SemiAnalysis article image on SpaceX AI cloud
SemiAnalysis (3 Nov 2026) newsletter on SpaceX datacenter expansion and power (paid post; based on visible text)., SemiAnalysis

要点

  • SemiAnalysis (3 Nov 2026) says the SpaceX “10GW in 2027” narrative discussed ~two months earlier is revised up to ~13GW; it reports ~3.6GW under construction and >9GW of capacity identified.
  • Build-time cases (as stated): Colossus 1 retrofit ~122 days/150MW; Colossus 2 ~91 days/210MW; Memphis MiniHard ~5 months; Mississippi “Gun Range” ~4 months; Starbase from late-Aug clearing to late-Sep vertical construction.
  • Power/kit: tracks >8GW of likely behind-the-meter supply under Elon’s management vs >9GW datacenter power coming online; Colossus 2 BESS cited ~5.9–6GWh; TeraFab may include a datacenter with possible third-party leasing. Paywalled remainder not used here.

The AI capex fight is shifting from “who can buy GPUs” to “who can deliver megawatt halls and power on a months clock.” SemiAnalysis’s paid note frames SpaceX/xAI-linked sites as a speed story toward possibly the world’s largest AI cloud.

Falsifiable read: do not swallow the “biggest” headline first—check observable MW under construction, interconnection/turbine/BESS progress, and IT MW actually energized by 2027. If permits, imagery, and utility records cannot support the critical nodes in the ~13GW narrative, or build calendars systematically miss the cited cases, discount “SpaceX becomes the largest AI cloud in that window.”

[1]

Visible hard claims (SemiAnalysis, 3 Nov 2026, paid post; this piece uses only fetched visible text and does not invent paywalled pricing/funding-model details):

  • Capacity path: earlier argued 10GW in 2027 was feasible and pointed at a “300b+ ARR” scale narrative; ~two months later cites ~3.6GW under construction, SpaceX estimate revised to ~13GW in 2027, and >9GW identified—rhetorically compared to “all of Google’s datacenter capacity from 1999–2026 in 1.5 years” (SA framing, not a Google statement).
  • Sites/timing: Colossus 2 said to add nearly two “Stargate Abilenes” (2GW) in under a year; MiniHard ~5 months; Gun Range ~4 months; Colossus retrofit records 122→91 days; Project SPAM (mid-Colossus 2) start 5 Oct 2026 targeting December first tranches. Starbase said to host datacenters and possibly well over 1GW on a 2027 (Vera Rubin) timeline.
  • TeraFab: ~100M sqft, 10GW+ Texas chip-fab project with medium-high confidence of an accompanying datacenter; local-official clip said not confirmed SpaceX-only, leaving third-party lease room; project allegedly began as a datacenter then pivoted to TeraFab.
  • Power: Energy Model notes >8GW supply under Elon’s management vs >9GW datacenter power coming online; Colossus 2 side cites Macrohard 700MW and MiniHard 450MW TVA service requests; GIS substation targeted complete by 31 Dec 2026 to Freeport; Southaven nameplate 2.7GW now → 1.24GW by Jul 2027; BESS ~6GWh; speculated PUE ~1.10–1.15 to help close gaps—analysis inference.
[1]

Evidence strength: industry-analysis newsletter plus claimed datacenter/energy models, local filings, and dated site imagery. Figures are largely SA model outputs and site observation—not a full public audit of an official SpaceX 13GW document; “fastest builder” / “most profitable neocloud” are author judgments. Paywalled contract pricing and funding-model detail are unused here.

Steelman: if turbines/BESS/modular MEP and vertically integrated labor (text cites ~3k workers per GW; Tesla may supply modular MEP) truly shorten the critical path, halving typical 8–12 month builds deserves scrutiny even if the headline is loud.

[1]

Strongest steelman: AI-cloud bottlenecks are often power and civil works, not chips alone; whoever parallelizes behind-the-meter supply, interconnection, and halls shapes the 2027 supply curve. Mapping Starbase reuse, Colossus BESS, and TeraFab-adjacent compute gives a trackable watchlist.

Steelman risks: 13GW / 3.6GW under construction / >9GW identified are analyst figures; the Google historical-capacity analogy is easy to misread as official parity; PUE and ~300MW gap closure are speculative; paywalled remainder is not freely third-party-checkable. Falsify via permits, interconnects, and energized IT MW.

[1]

For cloud and semiconductor supply chains, SpaceX/xAI site speed—if real—reprices 2027 colo and power premia; if schedules or supply falsify, paper GW unwind narratives. Watchlist: TVA/local interconnection MW, BESS install progress, energized halls at Starbase/Colossus, and whether TeraFab pads become fab support or leasable datacenters.

[1]

Downgrade “will be the biggest AI cloud” to a testable build plan: IT MW actually energized before 2027, calendar days from dirt to halls on named projects, and whether power (behind-the-meter + grid + BESS) covers the stated load. If verified, the speed story stands; if not, it remains a high-confidence industry hypothesis—not a realized map.

【No.2】

[1]