All podcasts / BG2 Pod / Summary

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

2026-06-11 - 81 min - source - Read full transcript
Brad Gerstner (host)Clark Tang (host)Gavin BakerAndrew Fox

Key insights

SpaceX's IPO math hinges on two levers: how fast it brings terrestrial data centers online and whether it achieves rapid Starship reusability for orbital compute.
Priced at $135/share ($1.77T) with banks projecting $160B revenue by 2028, Baker frames the bull case around Elon's demonstrated 122-day data-center buildout speed and, separately, the longer-dated orbital-compute optionality once Starship's second stage is reusable.
spacex-ipo
Speed of data center construction is itself a direct cost and monetization advantage.
SpaceX brought a 100,000-GPU cluster online in 19 days versus a normal 3-year planning cycle plus 1-year build for a comparable supercomputer; every day of delay is pure cost (electricians, plumbers), so faster buildout both lowers cost and lets SpaceX monetize compute sooner and at a premium.
ai-capex-buildout
Google is likely paying a premium for SpaceX terrestrial compute partly to secure a place in line for future orbital compute capacity.
The panel notes SpaceX's compute deals monetize at higher rates per gigawatt than most peers (Anthropic's deal reportedly higher still), and speculates that hyperscalers are paying up now to guarantee early access if orbital data centers become real, effectively a call option on space compute.
spacex-ipo
Orbital data centers could cut CapEx per gigawatt roughly 5x versus terrestrial once two-stage Starship reusability is achieved.
Terrestrial buildout costs roughly $60B per gigawatt today ($35B chips/silicon, $25B land/shell/power/cooling); the $25B non-chip portion could fall toward roughly $5B in space since power and cooling are close to free and reusable rockets amortize launch cost toward the price of fuel, implying total space CapEx near $30B per gigawatt.
orbital-compute
The most underappreciated part of the SpaceX story may be its model business, not its compute or launch business.
The Cursor acquisition brought a team with more proprietary coding tokens than exist on the public internet; that data, fed into Grok 4.3 pretraining (not just RL) on top of massive new compute access, briefly put Composer 2.5 Pareto-dominant on a coding benchmark within days, suggesting xAI/SpaceX AI has a real shot at the model frontier.
spacex-ipo
The frontier is shifting from single-pass benchmark scores to long-running, hours-long agentic task completion, and nobody can fully evaluate a model's ceiling before it's replaced.
Citing Noam Brown, the panel argues the right evaluation axis is now time or compute rather than a snapshot score, since models like Fable 5/Mythos can sustain multi-hour tasks; but because no lab runs a frontier model continuously for a year before the next release, true capability ceilings remain unknown.
frontier-vs-open-source
Contrary to years of predictions, frontier models have not lost share to cheap open-source tokens; if anything the gap has widened on economic value even as open source dominates raw token volume.
The panel estimates frontier models capture roughly 90% of AI economic value in 2026 while open source may account for roughly 80% of tokens consumed, because frontier models are the ones reliably carrying through full user intent on long-running, high-value tasks like coding, not just answering isolated prompts.
frontier-vs-open-source
2027 CapEx forecasts have risen to roughly $1.5 trillion against a currently modeled ~$300B in AI lab inference revenue, but the panel argues that revenue figure is understated and gross margins (50-70%+) make the unit economics work.
Morgan Stanley's 2027 CapEx forecast moved from $950B to $1.1T excluding SpaceX/CoreWeave, implying a total closer to $1.5T; against this, per-gigawatt monetization has risen from about $20B to $30-40B over roughly a year, and the panel expects 2026 inference revenue to end well over $200B, undercutting bear-case 'no ROI' arguments.
ai-capex-buildout
Compute demand may stay supply-constrained for years because agentic AI adoption is still extremely early.
Citing a stat that under 0.2% of people on Earth currently use AI in an agentic way, and describing individual usage patterns consuming hundreds of CPU cores and multiple GPUs continuously, the panel argues even modest growth in adoption implies a persistent compute shortage.
ai-capex-buildout
The ASIC-versus-Nvidia narrative has shifted from a binary contest to workload-specific accelerator selection, and Nvidia has held share better than expected.
Despite heavy ASIC investment from Broadcom, AMD, and OpenAI's own chip ('Jalapeno,' which needs more cooling than Nvidia GPUs), Nvidia has out-executed competitors because tokens-per-watt (i.e., revenue-per-watt) still favors its hardware in a power-constrained environment; the panel expects on-paper 2027 ASIC capacity (roughly 30% implied share) to undershoot in practice.
ai-capex-buildout
Altimeter has trimmed its AI/semis exposure from 'large' to 'medium-small' after a sharp rally, while remaining structurally bullish.
Gerstner cites elevated expectations after prices ran up in April-May, plus geopolitical and inflation concerns (CPI back above 4%), as reasons to size down within a 'set it and forget it' framework that adjusts position size with risk-reward rather than exiting core positions.
market-check
AI-driven markets have been unusually seasonal, cooling in summer as college-student token usage drops, which both investors treat as a normal near-term headwind rather than a thesis change.
Token consumption has plateaued for three consecutive summers per the panel's observation, coinciding with reduced student usage; combined with a modest shift toward cheaper open-source tokens in recent weeks, this is framed as a reason for near-term caution without undermining the longer-run compute-shortage thesis.
market-check

