"The SpaceX IPO and Data Centers in Space" — Ben Thompson
Why this is in the vault
Thompson's strategic frame for why the absurd $2T SpaceX IPO valuation could be defensible — data centers in space as the long-run AI-compute escape hatch — directly feeds RDCO's active SpaceX-IPO and Power Cycle (AI-infra power constraint) investing theses.
The core argument
The SpaceX S-1 is "silly" on its face: a $2T valuation on $18.67B revenue with $4.9B in losses, growth slowing 35%→33%, and a $28.5T TAM claim (incl. a $26.5T AI bucket Thompson openly mocks). No financial model justifies it today. But Thompson argues the relevant question isn't whether the numbers work now — it's whether the dream is possible and whether there's reason to think it happens. He thinks data centers in space clear that bar, on three axes:
- Possible (the "is it real" test): Don't picture earth-style buildings launched to orbit. Picture an individual satellite as a rack. A V2 Mini Direct-to-Cell satellite is already in the size range of an Nvidia NVL72 rack; the gap is power/heat dissipation (25kW vs 135kW), solvable with bigger solar arrays, 200m²+ radiators, and disposable larger/older chips that tolerate radiation and need less power. A laser-interconnected fleet of rack-satellites is feasible.
- Carrot (use case): Ties to his "Inference Shift" thesis — three workload types (training, answer inference, agentic inference). Agentic inference, where no human waits in the loop, tolerates latency, so it favors cheap/slow memory and "good enough" compute over bleeding-edge GPUs. That is the largest future market (scales with compute, not humans), and it's the workload that fits racks-in-space.
- Stick (forcing function): Terrestrial buildout is constrained not just by power but by zoning — communities now have veto power over data centers (unlike globalization-era factory closures). That raises terrestrial cost and, at sufficient compute demand, makes space "not just an alternative but the only choice."
Monetization anchor: SpaceX already earns $15B/yr from xAI's Colossus 1 (300MW ≈ "3,000 racks-in-space"); Anthropic would likely pay ~3x per unit of capacity. Even without xAI competing at the frontier, SpaceX could be a monopoly provider of marginal compute. Thompson still calls the IPO "nuts" given the stacked assumptions (Starship works, chip supply, demand explosion, agentic unbundling, zoning opposition succeeds), but is "glad it exists" — Musk's track record earns benefit-of-the-doubt, it's a hedge against an AI compute shortfall (his "replay of nuclear power" fear), and it's "a return to what an IPO should be": retail capital funding real build-out, retail investors getting to "play VC." Closing line: "you get Mars upside for free."
Mapping against Ray Data Co
This is a high-value cross-source corroboration for two active investing workstreams:
SpaceX-IPO thesis cluster. Thompson independently lands on the same ~$1.75–2T valuation bracket and the same "numbers are absurd but the dream might be real" posture already captured in [[2026-05-23-moonshots-ep-257-spacex-ipo-transcript]] (Moonshots EP257) and the S-1 financial teardown in [[2026-05-21-mostlymetrics-spacex-ipo-s1-breakdown]]. Where Moonshots is bullish-promotional and Mostly Metrics is numbers-first, Thompson supplies the strategic mechanism (rack-satellites + agentic inference + zoning forcing-function) that turns "Starlink → data centers → moon → Mars" from a slide into a falsifiable chain. The note triangulates the same event from three independent sources — useful evidence weight if a SpaceX position is ever sized.
Power Cycle thesis (AI-infra power constraint). Thompson's "stick" — terrestrial data-center buildout blocked by zoning and power, with $15B/yr Colossus 1 economics as the unit anchor — is the exact bottleneck [[2026-05-17-power-cycle-v1]] is built on, and it extends [[2026-05-18-stratechery-data-center-discontent]] (his own prior Update on data-center opposition) from "permission is the constraint" to "permission constraint → space is the release valve." Concretely: if Thompson's zoning-as-binding-constraint read holds, it strengthens the Power Cycle thesis (power/interconnect scarcity is real and worsening) but introduces a long-tail bear case — orbital compute as an eventual pressure-relief valve that caps terrestrial scarcity rents. That's a thesis-refinement input, not a thesis-killer (space is a 5-10yr+ optionality, not a near-term supply add).
Agentic-inference architecture shift. The "agentic inference favors cheap/slow memory + good-enough compute, unbundling the GPU" claim is a direct input to any AI-infra sub-thesis touching memory (DRAM vs HBM) or non-Nvidia compute (Cerebras/Groq). Worth carrying into hyperscaler-capex anchor work — if agentic inference de-emphasizes HBM/bleeding-edge GPUs, the capex mix shifts.
Related
- [[2026-05-23-moonshots-ep-257-spacex-ipo-transcript]] — Moonshots EP257, same SpaceX-IPO event, bullish framing; Thompson supplies the strategic mechanism their "stepping stones" pitch lacks.
- [[2026-05-21-mostlymetrics-spacex-ipo-s1-breakdown]] — CJ Gustafson's S-1 financial teardown; numbers-first companion to Thompson's strategy-first read.
- [[2026-05-18-stratechery-data-center-discontent]] — Thompson's own prior Update on zoning/permission as the data-center constraint; this article extends it to the space "release valve."
- [[2026-05-17-power-cycle-v1]] — RDCO Power Cycle investing thesis on AI-infra power/interconnect scarcity; Thompson's "stick" both corroborates and adds a long-tail bear case.
Note: Stratechery is subscription-funded with no third-party sponsor (sponsored: false). The email footer carries first-party Stratechery Plus bundle cross-promo (Sharp Tech, Dithering, Sharp China, GOAT, Asianometry podcasts) — not paid placement. Article body paraphrased from the authenticated email; no web reconstruction needed.