Wall Street's $500B bet on "compute" is really just GPU loans in disguise

Nvidia, Goldman Sachs, BlackRock, Apollo, and several other heavyweight financiers have announced a $500 billion financing package designed to turn computing chips into a legitimate investment asset. The move comes as the company seeks to reassure lenders that GPU hardware remains valuable longer than skeptics believe—and to shore up demand amid concerns about market saturation.

The reframing is strategic. Where Nvidia CEO Jensen Huang once dismissed older chips as nearly worthless, he now portrays them as "revenue-generating assets" with "long-lived" value. BlackRock's Larry Fink has drawn a direct parallel to his role in the 1970s mortgage-backed securities market—a comparison that should give observers pause, given how that market eventually collapsed.

The GPU loan by another name#

This is not, as the companies present it, a breakthrough financial innovation. It is essentially an upgraded version of GPU-backed lending that already exists. The core mechanism remains unchanged: financial firms lend money to data center operators and cryptocurrency miners using Nvidia chips as collateral.

The terminology shift from "GPU loans" to "compute assets" serves a purpose. Everyone knows that a graphics processor typically depreciates in two to five years, but calling it "compute" creates room to argue for longer asset life cycles. This matters enormously to lenders, because depreciation schedules directly determine loan terms and how much money borrowers can access.

Broadcom tested this playbook earlier this year with a $35 billion financing deal alongside Apollo and Blackstone. Nvidia appears to have taken note and is running the same strategy.

The depreciation sleight of hand#

The timing and framing of Huang's messaging reveals the stakes. Last year, he claimed that when Blackwell chips began shipping in volume, older Hopper processors "couldn't be given away." Now he argues that a 2020-era A100 chip remains commercially viable a decade after launch, with customers committing to multi-year deployments that could extend its "economic life toward a decade."

If Huang succeeds in establishing a ten-year depreciation timeline in the eyes of major lenders, it gives borrowers access to more favorable loan terms. Short seller Michael Burry has argued that two to three years is the appropriate depreciation window; IBM's Arvind Krishna suggests five years. But when the CEO of the chip maker himself insists on a decade-long valuation, banks have little reason to reject it.

The financial reality hinges on inference—running trained AI models on new data—which has driven sustained demand and kept rental prices for older chips high or even rising. CoreWeave, a pioneer in GPU lending and major Nvidia customer, has successfully extended contracts for 2020-era chips through 2029. This real-world data provides cover for Huang's longer depreciation claims.