that. Starlink broadband penetration currently sits below 1 percent of global households, and Wall Street models project connectivity revenue growing from roughly $10 billion to $50 billion by 2028, which Baker notes would still represent only 0.3 percent penetration of the global telecom market.

SpaceX has moved aggressively into AI compute. Clark Tang reports that xAI stood up a cluster of 100,000 GPUs in 19 days, described as easily the fastest supercomputer deployment on the planet compared to a conventional timeline of three years to plan and one year to deploy after equipment delivery. Baker attributes this to Musk re-engineering data center design from first principles, with the Colossus facility completed in 122 days. Monetization rates on compute deals have escalated sharply, with the Anthropic contract at $22 to $23 billion per gigawatt per year and the Google contract at $50 billion per gigawatt per year, against an implied blended rate of roughly $14 billion per gigawatt per year embedded in the leaked $160 billion revenue forecast. After the Google deal closed, SpaceX became approximately the fourth-largest hyperscaler, passing Oracle and CoreWeave in roughly 30 days. Altimeter calculated a 55 percent ARR on Colossus-class infrastructure, and Baker argues borrowing at 6 to 8 percent to invest in an asset returning 55 percent makes the math strongly favorable.

The Orbital Compute thesis rests on Starship cost-per-kilogram economics. Tang estimates current Falcon 9 launch costs at roughly $1,500 per kilogram to orbit, declining to approximately $250 per kilogram on Starship with full reusability. Each Starship launch can carry approximately 100 metric tons and deliver roughly 5 megawatts of compute capacity. Tang estimates the capital cost to put one gigawatt of compute in orbit at approximately $5 billion, compared to $20 to $25 billion for the non-silicon portion of a terrestrial data center. Baker puts total terrestrial cost at roughly $60 billion per gigawatt and his total orbital estimate at approximately $30 billion per gigawatt, noting that orbital land, power, and cooling are effectively free. Both Baker and Fox flag that two-stage Starship reusability is a prerequisite for Orbital Compute and is not guaranteed, and Baker adds that GPU failure rates and laser reliability in orbit remain known risks.

Baker described xAI's acquisition of Cursor as the biggest underappreciated upside in the SpaceX IPO. Cursor had 700 to 800 employees and was projected to exit the year at up to $10 billion in revenue. Baker says Cursor data is being injected directly into the pre-training process for the next model, not just reinforcement learning. He also states that Grok 4.3 is a 1.5 trillion parameter model currently in training, and that xAI has secured potentially up to 20 percent of Vera Rubin chip capacity.

Baker stated that frontier models are capturing approximately 90 percent of AI revenues, contrary to the earlier thesis that cheap open source tokens would close the gap, which he said has been decisively wrong based on revenue data through the first six months of the year. Total inference revenue will end the current year well over $200 billion, with Baker projecting $400 to $500 billion next year and a path to one trillion dollars or more by 2029. Morgan Stanley raised its 2027 CapEx forecast from $950 billion to $1.1 trillion, and including SpaceX, CoreWeave, and others the 2027 number is likely closer to $1.5 trillion. Baker said however bullish he was on compute before the Noam Brown post on long-running tasks, he is now significantly more bullish on compute demand.

Baker noted that internet stocks are down 16 percent and software is down 8 percent on the year while semis have seen doubles and triples, and that SPY and NASDAQ gains this year are largely attributable to AI and compute related components. He described NVIDIA and Broadcom as laggards relative to other semiconductor stocks in the recent run-up and predicted the current period is likely consolidation on the way to much higher highs.