Nvidia Could Become the ‘Federal Reserve Of AI,’ Gavin Baker Says: Will ‘Dark GPUs’ Become Its First Test?

Nvidia (NASDAQ:NVDA) is building what could become the “Federal Reserve of AI,” Atreides Management CIO Gavin Baker said on Friday’s All-In Podcast.

The investor says CEO Jensen Huang is transforming Nvidia’s chips into a financeable asset class, opening the door for Wall Street to finance a much larger AI buildout.

But co-host David Sacks warned there is one way the strategy could end badly for everyone: “dark GPUs.”

Nvidia last week unveiled a plan with Wall Street firms to mobilize more than $500 billion for AI infrastructure. On Monday, Nvidia backed SB Energy’s Ohio data-center buildout and said it will invest $1.5 billion in the developer.

Baker Says Nvidia Is Becoming AI’s ‘Federal Reserve’

Baker said financing has become a bottleneck for AI growth, with the market “constrained by the ability to finance this buildout.”

Nvidia is trying to remove that constraint by making GPUs easier for Wall Street to lend against. The chips generate rental income, can remain productive for years, and Nvidia may provide residual-value support of up to 25% on individual deals.

That makes Nvidia a “matchmaker” between compute buyers and lenders, Baker said, potentially turning it into the “central bank of AI, the Federal Reserve of AI.”

Sacks Warns of ‘Dark GPUs’

Making compute easier to finance raises an obvious question: what if too much gets financed?

Sacks, framing it as the downside scenario rather than a prediction, said the biggest risk is a “glut of compute” echoing the dot-com era, when unused fiber-optic lines became known as dark fiber.

“If you had dark GPUs, that’d be a disaster for everyone,” he said.

Easier financing could fuel a huge buildout, leaving supply ahead of demand and pushing down rental prices and GPU residual values.

Sacks added that permitting fights, power constraints and political resistance to data centers may act as a natural brake against overbuilding.

Prediction Markets Put a Price on Nvidia Compute

Kalshi traders aren’t pricing a compute glut: they give a 61% chance H100 SXM rental prices finish the year above $3.23 an hour, versus $2.79 currently.

A sustained decline in rental prices would be a warning sign of the oversupply Sacks describes.

Baker sees the financing push as removing the biggest constraint on the AI buildout. Sacks’ warning is that removing it too successfully could create a new problem: too much compute.

Image: Shutterstock

Kalshi and Benzinga have an existing data collaboration agreement.