SpaceX Ledger Shock: Profit Return Reshapes Power Economics

Profit Turnaround And Segment Mix

I read the quarter as a genuine turnaround : annual revenue up 161%, with quarterly operating profit of 431M$ against a 1.722B$ loss a year earlier. Diluted earnings reached 35 cents a share, yet the nine-month net still shows a 413M$ loss, and the AI division carries a nine-month operating loss of 3.295B$. The S-1 record on SEC confirms the shape: heavy 2025 spending damage, one profitable quarter breaking through. I treat that mix as progress wrapped in risk , not victory yet.

The filing splits SpaceX into three lines: space, connectivity, and AI, and AI is the engine , with revenue up 830% year on year. Quarterly investment hit 19.593B$, nearly double expectations: 16.44B$ for AI, 1.543B$ for Starlink, 1.606B$ for space, leaving AI with an 83.9% share. SEC annual figures add depth: 2025 revenue of 18.674B$, an operating loss of 2.589B$, adjusted EBITDA of 6.584B$, with AI revenue of 3.201B$ against an operating loss of 6.355B$. TechCrunch covered that same AI cash burn, and I read it as deliberate overbuild .

Capacity guidance is the shock : about 3 GW today, at least 13 GW next year, possibly 20 GW if conditions allow, against roughly 50 GW of installed data-center capacity worldwide. I anchor that in the SEC Colossus record and abiresearch.com analysis from ABI Research: the first cluster of near 1 GW went live in 122 days, the second in 91 days. The dated path runs 1.4 GW in June 2026, 1.7 GW in September, 2.5 to 3 GW by year end, then 10 GW more by end of 2027, near 1B$ per GW.

Musk claims each GW costs 40 to 50B$ yet returns that sum as revenue within about a year, not profit but revenue . I find the pricing split behind that claim: SpaceX sells near spot at about 50$ per MW of operating capacity, while CoreWeave still serves legacy contracts below 30$. Yahoo notes Nebius closed four deals above 1B$ at 20 to 25M$ per MW, and CoreWeave pricing pages show new contracts resetting higher. If expiring contracts reprice toward spot, rental income across the sector steps up quickly.

Capacity Race Meets Power Limits

The brake is not chips but power : electricity, distribution, and cooling set the pace, and Musk admits those links bind first. My fix watch starts with onsite generation, led by the bloomenergy.com story of Bloom Energy with Oracle: up to 2.8 GW of fuel cells, 1.2 GW already contracted, installed in 55 days. A 2026 Bloom study expects about one third of data centers to run fully onsite power by 2030, while new builds shift toward 800V DC designs that cut conversion losses.

Cooling and power delivery decide who monetizes the overbuild . Seeking Alpha relays a BNP Paribas view naming Vertiv and Eaton as cooling winners; Eaton just paid 9.5B$ for Boyd Thermal, while Schneider packages Motivair gear beside its own. Liquid cooling revenue grows from 5.52B$ toward 15.75B$ by 2030 in my notes. On power conversion, onsemi and Navitas stand out on 800V designs that waste less energy per rack. I rank these suppliers as leveraged plays on every GW commissioned.

Every AI dollar here currently flows to Nvidia hardware, with own silicon years away: the Terafab plant targets test output in 2027 and serial output in 2028, a schedule I would discount. TechCrunch and the SEC filing agree on the interim picture of single-vendor dependence plus heavy memory pull, with Micron as the lead memory supplier and a close commercial relationship. Capex lines confirm it: 12.727B$ of AI capex in 2025, 3.832B$ for space, 4.178B$ for connectivity, then 7.723B$ of AI capex in the first quarter of 2026 alone.

Scarcity shows in rental menus : Yahoo reports Nebius raised prices from October, with H100 moving 3.85 to 4.50$, H200 4.50 to 5.40$, B200 7.15 to 8.50$, and B300 7.85 to 9.50$. CoreWeave answers with volume, spending 9.4B$ of quarterly capex on 2.575B$ of quarterly revenue while guiding 35 to 39B$ for 2026. Musk says even SpaceX failed to rent outside capacity from Nebius, CoreWeave, or IREN-style clouds. I read empty shelves, not weak demand, as the binding constraint .

Orbital Cash And Software Options

Starlink funds patience elsewhere, and Basenor data frames the flywheel : about 12M subscribers at end-June heading to 13.6M by September, average revenue near 66$ a month, raised toward 69$. The 2025 base counts 8.9M subscribers and 11.4B$ of revenue, with enterprise and public income up 91% in a year. Aviation extends the runway from 3,008 fitted aircraft toward 12,000 contracted, while direct-to-cell counts 22M monthly users. I see a second growth engine beside the data-center build .

Pressure concentrates on rivals, starting with the Viasat venture Equatys with Space42: up to 1B$ of capital, 2,800 satellites, 60 orbital planes, and over 100 MHz of coordinated spectrum. Legacy satellite names such as Viasat, SES, and AST look exposed, alongside carriers T-Mobile, Verizon, and AT&T. Software tells the second story: the coding firm bought in August printed 680M$ from mid-August to September end, Grokbot holds 800 thousand weekly paid users, and new X tiers from 8$ to 200$ pool weekly revenue, per PricingSaaS. Grok did about 350M$ in 2025 with 64M monthly users, per the businessofapps.com record of Business of Apps, which leaves Microsoft under siege .