Gavin Baker — SpaceX IPO and AI CapEx Investment Thesis

Source: The SpaceX IPO, Fable 5, AI Capex Update & Market Check w/ Gavin Baker, Andrew Fox & Clark Tang, Bg2 Pod (Brad Gerstner), June 11, 2026.

The Framework: Two Levers, One Supercycle

Baker's lens for the SpaceX IPO and the broader AI trade collapses to two variables inside an accelerating compute supercycle: (1) how fast SpaceX/xAI can stand up and monetize terrestrial gigawatts, and (2) whether xAI/Cursor can stay on the Pareto frontier for coding intelligence. Everything else — launch, Starlink, orbital compute — compounds from Starship economics or serves as call options.

Layer	Baker's View	Investment Implication
Terrestrial compute	Not a commodity; first-principles design + 122-day build speed	SpaceX becomes hyperscaler #4 in 30 days; premium monetization per GW
Frontier models	Revenue accrues to Pareto frontier; long-running agents break benchmarks	Cursor data + Grok 4.3 pre-training = least-discussed upside
Orbital compute	Optional for IPO math; $5B/GW vs $25B terrestrial infrastructure	Starship reusability unlocks 5× infrastructure cost reduction
CapEx vs revenue	$300B AI revenue estimate is low; margins 60–70%	$1.5T 2027 capex math works if revenue trajectory holds
Near-term market	Bottleneck cliff-climbers need rest; NVDA/Broadcom lagged	Consolidation seasonally; hard to stay bearish on compute demand

Investment Thesis #1: SpaceX IPO — Must-Own Bet on Space and AI

Argument: Baker calls SpaceX a must buy, must own, set-it-and-forget-it position for institutional investors who need a real bet on both space and AI. The bull case is not one line item — it is three businesses (launch/Starlink, terrestrial AI compute resale, frontier xAI models) each with independent paths to the leaked ~$160B 2028 revenue scenario. Bears anchor on $18B trailing revenue vs 8× in three years; Baker breaks it part-by-part and finds Starlink connectivity, terrestrial compute deals, and the Cursor-powered model layer each "totally doable" with upside surprise in the model.

"I don't know another entrepreneur or another business that's a better bet on the future, right, than SpaceX... for most institutional investors, it's a must buy, a must own."

The IPO at $135/share ($1.77T) was ~100× trailing revenue before the Google and Anthropic deals; after ~$29B of new contracts in a month it fell to ~39×. Elon is locked up 365 days; employees and large investors have had liquidity every six months for a decade — an unprecedented supply picture versus typical IPO drawdown charts.

Contrarian element: Wall Street models still anchor on launch and Starlink; the Elon Web Services hyperscaler layer was not in forecasts six weeks ago and is now a major revenue line.

Trigger: Post-IPO: terrestrial GW monetization rates hold at or above signed Anthropic (~~$22–23B/GW) and Google (~~$50B/GW) levels; Grok 4.3 + Cursor pre-training reaches Pareto dominance on coding benchmarks.

Names: SpaceX (PRIVATE — IPO), Alphabet (GOOG) as compute customer, Anthropic (PRIVATE) as compute customer.

Investment Thesis #2: Terrestrial Compute Is Not a Commodity — Speed Is Margin

Argument: Baker rejects the belief that AI data centers are interchangeable commodities. Elon redesigned data centers from first principles the way he did rockets and EVs — 122 days from start to energized GPUs versus a normal three-year plan plus one year of integration for a 100,000-GPU cluster. Clark Tang's analysis shows xAI's Google cloud deal generates more operating profit per gigawatt than Anthropic, Meta, Google, or OpenAI; the Anthropic deal is second. Colossus One carried a 55% IRR at 6–8% borrowing costs (Altimeter's Fred Liu calculation).

SpaceX went from not being an AI hyperscaler to #4 in 30 days, passing Oracle and many NeoClouds. Only two or three players can reliably engineer behind-the-meter data centers; when GE Vernova has limited gas turbines, Baker asks who gets them — X.AI or a startup NeoCloud. Every supplier makes more money when GPUs energize faster: power, land, turbines, chips.

"Speed is literally cost because every day you're paying electricians and plumbers... speed is money for all of the suppliers."

