The Dark AI Factories: Powerless Data Centers And The Real Risk To Nvidia - III
Dumb and Dumber: The AI Boom That Ran Ahead of the Power Grid
IMPORTANT NOTE: This newsletter contains information for educational purposes only, and the content below should not be considered financial advice to readers. We DO NOT publish our short research and ideas on substack due to compliance reasons. We do not recommend shorting to retail investors. Our short ideas and research are only for institutional investors and not individual investors (retail investors). If you are institutional investor or sophisticated investor, and would like to become our client, email laks@unicusresearch.com.
One of the biggest challenges facing the utility companies was laid out in a recent report by Siemens, which noted that “forecast inputs are often overestimated or unreliable due to developers’ optimistic projections.”
On November 14, we walked through how AI data centers are quietly draining the basics from households: water, electricity, and land.
On November 18, we unpacked the next domino, how data center build-outs are chipping away 1.9 million family-run U.S. farms.
Today, we’re focusing on powerless data centers and the real risk to NVDA 0.00%↑. At the simplest level, Nvidia’s business model is selling shovels into the AI “data-center rush.” That’s a dangerous way to make a living when the grid can’t actually power the facilities those “shovels are meant to dig.”
IDLE EYESORES
Two gleaming new data centers in Santa Clara, a few minutes from Nvidia’s headquarters, are sitting dark. Digital Realty’s SJC37 and STACK Infrastructure’s SVY02A were built for high-density AI clusters, each designed for roughly 48 megawatts (MW) of critical load. But the city-owned utility, Silicon Valley Power, doesn’t have the grid capacity to energize them. Local plans call for about $450 million of new substations and lines, with full upgrades not expected until around 2028, which means these buildings may remain partly or fully idle for years.
This isn’t a quirky local problem. A Department of Energy–backed study estimates U.S. data centers used about 176 TWh of electricity in 2023—roughly 4.4% of all U.S. power—and could rise to as much as 12% by 2028 as AI workloads surge. Utilities and grid operators are scrambling to catch up: AEP Ohio has load-study requests for roughly 13 GW of new data center demand in its territory alone, while Amazon has formally accused PacifiCorp in Oregon of failing to deliver contracted power to four new data center campuses.
A 1-gigawatt (GW) “hyperscale AI” campus can cost on the order of $35 billion to fully build out, with roughly 39% of that in GPUs and about 10% in land and buildings, according to estimates shared by TD Cowen and Bernstein. At that scale, even a 100 MW slice of capacity represents several billion dollars of capital that only earns a return when the lights are on. And because Nvidia is now on an annual GPU release cadence—Hopper to Blackwell to Rubin and beyond—every year a rack of GPUs sits idle is a year of performance and pricing power that can never be recovered.
“If we delivered a bad quarter, it is evidence there’s an AI bubble. If we delivered a great quarter, we are fueling the AI bubble”- Huang
Huang knows the deeper problem and it has nothing to do with Nvidia.
The rest of this piece goes deeper: why the grid is so far behind, how much idle shells really cost, why projects are migrating to places like Texas, Louisiana, and New Mexico, and what this power bottleneck really means for Nvidia and its investors over the next cycle.
40% OFF OFFER: The below article is accessible to Confidential Insights members. If you want to subscribe, please note that there is a limited 40% offer for Thanksgiving. Regular confidential insights price is $1000/year- the thanksgiving sale is $600/year. The offer ends this Friday.
You can subscribe here: contrarianunicus.substack.com/subscribe.