EPISODE · Jul 16, 2026 · 47 MIN
Why Half of AI's Data Centers May Never Get Built | 50-Year Power Insider
from The Wall Street Skinny · host Kristen and Jen
With hyperscalers like Meta, Google, Amazon and SpaceXAI burning through cash, we decided to answer the question underneath all of it: what is this money actually buying? In this episode we start high level with a primer on the AI ecosystem or what Nvidia's CEO Jensen Huang calls the "five-layer cake" of AI — energy, chips, infrastructure, models, applications. We get into the vocabulary everyone uses and nobody defines: what a hyperscaler actually is, how it differs from a frontier model company like OpenAI or Anthropic, why Oracle only plays in one layer while Google plays in all five, and what a NeoCloud like CoreWeave is really doing when it borrows against its own chips. Then we get into the grid — all three of them — including how power prices get set, the difference between regulated and deregulated states, why Meta's $200 billion Project Hyperion campus in Louisiana needs enough electricity to power half of Manhattan in the summer, and why the new rule for data centers is essentially "bring your own electrons." We also dig into the tax incentives driving the timing of all this spend, and why states are competing so ferociously for projects that employ almost no one once the construction crews go home. Then we bring on an extra special guest: power expert. Ron Kelly, who spent 50 years in power and energy — as an engineer, at Calpine, and developing natural gas-fired power plants and solar plants all over the United States the country. He also happens to be Kristen's dad. His take is bracing: he's seen this movie before. Between 1995 and 2005, roughly 300 gigawatts of power projects were announced on the promise of the internet. 168 got built, 130 were canceled, the rest died, and Enron, Mirant, NRG, and Calpine all ended up in Chapter 11. Today's data center pipeline is about the same 300 gigawatts. Ron explains risks that could complicate the build out necessary to get all the needed power infrastructure online: the interconnection studies, transformer backlogs — plus what he really thinks about the security of the largest machine humans have ever built. Connect with Ron at / ronald-kelly-pe-mba-3587a718
What this episode covers
With hyperscalers like Meta, Google, Amazon and SpaceXAI burning through cash, we decided to answer the question underneath all of it: what is this money actually buying? In this episode we start high level with a primer on the AI ecosystem or what Nvidia's CEO Jensen Huang calls the "five-layer cake" of AI — energy, chips, infrastructure, models, applications. We get into the vocabulary everyone uses and nobody defines: what a hyperscaler actually is, how it differs from a frontier model company like OpenAI or Anthropic, why Oracle only plays in one layer while Google plays in all five, and what a NeoCloud like CoreWeave is really doing when it borrows against its own chips. Then we get into the grid — all three of them — including how power prices get set, the difference between regulated and deregulated states, why Meta's $200 billion Project Hyperion campus in Louisiana needs enough electricity to power half of Manhattan in the summer, and why the new rule for data centers is essentially "bring your own electrons." We also dig into the tax incentives driving the timing of all this spend, and why states are competing so ferociously for projects that employ almost no one once the construction crews go home. Then we bring on an extra special guest: power expert. Ron Kelly, who spent 50 years in power and energy — as an engineer, at Calpine, and developing natural gas-fired power plants and solar plants all over the United States the country. He also happens to be Kristen's dad. His take is bracing: he's seen this movie before. Between 1995 and 2005, roughly 300 gigawatts of power projects were announced on the promise of the internet. 168 got built, 130 were canceled, the rest died, and Enron, Mirant, NRG, and Calpine all ended up in Chapter 11. Today's data center pipeline is about the same 300 gigawatts. Ron explains risks that could complicate the build out necessary to get all the needed power infrastructure online: the interconnection studies, transformer backlogs — plus what he really thinks about the security of the largest machine humans have ever built. Connect with Ron at / ronald-kelly-pe-mba-3587a718
NOW PLAYING
Why Half of AI's Data Centers May Never Get Built | 50-Year Power Insider
No transcript for this episode yet
Similar Episodes
Dec 5, 2025 ·50m
Oct 9, 2025 ·33m
Oct 3, 2025 ·40m
Sep 11, 2025 ·31m
Aug 27, 2025 ·39m
Aug 18, 2025 ·54m