The AI boom has a circular capital problem nobody wants to talk about, and a power grid that's 70% end-of-life

Spent the last week going through the financial and operational data behind the AI industry. A few things that don't get enough coverage:

**The circular money loop**

OpenAI spent $12.43B on Azure inference between CY2024 and Q3 2025. Its total external revenue in all of 2024 was $3.76B. The inference costs have materially eclipsed revenues.

Microsoft funds OpenAI. OpenAI buys Azure. Azure revenues flow back to Microsoft. OpenAI is now contractually committed to purchasing $250B in Azure services going forward, and Microsoft owns 27% of OpenAI. The investor and the infrastructure vendor are the same entity.

Constellation Research's analysis of Microsoft's Q2 FY2026 earnings found that OpenAI alone represents 45% of Microsoft's entire remaining performance obligations, $625B total, up 110% YoY.

**The physical wall**

Northern Virginia data centers are now facing 4–7 year grid interconnection delays (WebProNews, March 18). Dominion Energy has publicly stated it cannot build fast enough. The US DOE projects data centers could hit 12% of total US electricity consumption by 2028, up from under 4% today.

Oracle is building Stargate data centers with on-site natural gas generation specifically to bypass the public grid. xAI built its own hybrid power plant in Memphis. When a $500B infrastructure project needs its own power station, the grid has become a limiting reagent.

**The ROI gap**

PwC's 29th Global CEO Survey (Davos, January 2026): 56% of CEOs say AI has produced neither increased revenue nor decreased costs in the past 12 months. Only 12% have achieved both. Gartner forecasts $2.52T in global AI spend for 2026, with infrastructure consuming more than half.