Why Nvidia is guaranteeing $250 billion of OpenAI's debt: the same reason showed up in the fiber-optic bubble 25 years ago

Why this story is back

A Bloomberg-sourced report today could move Korean markets too. Nvidia is reportedly discussing guaranteeing $250 billion of OpenAI's Ohio data-center debt, plus up to $350 billion more so OpenAI can buy Nvidia's own chips. OpenAI is an unprofitable private company with sub-investment-grade credit. A data-center builder needs its tenant's credit to borrow cheaply from banks, and Nvidia is filling that gap for OpenAI. The same dollars move from Nvidia to OpenAI, to Oracle, and back to Nvidia as chip revenue. Add Nvidia's stakes in CoreWeave, Nebius and IREN, and this circular financing tops $540 billion.

The same structure showed up 25 years ago

In the late 1990s, telecom firms raced to lay fiber-optic cable, betting the internet would make data traffic explode. Cash-poor startups borrowed from equipment makers to buy that equipment, and the loans were structured so the money flowed straight back as the lender's own sales. Nearly $30 billion in such vendor loans went out in 2000 alone.

Then data-compression techniques and DWDM (wavelength-division multiplexing) spread through the market, and the fiber already in the ground turned out to be more than enough. One new technology was all it took to pop the bubble. The Nasdaq, which had peaked at 5,048 in March 2000, fell 78% by 2002, and $2 trillion of telecom-stock value vanished. Roughly 95% of the fiber laid at the time was never lit even once.

Where it differs this time

The fiber bubble was built on demand that was overestimated from the start. Today's AI data-center demand is, at least so far, backed by real contracts and revenue. By scale, though, this round is already bigger. As of 2025, Nvidia's vendor financing equaled 67% of its own revenue, 2.8 times larger relative to revenue than Lucent's vendor loans at the telecom bubble's height.

Circular Financing: Does Nvidia's $110B Bet Echo the Telecom Bubble? (Tomasz Tunguz, 2025-10-03)

출처: tomtunguz.com

Relative to revenue, this loop has already passed the dot-com peak, so the shock could be larger too if the chain ever breaks.

The weakest link

The weakest part of this structure is OpenAI's own finances. OpenAI reportedly spends $1.22 for every dollar it earns, and its IPO-track filings project breaking even only around 2030. HSBC estimates OpenAI needs at least $207 billion more financing through 2030 to stay afloat.

OpenAI must find $207bn to meet AI data center spending commitments - HSBC (DataCenterDynamics, 2025-11-28)

출처: datacenterdynamics.com

A $25 billion-revenue company would have to carry $80-100 billion a year in compute commitments for five years running. Nvidia, Oracle and OpenAI are investor, supplier and customer to each other at once, so trouble at any one can spread like dominoes. Nvidia's cash generation should keep it standing, but OpenAI and Oracle are first in line.

What decides this

The question comes down to three things: can AI make enough money to justify hundreds of billions in infrastructure, can OpenAI keep raising cash until it does, and does a more efficient technology arrive that needs fewer GPUs. If OpenAI keeps raising money and grows into its contracts, this loop ends differently from 25 years ago. If signs appear it cannot close HSBC's $207 billion gap, the picture changes first. Last time, a single new technology was enough to make the demand itself disappear. Whether that happens again decides how long this circular-financing structure survives.