2 thoughts on “Hyperscaler AI Race: Soaring Capex Wipes Out Free Cash Flow; AGI and Digital Gods”

1. Power, not fiber, remains the dominant constraint on this cycle’s hyperscale buildout, but the two are increasingly the same story. New York’s one-year moratorium on hyperscale data center permitting (Jul 14) and PJM’s capacity auction clearing at its $325/MW-day price cap (Jul 14) show state and regional grid operators independently concluding that AI load is moving faster than the grid can absorb without political or ratepayer friction. Microsoft’s response at Chevron’s Project Kilby (Jun 22), a 2.67GW behind-the-meter gas plant in West Texas, and Brookfield’s fivefold expansion of its Bloom Energy fuel cell partnership to $25 billion (Jun 30), both point the same direction: hyperscalers are increasingly building their own power rather than waiting on interconnection queues. That’s a trend accelerated by NVIDIA’s own chip roadmap, as Vera Rubin’s power draw reportedly rose 500 watts specifically to outrun AMD.

   Fiber tells a parallel story of concentration. Keppel’s Bifrost cable, Lightstorm’s new I-2SEA system, and Lightpath’s builds for two more 1GW-plus campuses all show hyperscalers locking up dedicated capacity years ahead of delivery, while private capital consolidates the smaller providers left behind: Grain Management’s Rightfiber and Zayo’s leadership change both are designed with eventual private equity exits in mind, not organic growth.

   Finally, several key data center operators have published sustainability reports recently, and they are full of fascinating info about energy & emissions. These reports are not covered directly here. However, MTN Consulting is in the process of updating its annual study of hyperscaler energy & emissions (2025 study here). As part of this process, we can share some preliminary research findings.

   The chart below illustrates year-over-year (YoY) % change in energy consumption for several key hyperscalers. As shown, the growth rates for everyone except Microsoft increased in 2025. Alphabet’s jumped the most, from 25% growth in 2024 to a whopping 37% in 2025. The story is similar if you focus on carbon emissions. While Amazon continues to not report energy consumption, fortunately it does report emissions; that data verifies that Amazon is by far the worst polluter in the data center industry. A big portion of its emissions are due to its ecommerce unit, but certainly not all, and this fact doesn’t reduce the company’s impact on the environment. These latest ESG reports should be a clear and loud warning sign that the AI bubble is harming the environment. AI spending is set to counteract much of the hard work the tech sector has done on sustainability over the last decade.

   https://mtnconsulting.substack.com/p/optics-and-watts-who-owns-ais-bottlenecks

   Reply

2. Glaring example of circular AI revenues:
   1] Microsoft buys and licenses core technology from OpenAI. They pays to use OpenAI’s advanced artificial intelligence models and intellectual property, integrating them directly into products like Microsoft Copilot and its Azure cloud ecosystem.
   2]Then OpenAI buys so much of Microsoft cloud services that it accounted for 69% of the growth at Microsoft’s intelligent cloud business last year.

   There’s also the fake “other income” that hyperscalers record as profits from the mark ups on the assessed values of their AI private company investments, like Open AI and Anthropic

   Reply