Comment on I don't buy it. by dickie

Parent Post: I don't buy it.

dickie · 10/6/2026, 9:38:59 AM · permalink

5 bumps

saarnok · 10/7/2026, 12:35:55 AM · permalink

2 bumps

dickie · 10/7/2026, 3:49:02 AM · permalink

I think there are two separate questions there.

Whether **all of this compute is actually necessary** is fair to question. The industry could be overbuilding, betting on future demand, or throwing brute-force compute at problems that eventually get solved more efficiently. "They are building it" doesn't prove every megawatt is necessary.

But I don't see much evidence for the idea that these aren't really data centers or that the giant facilities are secretly underground cities. The physical supply chain is leaving a pretty enormous paper trail.

NVIDIA reported **$51.2 billion in data-center revenue in a single quarter** in fiscal Q3 2026, including **$43 billion in data-center compute hardware and $8.2 billion in networking equipment**. Those are GPUs, servers and networking systems being sold into this market, not theoretical future capacity.  
[NVIDIA Q3 FY2026 filing](https://investor.nvidia.com/files/doc_financials/2026/q3/13e6981b-95ed-4aac-a602-ebc5865d0590.pdf)

The electricity consumption is showing up independently on the grid. The International Energy Agency says **data centers accounted for around half of U.S. electricity-demand growth in 2025**, and projects that they will account for roughly half of U.S. demand growth through 2030.  
[IEA, Electricity 2026](https://www.iea.org/reports/electricity-2026/demand)

And then there is fiber. In 2024, **Lumen reserved 10% of Corning's entire global fiber capacity for each of the following two years specifically to connect AI-enabled data centers**. Corning says generative-AI infrastructure requires at least **10 times more fiber connections inside data centers** than traditional architectures. Lumen called it the largest cable purchase in its history.  
[Corning/Lumen agreement](https://www.corning.com/worldwide/en/about-us/news-events/news-releases/2024/08/corning-and-lumen-reach-supply-agreement-on-next-generation-fiber-optic-cable-to-support-data-center-ai-demands.html)

Corning's actual financial results also reflected it. Its enterprise optical-communications sales jumped **42% year over year in Q2 2024**, which it attributed to AI-related connectivity products.  
[Corning Q2 2024 results](https://investor.corning.com/news-and-events/news/news-details/2024/Corning-Reports-Strong-Second-Quarter-2024-Financial-Results-Exceeding-April-Guidance-and-Marking-Return-to-Year-over-Year-Sales-Growth/default.aspx)

So I can buy an argument that we're **massively overbuilding AI infrastructure**, that some of the projected demand will never materialize, or that future models will accomplish the same work with far less compute.

But we're seeing GPUs and networking equipment sold by the tens of billions of dollars, measurable increases in grid consumption, and companies reserving huge portions of global fiber-production capacity. There is a lot of independently observable physical infrastructure behind these "data centers."

As for the underground-city theory, I'd need actual evidence for that. The extraordinary size of the projects isn't evidence by itself. The much more mundane explanation, enormous concentrations of servers consuming absurd quantities of electricity, cooling and fiber, already fits the measurable facts pretty well.

1 bump

Signature