Deep|SPCX: xAI 8GW Buildout Mapping

Key Takeaways
- Demand-side 8GW is locked in: existing ~1.5GW + 13k GB300 racks (delivering through 2026, weighted to 2H, ≈2.5GW) + 15k Rubin racks in 2027 (≈4GW+) = ~8GW (per our channel checks). Chips, interconnects, and servers are already slotted into tier-1 OEM production schedules — there is no suspense left on the demand side.
- The existing base is pinned down by independent data: site-level power, chip mix, and tenant structure are triangulated across S-1 disclosures, Epoch AI satellite tracking, and our channel checks (see the reconciliation in Section 1.2). At ~946MW, Colossus 2 is currently the largest single-site AI data center in the world.
- Nobody on the supply side underwrites 8GW: six independent channel sources converge on 3.5–5GW. The gap sits at two gates — power generation (Gate 2) and environmental permitting (Gate 3); chips, transmission & distribution, and capital were each verified as non-bottlenecks.
- China’s supply chain is the only fix for the gap — and it is already in motion: the $1.465B turbine order to the Jereh × FTAI Aviation joint venture (>1GW, delivery by November 2027) is corroborated by public filings; transformer suppliers confirmed “no bottleneck” in interviews.
- Regulation is the biggest uncertainty: every turbine at Colossus 1 operates without a permit; the legal endgame is binary — a workaround (fines / retroactive permitting / relocation) or a shutdown.
- Capacity scenarios (nameplate basis, YE2027, our estimates): Bear 4GW (~25%) / Base 5GW (~50%) / Bull 8GW installed ≈ 5.5–6GW usable (~15–20%) / Tail <3GW (~5–10%). P(8GW installed) ≈ 12–17%.
- Key catalysts (all publicly trackable): ① a Solaris/ProEnergy acquisition announcement; ② a second GW-scale Jereh purchase order; ③ PSD permit approval or a negotiated regulatory settlement. The first concentrated checkpoint is SPCX earnings on August 4.
Executive Summary
评论
?
参与讨论