Contrarian element: The market prices NeoCloud scarcity uniformly; Baker argues execution differentiation concentrates margins at SpaceX/xAI and a tiny set of engineering-capable operators.

Trigger: Additional hyperscaler-style multi-GW deals; sustained sub-130-day data-center energization; monetization per GW continuing to rise from early-2026 ~$20B toward $30–40B+.

Names: NVIDIA (NVDA) as supplier, GE Vernova (GEV) as turbine vendor, SpaceX (PRIVATE), CoreWeave (CRWV) as passed competitor.

Investment Thesis #3: Orbital Compute — $5B/GW Call Option on Starship Reusability

Argument: Orbital compute is not required to underwrite the IPO — terrestrial land, power, and chips at high-probability yes gets you to leaked revenue numbers — but it is a massive long-term call option. Clark Tang's first-principles math: ~$5B/GW capex to launch AI satellites vs $20–25B/GW for terrestrial land, shell, power, and cooling (with ~$35B/GW for GPUs/silicon on top of ~$60B/GW all-in terrestrial). Power and cooling are effectively free in orbit; terrestrial infrastructure costs are inflationary while orbital launch costs deflate with reuse.

Starship must achieve two-stage reusability (second stage return later in 2026, refly in 2027), then rapid reusability (30–50 flights per vehicle). Cost per kg falls from Falcon's ~$1,500 toward ~$250, asymptoting toward fuel cost. Each Starship launch carries ~100 metric tons → roughly 5 MW of orbital compute capacity per launch. GPU melt and laser failure are known risks — math works if reliability is not "astronomically" worse than terrestrial.

"It's about $5 billion per gigawatt of capex to put these in space... terrestrially... about 25, 20 to 25 billion per gigawatt... you're talking about putting a gigawatt into space for 30 billion and having lower operating costs."

Contrarian element: Jeff Bezos cites ~six years; Elon says three to pull forward timelines. Baker frames orbital as optional for the IPO but transformative for the watts ceiling after terrestrial bottlenecks.

Trigger: Successful Starship second-stage catch and refly; published orbital-rack specs scaling toward 100–120 kW (Starlink V3 path from 20 kW today); Google-style premiums for "first in line" on space compute.

Names: SpaceX (PRIVATE), Alphabet (GOOG) as terrestrial customer with space call-option premium.

Investment Thesis #4: Cursor + Grok — The Least-Priced Frontier Upside

Argument: The market obsesses over Anthropic and Google compute resale contracts (90-day termination debates, revenue multiples). Baker says the least-discussed upside is xAI's frontier-model leap via the Cursor acquisition: Composer 2.5 was Pareto dominant for coding ~12 days before the podcast; Cursor and Anthropic each hold more proprietary coding tokens than exist on the public internet. Grok 4.3 (1.5T parameters) is training with Cursor data injected into pre-training, not just RL — a critical upcoming data point.

Grok 4.3 was on the Pareto frontier as the most intelligent 500B-parameter model; four frontier players remain: xAI, Google (Gemini 3.1 Pro), Anthropic, OpenAI. Amjad Masad (Replit founder) called strong coding "bitter lesson adjacent" — the fastest path to AGI/ASI because code can automate everything else. SpaceX/xAI can pull compute in-house to train and run models while monetizing excess capacity externally — the AWS/Black Friday analogy.

"If I had to say what the one piece that's being lost in the story... I think they've dramatically advanced their capability when it comes to building a frontier model... this is the place that's getting the least amount of attention and could have the biggest upside surprise."

Trigger: Grok 4.3 + Cursor pre-training release; coding benchmark Pareto leadership; xAI model revenue line exceeding NeoCloud resale in outer-year forecasts.

Names: SpaceX/xAI (PRIVATE), NVIDIA (NVDA), Anthropic (PRIVATE) as coding-data peer.

Investment Thesis #5: Long-Running Agents — The CapEx Math Finally Works

Argument: Fable 5, Mythos, and ChatGPT 5.5 represent a new model class: long-running agents that obsolete snapshot benchmarks. Noam Brown's post — nobody runs Mythos for a year before the next model ships — implies we may never fully measure each generation's intelligence. Baker's thought experiment: Einstein thinking about physics 24/7 for one year without eating or sleeping. That framing made him "just a lot more bullish" on compute than before.

