The SpaceX IPO, Fable 5, AI Capex Update & Market Check w/ Gavin Baker
Brad Gerstner hosts Gavin Baker (Atreides), Andrew Fox and Altimeter partner Clark Tang two days before SpaceX's record IPO ($135/share, **$1.77 trillion valuation**, banks modelling ~$160bn revenue by 2028). The frame: don't pump it, decompose it from first principles. The bull case rests on three terrestrial legs that all 'pencil' before you need the moonshot. (1) **Launch + Starlink** — rapid Starship reusability drives cost per kg from ~$1,500 on Falcon toward sub-$250, and Starlink direct-to-cell can plausibly 5x connectivity revenue from ~$10bn to ~$50bn (still only 0.3% of the global telecom market). (2) **AI compute ("Elon Web Services")** — the surprise of the last six weeks: SpaceX struck huge compute resale deals with Anthropic and Google and **went from non-hyperscaler to #4 in 30 days, passing Oracle**, monetising at $22-23bn/GW (Anthropic) and ~$50bn/GW (Google) vs an implied $14bn/GW in the model. (3) **The model itself — the panel's highest-conviction overlooked upside:** the Cursor acquisition (~700-800 people, ~$10bn run-rate) injects proprietary coding data into Grok 4.3 training; Composer 2.5 was Pareto-dominant on coding 12 days ago. Orbital compute is a call option, not a requirement: Fox lays out **~$5bn/GW capex to put compute in space vs ~$20-25bn/GW terrestrially** — a 5x cut on half the bill of materials, assuming satellite reliability holds. On the broader capex debate, Gerstner flags Morgan Stanley lifting 2027 capex to $1.1tn (he thinks ~$1.5tn incl. SpaceX/CoreWeave) against ~$300bn inference revenue; the panel argues the math 'maths' because inference gross margins are 50-70%+ and revenue is being underestimated (>$200bn this year). On models: Fable 5 / Mythos and ChatGPT 5.5 mark a new class defined by long-running tasks — Noam Brown's point that benchmarks should be measured in time/tokens, not snapshots, made Baker 'a lot more bullish' on compute. The open-source debate resolves to 'two things can be true': **frontier captures ~90% of economic value while open source is ~80% of tokens** — bearish for frontier-lab margins but bullish for compute. Market check: Gerstner trims from 'large' to 'medium-small' into a seasonally weak, inflation-pressured summer (CPI ~4.2%), but stays structurally long.
Key points
- SpaceX IPO priced at $135/share, ~$1.77tn valuation; Goldman/banks model ~$160bn revenue by 2028 vs ~$18bn last year (an ~8x); Elon owns ~50% and is locked up 365-366 days.
- SpaceX went from non-hyperscaler to the #4 AI hyperscaler in 30 days after the Google deal, passing Oracle — 'Elon Web Services'; deals signed at $22-23bn/GW (Anthropic) and ~$50bn/GW (Google) vs an implied ~$14bn/GW monetisation in the leaked model.
- Clark Tang's analysis: xAI's Google cloud-compute deal generates more operating profit per gigawatt than Anthropic, Meta, Google or OpenAI; Altimeter's Freda calculated a 55% ARR on Colossus 1 (borrow at 6-8%, invest at 55%).
- Orbital compute economics (Fox): two-stage rapid reusability cuts launch to ~$250/kg from ~$1,500/kg on Falcon; ~5MW per Starship launch yields ~$5bn/GW capex in space vs ~$20-25bn/GW terrestrially — a 5x cut on the ~$25bn non-silicon half of a ~$60bn/GW data centre.
- Cursor acquisition (~700-800 people, projected up to $10bn revenue this year) is the panel's biggest overlooked upside: its proprietary coding data is being injected into Grok 4.3 (1.5tn-parameter) pre-training; Composer 2.5 was Pareto-dominant on coding 12 days ago.
- Fable 5 (Mythos with cyber/bio/chem classifiers, falls back to Opus 4.8) and ChatGPT 5.5 define a new model class judged on long-running tasks; Karpathy called it SOTA, Noam Brown argues benchmarks should be measured in time/tokens/compute, not snapshots.
- Open-source resolution: 'two things can be true' — frontier models capture ~90% of economic value while open source is ~80% of tokens; bearish for frontier-lab margins but bullish for compute/hardware (Harvey beat Opus 4.7/4.8 at lower cost via fine-tuned open source + a router).
- AI capex vs revenue: Morgan Stanley lifted 2027 capex to ~$1.1tn (Gerstner estimates ~$1.5tn incl. SpaceX/CoreWeave) vs ~$300bn inference revenue — panel says the math works given 50-70%+ inference gross margins and revenue >$200bn this year; ~35% of spend is non-revenue (training).
- Nvidia is gaining, not losing, share: maintained share 'handsomely' (gained in 25/26 ex-Anthropic); ASIC landscape nuanced — MediaTek V8T vs Broadcom V8I for TPUs; OpenAI's chip is 'great' but runs hotter (more cooling/power); Meta and Microsoft ASICs 'disappointing'.
- Total capital being raised across Anthropic + OpenAI + SpaceX (~$250bn) is only ~1% of the Mag 7 market cap; post-deals SpaceX trades ~39x trailing revenue vs ~100x a month earlier after adding ~$29bn in a month.
- Market check: semis ripped (doubles/triples) but huge dispersion — Internet -16%, software -8% YTD; CPI ~4.2%, oil ~$100, war with Iran; Gerstner trimmed from 'large' to 'medium-small' expecting consolidation before higher highs; AI usage is seasonally weak in summer (college kids).
Notable quotes
In 30 days we went from not being an AI hyperscaler to being number four and we passed a lot of companies, including Oracle.
The implied monetization rate on that number is something like $14 billion per gigawatt per year for the AI business.
But the math you get to is it's about $5 billion per gigawatt of capex to put these in space.
And then it is a little funny to me that it was 100 times trailing TTM revenue.
And so I think for most institutional investors, it's a must buy, A must own is, set it and forget it, right?
Because nobody has run Mythos for a year continuously.
90% That has been decisively wrong, probably more than 90% and it may continue to be decisive.
So the better open source does, the better it is for compute providers.
And especially as long as we are in a Watt constrained world, if you can get more tokens per watt, which is literally revenue with Nvidia, then a lot of alternatives, just if you build your factory with another chip, you may save some money, but you're going to have less revenue and the margins may be lower.
Less than 0.2% of people on Earth are actually using AI in an agentic way.
Themes
- AI-capex sustainability
- IPO-window reopening
- compute-as-the-binding-constraint
- frontier-vs-open-source value capture
- orbital data-centre economics
Mentioned
People
Companies
Ideas
- Pareto frontier of intelligence-per-cost
- orbital / space-based compute
- rapid Starship reusability
- Starlink direct-to-cell
- AI compute resale ('Elon Web Services')
- long-running-task model class
- frontier value vs open-source token volume
- monetization per gigawatt
- AI capex vs inference revenue math
- model routing
- Nvidia tokens-per-watt in a watt-constrained world
- set-it-and-forget-it position sizing