Media referenced

Companies

Techniques and frameworks

Summary

Brad Gerstner and Clark Tang host Gavin Baker and Andrew Fox of Atreides Management for a deep dive on the SpaceX IPO, two days before it prices at $135/share ($1.77 trillion valuation) against Wall Street forecasts of $160 billion in 2028 revenue. The panel breaks the bull case into three business lines: Starlink/launch (foundational, with rapid two-stage Starship reusability seen as the key unlock for driving cost per kilogram from roughly $1,500 down toward $250 and eventually the cost of fuel), a suddenly massive AI-compute resale business built on Elon's unmatched speed at standing up data centers (a 100,000-GPU cluster online in 19 days versus a normal multi-year build), and, least discussed but potentially highest-upside, the model business itself following the Cursor acquisition, which brought xAI a trove of proprietary coding data that helped Composer 2.5 briefly reach Pareto-dominance on a coding benchmark.

Baker and Fox walk through the economics of orbital data centers: once Starship achieves reusability, launching AI-satellite compute into space could cost roughly $5 billion per gigawatt of CapEx versus $25-30 billion for the equivalent terrestrial land, shell, power, and cooling infrastructure, because power and cooling are effectively free in space. They stress this optionality isn't required to justify the IPO valuation, since the terrestrial buildout alone, at current per-gigawatt monetization rates, can support the leaked revenue numbers.

The conversation pivots to frontier AI more broadly following Anthropic's Fable 5 release (essentially Mythos with added safety classifiers). Citing a Noam Brown post, the group argues that snapshot benchmarks are becoming less meaningful as models gain the ability to sustain coherent work over hours, and that no lab has run a frontier model long enough to know its true intelligence ceiling before the next model supersedes it. Against years of predictions that cheap open-source tokens would erode frontier-model economics, the panel notes the opposite has happened in 2026: frontier models likely capture around 90% of AI's economic value even as open source may represent the majority of raw tokens consumed, because paying customers value models that reliably carry through complex, long-running intent rather than answer isolated queries.

On capital spending, the group works through the widely cited concern that roughly $1.5 trillion in projected 2027 AI CapEx looks large against roughly $300 billion in modeled inference revenue. They argue the revenue figure is likely understated (expecting 2026 to close well over $200 billion) and that rising per-gigawatt monetization (from about $20 billion to $30-40 billion in roughly a year) at 50-70%+ gross margins means the math works, especially with less than 0.2% of the world's population currently using AI in an agentic way. They also revisit the ASIC-versus-Nvidia debate, concluding that despite heavy custom-silicon investment from Broadcom, AMD, and OpenAI, Nvidia has held share better than expected because tokens-per-watt still drives more revenue for compute buyers.

The episode closes with a market check: after a sharp AI/semis rally, Gerstner says Altimeter has trimmed exposure from "large" to "medium-small," citing elevated expectations, inflation data (CPI back above 4%), and geopolitical risk, while Baker frames the market as a runner that sprinted uphill and now needs to rest, without turning bearish on the multi-year thesis. Both frame their approach as "set it and forget it" on core positions, sizing up or down with risk-reward rather than trading around news.

Notable Quotes

"I think we're all pretty AI pilled. And if you're AI pilled, that means we got to build a lot more compute than the world thinks. And that these models are going to be a lot more valuable than people think." - Clark Tang

"Nobody has run Mythos for a year continuously. And we may never know how smart each generation of models actually is or was, because we don't have time to appropriately evaluate their intelligence before the next model comes out." - Gavin Baker

"This whole category of taking all of this compute, which he's uniquely good at standing up, and then reselling it in a way that's highly profitable was not in a lot of people's forecast. Now it's a major component of the forecast." - Brad Gerstner

"Less than 0.2% of people on Earth are actually using AI in an agentic way." - Andrew Fox

"I always assume a bullet is coming for me. Head on a swivel. It's the bullet you don't see that gets you." - Gavin Baker