Revenue proved the thesis first with Opus 4.6 in January; enterprise and consumer spend followed. ~90% of economic value accrues to frontier models despite open-source potentially consuming ~80% of tokens (Harvey legal stack: open-source + router beat Opus 4.7/4.8 at lower cost on proprietary legal data — but still used frontier tokens). Open-source strength is bullish for hardware, not bearish: less margin at the model layer → more spend on compute.

CapEx debate: Morgan Stanley raised 2027 capex to $1.1T; Baker thinks ~$1.5T including SpaceX and CoreWeave. AI lab revenue ~$300B in 2027 vs that capex scares the market — Baker calls $300B low, expects well over $200B inference revenue by end of 2026, and 60–70% gross margins on lab revenue. Jensen's ~$1T capex call two years ago was conservative. Less than 0.2% of people on Earth use AI agentically (Alex Sacerdote / Whale Rock framing); monetization per GW rose from ~$20B to $30–40B+ as deals signed. Deals signed in November 2025 may deliver triple the expected return today — "accidental profitability" because labs could not secure enough compute fast enough.

"If you're AIDL, 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."

Trigger: Anthropic/OpenAI run-rate disclosures above $200B trajectory; enterprise survey data showing rising frontier token consumption despite routing optimization; continued monetization-per-GW increases on new deals.

Names: NVIDIA (NVDA), Anthropic (PRIVATE), OpenAI (PRIVATE), Google (GOOG), SpaceX (PRIVATE).

Investment Thesis #6: Nvidia Still Wins the ASIC Debate — But Semis Need a Breather

Argument: Clark Tang's Taiwan takeaway: the "Nvidia vs ASIC winner-take-all" framing is dead — Nvidia maintained share and likely gained in 2025–26 if you adjust for Anthropic's non-Nvidia usage. On paper OpenAI's 27 GW build splits ~10 GW Nvidia, ~10 GW Broadcom, ~6 GW AMD, ~1 GW Cerebras (~30% Nvidia) — Baker calls that outcome "extremely unlikely," especially in a watt-constrained world where tokens-per-watt equals revenue and non-Nvidia fabs yield lower monetization.

OpenAI's Jalapeno ASIC is genuinely good but runs at lower temperature → more cooling → more power. Meta and Microsoft ASIC efforts were disappointing; OpenAI succeeded but Baker questions vertical integration if the superintelligence race ends in two to three years — focus on intelligence, not chip design. Jensen may release frontier open-source models, becoming a cloud competitor to his customers and pressuring ASIC economics ("How would you like open source to join the frontier?").

Near-term portfolio stance: Baker aligns with Gerstner — after a steep semi run, "that game is over" for bottleneck cliff-climbers; many names went straight up a cliff and are tired. Nvidia and Broadcom ironically lagged. Internet −16%, software −8% YTD; CPI 4.2%; seasonally weak period ahead; AI token consumption sometimes plateaus in summer — but Fable/Mythos capabilities make Baker hard-pressed to get structurally bearish on compute.

Trigger: Actual GW deployment shares vs paper OpenAI splits; NVDA Vera Rubin allocation (SpaceX/xAI reportedly ~20% of early scarce supply); semi complex consolidation vs continued index leadership.

Names: NVIDIA (NVDA), Broadcom (AVGO), AMD, Cerebras (PRIVATE), Meta (META), Microsoft (MSFT).

The Ecosystem Map

SpaceX IPO: Baker shareholder; "set and forget" core; ballast can shift around lock-up/dribble-out uncertainty
Atreides portfolio: Cerebras (1 GW on OpenAI paper plan); NeoCloud exposure (CoreWeave)
Altimeter: Colossus 55% IRR analysis; Gerstner buying IPO; medium-small exposure after April/May rip
Frontier labs on Pareto curve: Anthropic (Fable 5, Mythos), OpenAI (ChatGPT 5.5, Jalapeno), Google (Gemini 3.1 Pro), xAI (Grok 4.3, Cursor)
Compute customers of SpaceX/xAI: Google (~~$50B/GW premium), Anthropic (~~$22–23B/GW)
Starlink: Baker uses globally — "fastest, lowest latency"; broadband <1% penetrated; direct-to-cell path to ~$50B by 2028 framed as 0.3% of global telecom
Launch cadence: ~160 launches 2025 → high hundreds → thousands (aspiration: 2–3 Starship launches/day)
Not invested: Reflection AI — Baker wants a US frontier open-source lab but is not an investor

