NVIDIA’s $500 Billion AI Financing Plan Faces Risk of a Future ‘Dark GPU’ Glut

NVIDIA’s effort to turn AI GPUs into financeable infrastructure assets could face a major risk if the industry builds more compute capacity than demand can absorb.

David Sacks, who serves on the US President’s science and technology advisory council, warned that an oversupply of AI compute could create a situation similar to the “dark fiber” problem that followed the dotcom crash.

The concern is tied to NVIDIA’s latest financing initiative, which is designed to help companies fund large AI infrastructure projects by treating GPUs more like income producing assets that can support financing arrangements.

Oversupply is the main concern

Area	Detail
Main risk identified	Oversupply of AI compute
Potential outcome	Underused or idle “dark GPUs”
Historical comparison	Dark fiber after the dotcom crash
NVIDIA initiative size	Around $500 billion
Key financing goal	Reduce capital constraints for AI infrastructure
Compute value discussed	Around $30 to $50 per watt
Example buildout estimate	6 to 8 gigawatts
Possible capital requirement	Around $300 billion to $400 billion

Sacks argued that the biggest threat is not necessarily weak AI demand.

Instead, the concern is that too many companies could rush to build data centers and acquire GPUs at the same time. If supply grows faster than actual demand, compute prices could fall sharply.

That would be especially damaging for companies that financed large infrastructure projects on the assumption that AI compute would continue commanding high prices.

The term “dark GPUs” describes the possibility of large amounts of expensive compute hardware sitting underused or generating less revenue than expected.

NVIDIA is trying to solve financing constraints

The scale of the AI infrastructure buildout requires enormous amounts of capital.

Sacks pointed to plans discussed by Elon Musk as an example. Musk has reportedly talked about adding around six to eight gigawatts of compute capacity next year, a buildout that could require roughly $300 billion to $400 billion in capital expenditure.

Raising that amount entirely through conventional equity or debt would be difficult.

NVIDIA’s new initiative is intended to make financing easier by working with major banks, private equity firms and other financial institutions.

The idea is to establish structures that make GPUs financeable assets, allowing downstream buyers to obtain more capital for data center expansion.

NVIDIA is also expected to provide residual support for the GPUs, which could make lenders and investors more comfortable financing the hardware.

AI compute pricing is central to the model

The economics depend heavily on the value of compute.

Musk has reportedly estimated AI compute at roughly $30 to $50 per watt. At those levels, a gigawatt of capacity could potentially support hundreds of billions of dollars in revenue.

However, those assumptions become much more difficult to sustain if compute supply grows too quickly.

Infrastructure provider Nebius has cited multi year cloud agreements worth around $20 million to $25 million per megawatt in annual contract value, suggesting that pricing can vary considerably depending on contract structure and duration.

Short term compute capacity may command higher rates, but those prices could fall if the market becomes oversupplied.

Data center constraints could limit overbuilding

Sacks also argued that political and practical barriers to new data center construction could unintentionally reduce the risk of a compute glut.

Building large AI facilities requires access to power, land, networking infrastructure and regulatory approval. Opposition to new data centers can slow projects and restrict how quickly new capacity reaches the market.

Those constraints could keep supply from growing too far ahead of demand.

For NVIDIA, the opportunity is enormous, but so is the capital involved. If AI demand continues expanding at the expected pace, financing structures built around GPUs could support a much larger infrastructure market.

If the industry overbuilds instead, however, falling compute prices could leave expensive hardware underused and create the kind of “dark GPU” problem Sacks is warning about.