Key Risks

8× revenue skepticism: Few companies in history 8× revenue in three to four years — bears anchor on $18B → $160B leap
Post-IPO drawdown pattern: Historical mega-IPO max drawdowns often >50% — Baker acknowledges unprecedented lock-up/dribble structure but short-term path unknown
Starship reusability failure: Rapid two-stage reusability is "extremely difficult"; orbital and Starlink scale assumptions depend on it
Orbital reliability: GPU melt and laser failure in space; maintenance at scale unproven
Open-source token share: Majority of tokens may stay open-source (Harvey router example) — bullish for compute but compresses frontier model margins
Jensen open-source + cloud competition: Nvidia frontier open-source models could disrupt ASIC ROI math for hyperscaler custom silicon
CapEx ROI skepticism: Chamath and others argue token-maxing / no ROI — Baker counters with millions of independent rational economic actors paying up
Macro / seasonal: CPI 4.2%, Iran war, oil ~$100, seasonally weak summer period; AI token use may plateau when college users leave
Geopolitics and inflation: Short-term pressure on rates and PCE; market held up only because AI revenue surprised — without Anthropic inflection, Baker thinks the whole market could be down YTD
Meta / Microsoft ASIC disappointment: Custom silicon diverting focus from frontier model race

Investment Opportunities at a Glance

Tier	Name / Category	Core Thesis	Conviction Signal
1	SpaceX (PRIVATE)	Must-own space + AI bet; EWS #4 hyperscaler in 30 days; orbital call option	Signed Google/Anthropic deals cut valuation from ~100× to ~39× TTM in one month
1	NVIDIA (NVDA)	Maintained/accelerated share; tokens-per-watt in watt-short world; ~20% early Vera Rubin to xAI	Clark: "Nvidia is accelerating... outexecute competitors"
1	Anthropic (PRIVATE)	Fable 5 / Mythos; long-running agents; revenue inflection proved 2026 thesis	Multi-agent orchestration unlocking enterprise workflows
2	Alphabet (GOOG)	Pays ~$50B/GW premium for xAI compute; first-in-line call option on orbital	Highest operating profit/GW deal in Clark's analysis
2	GE Vernova (GEV)	Finite turbines routed to proven operators (X.AI vs NeoClouds)	Behind-the-meter engineering bottleneck
2	Micron (MU)	Compute shortage + accidental lab profitability → more hardware spend	Implied by open-source-bullish-for-hardware thesis
3	Broadcom (AVGO)	10 GW on OpenAI paper plan; custom TPU/MediaTek V8T competition	Nvidia ASIC share likely overstated at 30%
3	Cerebras (PRIVATE)	Atreides holding; 1 GW OpenAI allocation; prefill/decode role	Named on OpenAI gigawatt chart
3	CoreWeave (CRWV)	NeoCloud beneficiary of compute shortage — but passed by SpaceX scale	"Huge business" yet hyperscaler #4 emerged in 30 days
4	AMD	6 GW on OpenAI paper; Baker skeptical of deployed share	Warrants tied to OpenAI build

Monitoring Checklist

Bottom Line

SpaceX is Baker's must-own compound bet — launch/Starlink core plus a hyperscaler business that did not exist in models six weeks ago, plus a Cursor-powered frontier model upside most investors are ignoring.
Terrestrial compute economics already clear the IPO — $22–50B/GW signed deals and 55% IRR Colossus math mean you do not need orbital faith to underwrite the story; orbital is a 5× infrastructure cost option on Starship reuse.
Long-running agents broke the revenue debate — Fable/Mythos and Noam Brown's benchmark critique imply compute demand scales with time and tokens, not snapshots; Baker thinks AI revenue estimates are still too low for the capex build.
Nvidia's ASIC share fears are overstated — Paper 30% of OpenAI's 27 GW is not Baker's expected equilibrium in a watts-constrained, tokens-per-watt world; Meta/Microsoft custom silicon disappointed while OpenAI's Jalapeno trades cooling costs for performance.
Near-term: rest at the cliff top, not abandon the trade — Bottleneck semis went vertical and need consolidation; Baker and Gerstner both dialed exposure down — but post-Fable compute bullishness makes structural bearishness hard to justify.